# text-to-image

A machine-learning technology that generates images from text prompts.

This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles.

## VerifyPrompt: Towards Text-to-Image Models Verification Behind Black-Box APIs

DevFeed: [VerifyPrompt: Towards Text-to-Image Models Verification Behind Black-Box APIs](<https://devfeed.tech/articles/verifyprompt-towards-text-to-image-models-verification-behind-black-box-apis-143098.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2410.22725>)

Author: Ji Guo, Rui Zhang, Guomin Lu, Aiguo Chen, Tianwei Zhang, Hongwei Li, Wenbo Jiang

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>), [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>)

Tags: [adversarial](<https://devfeed.tech/tags/adversarial.md>), [apis](<https://devfeed.tech/tags/apis.md>), [black-box](<https://devfeed.tech/tags/black-box.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [fraud](<https://devfeed.tech/tags/fraud.md>), [image](<https://devfeed.tech/tags/image.md>), [learning](<https://devfeed.tech/tags/learning.md>), [model](<https://devfeed.tech/tags/model.md>), [model-providers](<https://devfeed.tech/tags/model-providers.md>), [model-verification](<https://devfeed.tech/tags/model-verification.md>), [models](<https://devfeed.tech/tags/models.md>), [oracle](<https://devfeed.tech/tags/oracle.md>), [platforms](<https://devfeed.tech/tags/platforms.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [search](<https://devfeed.tech/tags/search.md>), [security](<https://devfeed.tech/tags/security.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>), [third-party](<https://devfeed.tech/tags/third-party.md>), [verification](<https://devfeed.tech/tags/verification.md>)

### AI overview

VerifyPrompt tests whether a black-box text-to-image service matches a claimed reference model, addressing possible model substitution by third-party platforms. It searches for a model-specific adversarial verify prompt designed to make the target generate a particular image while other models generate different ones. VerifyPrompt-RL uses reinforcement learning, multi-sample exploration, and group relative error estimation to stabilize and speed that black-box search. The abstract reports over 90% verification accuracy across various T2I platforms; it does not provide further experimental details here.

### Source excerpt

arXiv:2410.22725v5 Announce Type: replace-cross Abstract: With the rise of diverse Text-to-Image (T2I) models such as DALL-E 3, third-party platforms increasingly integrate multiple black-box APIs to deliver cost-effective services. This also raises a new and practical security question: are these third-party platforms truly offering the models they claim? Such model substitution fraud, in which a malicious third-party platform operator acts as the adversary, enables the platform to profit illicitly while undermining user interests and damaging the commercial reputation of legitimate model providers. To address this practical concern, we first define the task of T2I model verification, which determines whether a black-box target model matches a black-box oracle reference model. After that, we propose VerifyPrompt, which performs T2I model verification through a specially designed verify prompt. Intuitively, the verify prompt is an adversarial prompt for the target model, without transferability to other models. It makes the target model generate a specific image while making other models produce entirely different images. Our key idea for generating verify prompts is to enforce strong interactions between adversarial perturbations and the target model, thereby overfitting the target and producing model-specific perturbations. To enable efficient black-box search for verify prompts, we further introduce VerifyPrompt-RL, a reinforcement learning-based strategy that leverages multi-sample exploration and group relative error estimation to stabilize and accelerate the search process. Experiments show that VerifyPrompt achieves over 90\% accuracy across various T2I platforms, confirming its effectiveness in real-world third-party services.

## Beyond Semantic Prompts: Unified Watermarking and Access Control for Text-to-Image Diffusion Models

DevFeed: [Beyond Semantic Prompts: Unified Watermarking and Access Control for Text-to-Image Diffusion Models](<https://devfeed.tech/articles/beyond-semantic-prompts-unified-watermarking-and-access-control-for-text-to-image-diffusion-models-139302.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.20470>)

Author: Fangming Shi, Li Li, Guorui Feng, Xinpeng Zhang

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [stable-diffusion](<https://devfeed.tech/topics/stable-diffusion.md>), [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>), [Access Control](<https://devfeed.tech/topics/access-control.md>), [watermarking](<https://devfeed.tech/topics/watermarking.md>)

Tags: [access-control](<https://devfeed.tech/tags/access-control.md>), [autoencoder](<https://devfeed.tech/tags/autoencoder.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [customizable](<https://devfeed.tech/tags/customizable.md>), [decoder](<https://devfeed.tech/tags/decoder.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [diffusion-models](<https://devfeed.tech/tags/diffusion-models.md>), [generation](<https://devfeed.tech/tags/generation.md>), [image-processing](<https://devfeed.tech/tags/image-processing.md>), [images](<https://devfeed.tech/tags/images.md>), [models](<https://devfeed.tech/tags/models.md>), [prompt-structure](<https://devfeed.tech/tags/prompt-structure.md>), [prompts](<https://devfeed.tech/tags/prompts.md>), [protocol](<https://devfeed.tech/tags/protocol.md>), [quality](<https://devfeed.tech/tags/quality.md>), [recognition](<https://devfeed.tech/tags/recognition.md>), [recovery](<https://devfeed.tech/tags/recovery.md>), [robustness](<https://devfeed.tech/tags/robustness.md>), [services](<https://devfeed.tech/tags/services.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>), [watermarking](<https://devfeed.tech/tags/watermarking.md>)

### AI overview

The proposed framework combines multi-bit watermarking and tiered access control in a single text prompt. It encodes watermark messages as designated words with dedicated learnable tokens, then jointly adapts the diffusion U-Net and token embeddings using a frozen reference model and pretrained watermark codec. A prompt-structure recognition network controls the VAE decoder to produce clear watermarked images, blurred previews, or black outputs according to access state. The abstract reports reliable watermark recovery, robustness to tested image-processing operations and model modifications, and preserved decoding fidelity and watermark recovery under full access. It does not provide quantitative results in the supplied excerpt.

### Source excerpt

arXiv:2611.20470v1 Announce Type: new Abstract: Text-to-image diffusion services require both traceability of generated content and differentiated access to generation capabilities. However, existing watermarking methods commonly expose watermark control through mechanisms beyond the text-prompt interface, while credential-based access control does not inherently enforce distinct output quality levels. We propose a unified framework that enables customizable multi-bit watermark embedding and tiered access control through a single text prompt. A unified prompt represents watermark messages using designated words and associates protocol structures with different access states. The framework adopts a two-stage training strategy. First, watermark payload partitioning represents messages as designated watermark words, while vocabulary extension assigns each protocol word a dedicated token with a learnable input embedding. Guided by a frozen reference model and a pretrained watermark codec, joint adaptation of the U-Net and protocol-token embeddings transfers message-dependent watermark patterns into the generation process. Second, a prompt-structure recognition network controls feature modulation within the variational autoencoder decoder, which is trained to produce clear watermarked images, blurred previews, or black outputs according to the access state. Experiments demonstrate reliable watermark recovery, robustness to common image-processing operations and evaluated model modifications, and consistent performance across diverse generation settings. The controlled decoder also achieves distinct output behaviors under restricted access while closely preserving reference decoding fidelity and watermark recoverability under full access. Our open-source code and model weights will be public after acceptance.

## Rethinking Visual Reasoning in Text-to-Image Reward Modeling

DevFeed: [Rethinking Visual Reasoning in Text-to-Image Reward Modeling](<https://devfeed.tech/articles/rethinking-visual-reasoning-in-text-to-image-reward-modeling-135083.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.17585>)

Author: Shiye Su, Xiaohan Wang, Ludwig Schmidt, Serena Yeung-Levy

Published: 2026-11-16T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>), [human feedback](<https://devfeed.tech/topics/human-feedback.md>), [post-training](<https://devfeed.tech/topics/post-training.md>), [vlm](<https://devfeed.tech/topics/vlm.md>)

Tags: [continual-learning](<https://devfeed.tech/tags/continual-learning.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [data](<https://devfeed.tech/tags/data.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [evaluations](<https://devfeed.tech/tags/evaluations.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [generation](<https://devfeed.tech/tags/generation.md>), [image](<https://devfeed.tech/tags/image.md>), [image-generation](<https://devfeed.tech/tags/image-generation.md>), [images](<https://devfeed.tech/tags/images.md>), [improvement](<https://devfeed.tech/tags/improvement.md>), [inference](<https://devfeed.tech/tags/inference.md>), [modeling](<https://devfeed.tech/tags/modeling.md>), [pairing](<https://devfeed.tech/tags/pairing.md>), [pairs](<https://devfeed.tech/tags/pairs.md>), [post-training](<https://devfeed.tech/tags/post-training.md>), [prediction](<https://devfeed.tech/tags/prediction.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [reward](<https://devfeed.tech/tags/reward.md>), [text](<https://devfeed.tech/tags/text.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>)

### AI overview

Reward models used for text-to-image selection and post-training can lose visual reasoning ability during preference fine-tuning, while continual-learning baselines do not resolve the reported tradeoff with preference accuracy. The paper pairs existing visual reasoning examples by keeping the image fixed and varying the prompt, then mixes these pairs with conventional human preferences, which pair images under a shared prompt. The authors report improved visual understanding and preference prediction, plus average GenEval gains of 4.0% in best-of-n selection and 5.5% in RL post-training. Their model also outperforms HPSv3, despite HPSv3 having been trained on six times more preference pairs.

### Source excerpt

arXiv:2611.17585v1 Announce Type: cross Abstract: Reward models are a core component of modern text-to-image systems, serving as the objective for inference-time selection and RL-based post-training. A key prerequisite is visual understanding: to judge prompt adherence and ground aesthetic preferences, a reward model must bind attributes, resolve relations, and count reliably. We find that widely-used reward models perform poorly on targeted visual reasoning evaluations. Even when starting from strong vision-language backbones, preference fine-tuning erodes these capabilities, and continual-learning baselines fail to mitigate a steep tradeoff between visual reasoning and preference accuracy. We propose a data-centric intervention that injects visual reasoning supervision into preference learning by flipping the pairing axis. To complement conventional human preference data (paired images under a shared prompt), we convert existing visual reasoning datasets into preference pairs that share the same image but differ in prompt, and train on the mixture. This yields a Pareto improvement in both visual understanding and preference prediction. Across inference-time best-of-n selection and RL post-training, the resulting reward model improves text-to-image generation on challenging compositional prompts (GenEval) by an average of 4.0% and 5.5% respectively, outperforming HPSv3 which was trained on 6x more preference pairs.

## When Chekhov's Gun Fires Anyway: Over-Realization in Text-to-Image Models

DevFeed: [When Chekhov's Gun Fires Anyway: Over-Realization in Text-to-Image Models](<https://devfeed.tech/articles/when-chekhov-s-gun-fires-anyway-over-realization-in-text-to-image-models-132264.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.16630>)

Author: Jiahao Huang, Kon Woo Kim, Song-Duo Ma, Yuka Kitamura, Felix Wagner, Yunshan Huang, Sunisth Kumar, Xanh Ho, Yun-Ang Wu, Zhijie Chen, Yunjia Xu, Junfeng Jiang, Zefan Yu, Fei Cheng, Akiko Aizawa

Published: 2026-11-13T05:00:00Z

Content type: news

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>)

Tags: [attention](<https://devfeed.tech/tags/attention.md>), [computer-vision](<https://devfeed.tech/tags/computer-vision.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [image-generation](<https://devfeed.tech/tags/image-generation.md>), [metrics](<https://devfeed.tech/tags/metrics.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>)

### AI overview

Text-to-image models often depict a mentioned target even when the prompt negates it, places it in a character's imagination or a figurative comparison, or hides it from view. The paper calls this over-realization and introduces OverReal: 698 prompts to measure rates across these four situations, plus 2,841 human-labeled images to verify its automatic judge. Across eight systems, all over-realize; conventional metrics such as CLIPScore reward the failure, and prompt expansion only partly reduces it. The authors report that diffusion backbones attend nearly as much to merely mentioned targets as to requested ones, while cues meant to suppress those targets receive less attention than ordinary words.

### Source excerpt

arXiv:2611.16630v1 Announce Type: cross Abstract: In text-to-image generation, a target that the prompt mentions often appears in the generated image in a way the prompt does not call for. We call this failure over-realization. Mental-spaces theory points to the situations where it can happen: under a negation, inside what a character believes or imagines, as the vehicle of a figurative comparison, or behind a viewpoint that hides the target. Prior work has studied only negation in depth, and the conventional alignment and preference metrics do not penalize the failure. We build OverReal to measure it: OverReal-Gen, 698 prompts covering the four situations, on which we measure each generator's over-realization rate; and OverReal-Det, 2,841 generated images human-labeled for whether and how the target is over-realized, on which we verify the automatic judge that produces this rate. Our evaluation of eight text-to-image systems reveals that all of them over-realize, and that conventional metrics such as CLIPScore reward the failure rather than penalize it. Prompt expansion only partly mitigates the failure. Inside the generator, the diffusion backbone attends to a mentioned target almost as much as to a requested one. The cue words that should hold the target back receive less attention than ordinary words.

## Verifier Choice is a Benchmark Design Variable: Auditing Structural Counting Evaluation in Text-to-Image Models

DevFeed: [Verifier Choice is a Benchmark Design Variable: Auditing Structural Counting Evaluation in Text-to-Image Models](<https://devfeed.tech/articles/verifier-choice-is-a-benchmark-design-variable-auditing-structural-counting-evaluation-in-text-to-image-models-132994.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.15433>)

Author: Shurun Li, Michael Hua Wang, Haibo Zeng, Long Wang

Published: 2026-11-13T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [vlm](<https://devfeed.tech/topics/vlm.md>), [Benchmark](<https://devfeed.tech/topics/benchmark.md>), [Trustworthy AI](<https://devfeed.tech/topics/trustworthy-ai.md>), [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>), [flux](<https://devfeed.tech/topics/flux.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [Claude](<https://devfeed.tech/topics/claude.md>)

Tags: [aggregate](<https://devfeed.tech/tags/aggregate.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [audit](<https://devfeed.tech/tags/audit.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [automatic](<https://devfeed.tech/tags/automatic.md>), [behavior](<https://devfeed.tech/tags/behavior.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [benchmark-design](<https://devfeed.tech/tags/benchmark-design.md>), [benchmarking](<https://devfeed.tech/tags/benchmarking.md>), [bias](<https://devfeed.tech/tags/bias.md>), [chairs](<https://devfeed.tech/tags/chairs.md>), [claude](<https://devfeed.tech/tags/claude.md>), [claude-opus](<https://devfeed.tech/tags/claude-opus.md>), [clover](<https://devfeed.tech/tags/clover.md>), [compliance](<https://devfeed.tech/tags/compliance.md>), [counts](<https://devfeed.tech/tags/counts.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [image](<https://devfeed.tech/tags/image.md>), [models](<https://devfeed.tech/tags/models.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>)

### AI overview

The paper finds that structural-counting benchmark scores depend on the verifier and how it is used, not just on the image generator. Across 6,240 images and seven object or substructure categories, the authors compare open MLLM verifiers under count-only and reject-aware, zero-shot and few-shot protocols, alongside TIFA-like queries, a detector baseline, and closed-model checks on two difficult objects. On a balanced 2,280-image split, similar aggregate scores can hide image-level disagreement; every tested verifier-protocol setup narrows the human-measured gap between Qwen-Image-2512 and FLUX.2-dev by 7.5-25 percentage points. With Z-Image added, all four tested protocols reverse its human ranking against FLUX.2-dev on the fixed prompt suite. Closed-model checks retain failures: GPT-5.5 achieves 59.92% valid-count fidelity on clover, while Claude Opus 4.7 has κ=0.432 same-slice agreement. The authors propose reporting verifier reliability across score gaps, image-level agreement, generator-comparison distortion, valid-count fidelity, rejection, output compliance, and directional bias.

### Source excerpt

arXiv:2611.15433v1 Announce Type: cross Abstract: Automatic evaluation is now central to text-to-image benchmarking, but the verifier that converts generated images into scores is often treated as an implementation detail. We argue that this is unsafe: an automatic benchmark score is a property of the image generator, prompt distribution, verifier, verifier-use protocol, and scoring rule, not of the generator alone. We study this issue in structural counting, a controlled setting where human-visible counts can be audited. We construct a verifier-centric audit suite over seven counting objects: cell phones, chairs, and traffic lights for object counting, and clover leaflets, dresser drawer fronts, power-strip sockets, and turbine blades for substructure counting. Across 6,240 generated images from Qwen-Image-2512 and FLUX.2-dev, we compare open MLLM verifiers under count-only, reject-aware, zero-shot and few-shot protocols, plus TIFA-like target queries, a detector baseline, and frontier closed-model checks on two hard objects. On a balanced 2,280-image evaluation split, aggregate score matching can coexist with image-level disagreement, and verifier-use protocols and generator distributions change the measured reliability. Every verifier-protocol configuration also shrinks the human score difference between the two generators, by 7.5-25.0 pp, and with a third generator, Z-Image, all four tested protocols reverse the human ordering of Z-Image and FLUX.2-dev on our fixed prompt suite. GPT-5.5 and Claude Opus 4.7 checks do not remove the failure modes: GPT-5.5 still has only 59.92% clover valid-count fidelity, while Claude Opus 4.7 has lower same-slice agreement ($\kappa=0.432$). We propose reporting automatic benchmark scores with a factorized verifier reliability profile: aggregate score gap, image-level agreement, generator-comparison distortion, valid-image count fidelity, rejection behavior, output compliance, and directional bias.

## Refining Few-Step Text-to-Multiview Diffusion via Reinforcement Learning

DevFeed: [Refining Few-Step Text-to-Multiview Diffusion via Reinforcement Learning](<https://devfeed.tech/articles/refining-few-step-text-to-multiview-diffusion-via-reinforcement-learning-132721.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2505.20107>)

Author: Ziyi Zhang, Li Shen, Deheng Ye, Yong Luo, Huangxuan Zhao, Meng Liu, Wei Yu, Lefei Zhang

Published: 2026-11-13T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>), [Model Development](<https://devfeed.tech/topics/model-development.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [text-to-image](<https://devfeed.tech/topics/text-to-image.md>)

Tags: [code](<https://devfeed.tech/tags/code.md>), [consistency](<https://devfeed.tech/tags/consistency.md>), [constraint](<https://devfeed.tech/tags/constraint.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [finetuning](<https://devfeed.tech/tags/finetuning.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [reward](<https://devfeed.tech/tags/reward.md>), [text](<https://devfeed.tech/tags/text.md>)

### AI overview

MVC-ZigAL is an RL fine-tuning framework for few-step text-to-multiview diffusion models, which can generate views in real time but may lose per-view fidelity and cross-view consistency. It jointly models generated views and scores their collective quality, uses gains from self-refinement sampling over standard sampling to strengthen advantage-learning signals, and applies Lagrangian dual optimization with adaptive primal-dual updates and a self-paced threshold curriculum to balance single-view and multiview objectives. The abstract reports gains in both fidelity and consistency; it provides no quantitative results or evaluation details.

### Source excerpt

arXiv:2505.20107v3 Announce Type: replace Abstract: Text-to-multiview (T2MV) diffusion models have shown great promise in generating multiple views of a scene from a single text prompt. While few-step backbones enable real-time T2MV generation, they often compromise key aspects of generation quality, such as per-view fidelity and cross-view consistency. Reinforcement learning (RL) finetuning offers a potential solution, yet existing approaches designed for single-image diffusion do not readily extend to the few-step T2MV setting, as they neglect cross-view coordination and suffer from weak learning signals in few-step regimes. To address this, we propose MVC-ZigAL, a tailored RL finetuning framework for few-step T2MV diffusion models. Specifically, its core insights are: (1) a new MDP formulation that jointly models all generated views and assesses their collective quality via a joint-view reward; (2) a novel advantage learning strategy that exploits the performance gains of a self-refinement sampling scheme over standard sampling, yielding stronger learning signals for effective RL finetuning; and (3) a unified RL framework that extends advantage learning with a Lagrangian dual formulation for multiview-constrained optimization, balancing single-view and joint-view objectives through adaptive primal-dual updates under a self-paced threshold curriculum that harmonizes exploration and constraint enforcement. Collectively, these designs enable robust and balanced RL finetuning for few-step T2MV diffusion models, yielding substantial gains in both per-view fidelity and cross-view consistency. Code is available at https://github.com/ZiyiZhang27/MVC-ZigAL.

## Steering Away from Memorization: Reachability-Constrained Reinforcement Learning for Text-to-Image Diffusion

DevFeed: [Steering Away from Memorization: Reachability-Constrained Reinforcement Learning for Text-to-Image Diffusion](<https://devfeed.tech/articles/steering-away-from-memorization-reachability-constrained-reinforcement-learning-for-text-to-image-diffusion-131887.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2603.00140>)

Author: Sathwik Karnik, Juyeop Kim, Sanmi Koyejo, Jong-Seok Lee, Somil Bansal

Published: 2026-11-13T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>)

Tags: [alignment](<https://devfeed.tech/tags/alignment.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [diversity](<https://devfeed.tech/tags/diversity.md>), [mitigation](<https://devfeed.tech/tags/mitigation.md>), [quality](<https://devfeed.tech/tags/quality.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [steering](<https://devfeed.tech/tags/steering.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>)

### AI overview

RADS aims to reduce memorization in text-to-image diffusion without sacrificing generation fidelity. It treats denoising as a dynamical system and learns an offline policy-conditioned reachability value function to estimate future memorization risk. During sampling, the policy steers trajectories away from risk by perturbing caption embeddings, leaving the pretrained model separate from the controller. The abstract reports a stronger empirical Pareto frontier than state-of-the-art inference-time baselines across diversity (SSCD), quality (FID), and alignment (CLIP); it gives no quantitative results or implementation details.

### Source excerpt

arXiv:2603.00140v2 Announce Type: replace-cross Abstract: Text-to-image diffusion models are susceptible to memorization, revealing a fundamental failure to generalize beyond the training set. Current mitigation approaches often sacrifice image quality or prompt alignment to reduce memorization. To address this, we propose Reachability-Aware Diffusion Steering (RADS), an offline-trained, sampling-time control framework that mitigates memorization while maintaining generation fidelity. RADS models the diffusion denoising process as a dynamical system and learns a policy-conditioned reachability value function that estimates future memorization risk along the denoising trajectory. This motivates a constrained reinforcement learning (RL) formulation, where a policy learns to steer the trajectory away from memorization via perturbations in the caption embedding space. Empirical evaluations show that RADS achieves a stronger empirical Pareto frontier between generation diversity (SSCD), quality (FID), and alignment (CLIP) compared to state-of-the-art inference-time baselines. Crucially, RADS applies the learned controller during sampling time, separating memorization mitigation from the pretrained model. Our website is available at: https://s-karnik.github.io/rads-memorization-project-page/.

## ORCA: Hunting Compositional Failures in Text-to-Image Diffusion

DevFeed: [ORCA: Hunting Compositional Failures in Text-to-Image Diffusion](<https://devfeed.tech/articles/orca-hunting-compositional-failures-in-text-to-image-diffusion-128585.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.09841>)

Author: Arshia Hemmat, Amirhossein Vahidi, Amitis Shidani, Mohammad Vali Sanian, Hesam Asadollahzadeh, Aryan Yazdan Parast, Mohammad Lotfollahi

Published: 2026-11-11T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [diffusion-transformers](<https://devfeed.tech/topics/diffusion-transformers.md>)

Tags: [alignment](<https://devfeed.tech/tags/alignment.md>), [bind](<https://devfeed.tech/tags/bind.md>), [clip](<https://devfeed.tech/tags/clip.md>), [covariance](<https://devfeed.tech/tags/covariance.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [denoising](<https://devfeed.tech/tags/denoising.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [diffusion-models](<https://devfeed.tech/tags/diffusion-models.md>), [embedding](<https://devfeed.tech/tags/embedding.md>), [failures](<https://devfeed.tech/tags/failures.md>), [features](<https://devfeed.tech/tags/features.md>), [inference](<https://devfeed.tech/tags/inference.md>), [information](<https://devfeed.tech/tags/information.md>), [language](<https://devfeed.tech/tags/language.md>), [modal](<https://devfeed.tech/tags/modal.md>), [object](<https://devfeed.tech/tags/object.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [rank](<https://devfeed.tech/tags/rank.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>), [training](<https://devfeed.tech/tags/training.md>), [transformer](<https://devfeed.tech/tags/transformer.md>)

### AI overview

ORCA targets compositional failures in text-to-image diffusion by adding one auxiliary training loss that aligns a diffusion transformer's latent with a low-rank target from a frozen visual encoder. A learned residual between T5 and CLIP embeddings parameterizes an orthogonal basis, providing a prompt-dependent signal for selecting the visual readout subspace. Across DiT-B/2, DiT-L/2, and U-ViT-L, the abstract reports improved FID and GenEval over vanilla and REPA baselines, with no added inference-time cost. On DiT-L/2, ORCA reaches FID 16.65 and GenEval 0.291 at 200K steps, surpassing the strongest 400K baseline at half the training cost; reported gains are largest for attribute binding, spatial relations, and multi-object prompts.

### Source excerpt

arXiv:2610.09841v2 Announce Type: replace-cross Abstract: Text-to-image diffusion models fail predictably on compositional prompts: attributes bind to the wrong objects, spatial relations invert, and multi-object scenes lose count. Recent architectures already augment CLIP with a T5 encoder precisely because CLIP's contrastive embedding loses compositional structure, yet these failures persist. We argue the binding problem is therefore not one of missing information but of misaligned information: a text encoder preserves compositional structure, but in a representation space shaped by language modelling rather than vision, and the denoising objective does not directly reward aligning the two. We show this correspondence can be supplied as an explicit training signal, that the relevant cross-modal information is concentrated in a low-rank subspace of self-supervised visual features, and that supplying it can be folded into diffusion training as a single auxiliary loss. Our method, ORCA (Orthogonal Residual Compositional Alignment), aligns the latent of a diffusion transformer with a low-rank target derived from a frozen visual encoder, through a predictor whose orthogonal basis is parameterised by a learned residual between T5 and CLIP embeddings, which provides a prompt-dependent signal for selecting the visual readout subspace. We prove that the cross-modal information recoverable at a given rank is bounded by the spectral mass of the visual encoder's covariance in the top components. Across three diffusion-transformer backbones (DiT-B/2, DiT-L/2, U-ViT-L), ORCA improves FID and GenEval over both vanilla and REPA baselines at zero inference-time cost; on DiT-L/2 it reaches FID 16.65 and GenEval 0.291 at 200K steps, exceeding the strongest 400K baseline at half the training cost, with the largest gains concentrated on attribute binding, spatial relations, and multi-object prompts.

## The RenAIssance: Caption-Guided Diffusion Models for Historical-Art Visual Media

DevFeed: [The RenAIssance: Caption-Guided Diffusion Models for Historical-Art Visual Media](<https://devfeed.tech/articles/the-renaissance-caption-guided-diffusion-models-for-historical-art-visual-media-125223.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.10599>)

Author: Zhanyu Tuo

Published: 2026-11-10T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [stable-diffusion](<https://devfeed.tech/topics/stable-diffusion.md>), [lora](<https://devfeed.tech/topics/lora.md>)

Tags: [adapter](<https://devfeed.tech/tags/adapter.md>), [ai](<https://devfeed.tech/tags/ai.md>), [art](<https://devfeed.tech/tags/art.md>), [authoring](<https://devfeed.tech/tags/authoring.md>), [clip](<https://devfeed.tech/tags/clip.md>), [computer-science](<https://devfeed.tech/tags/computer-science.md>), [computer-vision](<https://devfeed.tech/tags/computer-vision.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-mm](<https://devfeed.tech/tags/cs-mm.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [diffusion-models](<https://devfeed.tech/tags/diffusion-models.md>), [generative-ai](<https://devfeed.tech/tags/generative-ai.md>), [interactive](<https://devfeed.tech/tags/interactive.md>), [lora](<https://devfeed.tech/tags/lora.md>), [media](<https://devfeed.tech/tags/media.md>), [models](<https://devfeed.tech/tags/models.md>), [pipeline](<https://devfeed.tech/tags/pipeline.md>), [stable-diffusion](<https://devfeed.tech/tags/stable-diffusion.md>), [style](<https://devfeed.tech/tags/style.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>), [visual-language](<https://devfeed.tech/tags/visual-language.md>)

### AI overview

The RenAIssance pipeline turns historical European artworks into a reusable text-to-image resource: BLIP generates scene-level captions, and about 15,000 DEArt image-text pairs train a compact LoRA adapter for Stable Diffusion v1.5. The adapter steers generations toward the historical-art distribution while retaining the base model's ability to render novel concepts. Across historical and contemporary prompts of varying lengths, it improves FID against DEArt references in every reported setting, with a modest decrease in CLIP text-image similarity. Qualitative results show period-appropriate subjects and contemporary concepts rendered in a coherent historical visual language. The supplied abstract does not provide metric values or further implementation details.

### Source excerpt

arXiv:2611.10599v1 Announce Type: cross Abstract: Generative AI can expand cultural collections into prompt-responsive visual media, but historical artworks rarely provide the paired descriptions needed to adapt modern image generators. We present The RenAIssance, a scalable pipeline that turns a collection of historical European artworks into a reusable text-to-image media resource. The pipeline generates scene-level captions with BLIP and uses the resulting approximately 15,000 image--text pairs from DEArt to train a compact LoRA adapter for Stable Diffusion v1.5. The adapter shifts generated images toward the historical-art distribution while retaining the base model's ability to render novel concepts. On historical and contemporary prompts of different lengths, it improves FID to the DEArt reference distribution in every setting with only a modest decrease in CLIP text--image similarity. Qualitative results show that the same adapter can render both period-appropriate subjects and contemporary concepts in a coherent historical visual language. By decoupling a reusable style representation from the base generator, the approach offers a practical building block for controllable historical-art media creation, interactive authoring, and rich-media experiences.

## How Text-to-Image AI Generators Work, and 5 Tools That Deliver Good Results

DevFeed: [How Text-to-Image AI Generators Work, and 5 Tools That Deliver Good Results](<https://devfeed.tech/articles/how-text-to-image-ai-generators-work-and-5-tools-that-deliver-good-results-123748.md>)

Original publisher: [Read original article](<https://www.sitepoint.com/how-text-to-image-ai-generators-work-and-5-tools-that-deliver-good-results/>)

Author: Div John

Published: 2026-11-09T07:15:58Z

Content type: article

Language: en

Sources: [SitePoint](<https://devfeed.tech/sources/sitepoint.md>)

Topics: [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>), [generative-ai-tools](<https://devfeed.tech/topics/generative-ai-tools.md>), [stable-diffusion](<https://devfeed.tech/topics/stable-diffusion.md>), [Image](<https://devfeed.tech/topics/image.md>), [Tool](<https://devfeed.tech/topics/tool.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-tools](<https://devfeed.tech/tags/ai-tools.md>), [ai-video](<https://devfeed.tech/tags/ai-video.md>), [developer-tools](<https://devfeed.tech/tags/developer-tools.md>), [generators](<https://devfeed.tech/tags/generators.md>), [image](<https://devfeed.tech/tags/image.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>), [tools](<https://devfeed.tech/tags/tools.md>), [web](<https://devfeed.tech/tags/web.md>)

### AI overview

Text-to-image generators encode prompts as numerical representations, then guide a diffusion model as it removes noise step by step; latent representations make this process more efficient. Clear, focused prompts and changing one variable at a time can improve iteration, while hands, small text, crowded prompts and exact spatial relationships remain common failure points. Tool strengths differ: Pollo AI offers varied styles and an image-to-video workflow; Midjourney emphasizes polished artistic results but may follow literal instructions less closely; Adobe Firefly connects to Adobe editing tools and describes licensed training data and Content Credentials; Ideogram is suited to short readable text in images; Leonardo AI offers models, references and editing controls for consistent assets, with a steeper learning curve. Plans and rights vary, so the text advises checking current terms and reviewing outputs for artifacts before use.

### Source excerpt

null Continue reading How Text-to-Image AI Generators Work, and 5 Tools That Deliver Good Results on SitePoint.

## Panoptic Scene Program Diffusion Transformer

DevFeed: [Panoptic Scene Program Diffusion Transformer](<https://devfeed.tech/articles/panoptic-scene-program-diffusion-transformer-123211.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2609.31780>)

Author: Chika Maduabuchi

Published: 2026-11-09T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [diffusion-transformers](<https://devfeed.tech/topics/diffusion-transformers.md>), [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [Image](<https://devfeed.tech/topics/image.md>)

Tags: [architecture](<https://devfeed.tech/tags/architecture.md>), [consistency](<https://devfeed.tech/tags/consistency.md>), [counts](<https://devfeed.tech/tags/counts.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [grounding](<https://devfeed.tech/tags/grounding.md>), [identity](<https://devfeed.tech/tags/identity.md>), [image](<https://devfeed.tech/tags/image.md>), [inference](<https://devfeed.tech/tags/inference.md>), [models](<https://devfeed.tech/tags/models.md>), [object](<https://devfeed.tech/tags/object.md>), [ordering](<https://devfeed.tech/tags/ordering.md>), [overhead](<https://devfeed.tech/tags/overhead.md>), [ownership](<https://devfeed.tech/tags/ownership.md>), [program](<https://devfeed.tech/tags/program.md>), [psp](<https://devfeed.tech/tags/psp.md>), [quality](<https://devfeed.tech/tags/quality.md>), [robust](<https://devfeed.tech/tags/robust.md>), [settings](<https://devfeed.tech/tags/settings.md>), [spatial](<https://devfeed.tech/tags/spatial.md>), [transformer](<https://devfeed.tech/tags/transformer.md>)

### AI overview

PSP-DiT targets compositional prompt errors by jointly denoising image latents and panoptic scene-program latents in coupled transformer streams. Panoptic grounding and cycle-consistency objectives connect instances, attributes, relations, and counts to visual evidence in the generated image. Under matched training and inference settings, it outperforms a strong flat-text baseline on GenEval 2, SANEval-Simple, PSG-Score, and DetailMaster, with the largest gains reported for counting, attribute binding, role-sensitive relations, and long structured prompts. The abstract says image quality is preserved, inference overhead is modest, and performance remains robust to imperfect scene programs.

### Source excerpt

arXiv:2609.31780v2 Announce Type: replace-cross Abstract: Modern text-to-image models produce high-fidelity images but still struggle with compositional prompts that require instance identity, attribute ownership, counting, spatial ordering, and role-sensitive relations. We introduce Panoptic Scene Program Diffusion Transformer (PSP-DiT), a diffusion-transformer architecture that treats a panoptic scene program as a first-class latent variable rather than an external control signal or post-hoc parse. PSP-DiT jointly denoises image latents and scene-program latents through coupled transformer streams, while panoptic grounding and cycle-consistency objectives tie object instances, attributes, relations, and counts to visual support in the generated image. Under matched training and inference settings, PSP-DiT improves over a strong flat-text baseline across GenEval 2, SANEval-Simple, PSG-Score, and DetailMaster, with the largest gains on counting, attribute binding, role-sensitive relations, and long structured prompts. The method preserves image quality, adds modest inference overhead, and remains robust to imperfect scene programs.

## A Proxy-Based Method for Measuring Racial Disparities in~Text-to-Image~Generators

DevFeed: [A Proxy-Based Method for Measuring Racial Disparities in~Text-to-Image~Generators](<https://devfeed.tech/articles/a-proxy-based-method-for-measuring-racial-disparities-in-text-to-image-generators-120562.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.07132>)

Author: Nia Mucher, Ngan Pham, Jason Chin, Aaron Brown, David Evans

Published: 2026-11-06T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>), [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [Image](<https://devfeed.tech/topics/image.md>), [Ground truth / benchmark quality](<https://devfeed.tech/topics/ground-truth-benchmark-quality.md>), [stable-diffusion](<https://devfeed.tech/topics/stable-diffusion.md>), [flux](<https://devfeed.tech/topics/flux.md>)

Tags: [auditing](<https://devfeed.tech/tags/auditing.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [bias](<https://devfeed.tech/tags/bias.md>), [cosine-similarity](<https://devfeed.tech/tags/cosine-similarity.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [fairness](<https://devfeed.tech/tags/fairness.md>), [flux](<https://devfeed.tech/tags/flux.md>), [models](<https://devfeed.tech/tags/models.md>), [proxy-variables](<https://devfeed.tech/tags/proxy-variables.md>), [range](<https://devfeed.tech/tags/range.md>), [skin](<https://devfeed.tech/tags/skin.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>)

### AI overview

The paper proposes an automated audit of text-to-image generators that uses geographic, occupational, and surname proxies drawn from U.S. antidiscrimination law. It classifies generated images by race and compares representation with public demographic benchmarks, while CIELAB skin-tone measurements and image-group cosine similarity capture differences such as tone diversity and image similarity. In tests of Imagen 4 Fast, Stable Diffusion 3.5 Large Turbo, and Flux Schnell across 50 proxy variants per category, results varied by model and proxy. Imagen 4 Fast most overrepresented white-classified images and produced lighter, more homogeneous skin tones; it also showed the largest Black-representation disparity for surname and geographic prompts. Stable Diffusion 3.5 Large Turbo diverged most overall from census data but produced the most diverse skin tones. The abstract presents this as a legally grounded audit that does not depend on an implicit definition of fairness.

### Source excerpt

arXiv:2611.07132v1 Announce Type: cross Abstract: We present an automated method for auditing text-to-image generators for representational bias against ground-truth demographic benchmarks. Our approach adapts the racial proxy variables used in U.S. antidiscrimination law to provide a fine-grained comparative analysis of variations in racial representation in generated images. We classify each generated image by race and then compare the model's distribution to publicly available demographic benchmark data to identify patterns of over- and under-representation. We also measure skin tone using CIELAB values and calculate cosine similarity amongst groups of images in order to capture biases that don't show up in demographic comparisons such as differences in image quality across races and skin tones. We report on experiments testing three models (Imagen 4 Fast, Stable Diffusion 3.5 Large Turbo, and Flux Schnell) using three categories of proxies (geographic, occupational, and surname) and 50 variants within each category. We find notable variation across the models as well as by proxy category. For example, Imagen 4 Fast overrepresents white classified images more than the other models tested and produces images with lighter, more homogeneous skin tones. Imagen 4 Fast also had the largest representational disparity for Black demographics among surname and geographical prompts. Stable Diffusion 3.5 Large Turbo produces results that overall diverge most from census data but yield the most diverse range of skin tones. Our method yields a generalizable and legally-grounded auditing tool that surfaces detailed information about model outputs without relying on implicit normative definitions of fairness.

## Segment What It Can Draw: Querying a Text-to-Image Generator with Its Own Images for Skin Lesion Segmentation

DevFeed: [Segment What It Can Draw: Querying a Text-to-Image Generator with Its Own Images for Skin Lesion Segmentation](<https://devfeed.tech/articles/segment-what-it-can-draw-querying-a-text-to-image-generator-with-its-own-images-for-skin-lesion-segmentation-120154.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.07061>)

Author: Hongyu Cao, Guangfu Guo, Jinghan Zhang, Dakshak Keerthi Chandra, Ying Zhang, Kunpeng Liu

Published: 2026-11-06T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [diffusion-transformers](<https://devfeed.tech/topics/diffusion-transformers.md>), [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [Image](<https://devfeed.tech/topics/image.md>), [Image processing](<https://devfeed.tech/topics/image-processing.md>), [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>)

Tags: [annotation](<https://devfeed.tech/tags/annotation.md>), [annotations](<https://devfeed.tech/tags/annotations.md>), [attention](<https://devfeed.tech/tags/attention.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [model](<https://devfeed.tech/tags/model.md>), [patches](<https://devfeed.tech/tags/patches.md>), [pipeline](<https://devfeed.tech/tags/pipeline.md>), [prototypes](<https://devfeed.tech/tags/prototypes.md>), [regions](<https://devfeed.tech/tags/regions.md>), [rules](<https://devfeed.tech/tags/rules.md>), [skin](<https://devfeed.tech/tags/skin.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>), [training](<https://devfeed.tech/tags/training.md>), [transformer](<https://devfeed.tech/tags/transformer.md>), [zero-shot](<https://devfeed.tech/tags/zero-shot.md>)

### AI overview

GENIQ uses a text-to-image diffusion transformer's generated healthy-skin and lesion images to form initial prototypes, then refines them with the most skin-like and lesion-like patches from unlabelled dermoscopy images. The paper reports that this generated-image direction aligns better with real skin-versus-lesion differences than the direction derived from concept words. GENIQ scores test patches along the refined direction and uses a gated threshold to reduce contextual spill-over around small lesions. It requires no inference-time annotations or additional models and achieves Dice scores of 0.85-0.88 across three test sets, outperforming zero-shot and word-based attention methods on all three; it beats a hand-designed dermoscopy pipeline in Dice and IoU on two.

### Source excerpt

arXiv:2611.07061v1 Announce Type: cross Abstract: Skin lesion segmentation without annotated masks requires the target to be specified through other cues. Existing approaches rely on dermoscopy-specific rules, annotation-derived model priors, or user-selected regions. A general text-to-image diffusion transformer offers a text interface to the target, but its concept-word attention is unreliable on dermoscopy images: in the model's attention-output space, the word-derived skin-to-lesion direction is poorly aligned with the direction separating real skin from real lesions. We show that the direction derived from generated healthy-skin and lesion images is substantially better aligned. Based on this finding, we propose GENIQ, a training-free method that constructs initial skin and lesion prototypes from generated images and refines them using the most skin-like and lesion-like patches from unlabelled dermoscopy images. It scores test patches along the refined direction and applies a gated threshold to suppress contextual spill-over around small lesions. Without inference-time annotations or additional models, GENIQ achieves Dice scores of 0.85-0.88 on three dermoscopy test sets. It outperforms zero-shot and word-based attention methods on all three sets, and a hand-designed dermoscopy pipeline in Dice and IoU on two.

## Multi-Axis Max@K Reinforcement Learning for Representative Diversity in Text-to-Image Generation

DevFeed: [Multi-Axis Max@K Reinforcement Learning for Representative Diversity in Text-to-Image Generation](<https://devfeed.tech/articles/multi-axis-max-k-reinforcement-learning-for-representative-diversity-in-text-to-image-generation-116728.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2607.14962>)

Author: Ku Onoda, Paavo Parmas, Hiroki Furuta, Soichiro Nishimori, Yuta Oshima, Shohei Taniguchi, Yutaka Matsuo

Published: 2026-11-04T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [stable-diffusion](<https://devfeed.tech/topics/stable-diffusion.md>), [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>), [Generative AI](<https://devfeed.tech/topics/generative-ai.md>)

Tags: [colors](<https://devfeed.tech/tags/colors.md>), [coverage](<https://devfeed.tech/tags/coverage.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [diversity](<https://devfeed.tech/tags/diversity.md>), [fairness](<https://devfeed.tech/tags/fairness.md>), [image-generation](<https://devfeed.tech/tags/image-generation.md>), [learning](<https://devfeed.tech/tags/learning.md>), [quality](<https://devfeed.tech/tags/quality.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>)

### AI overview

MultiMaxPO trains diffusion text-to-image models to cover more predefined modes across a set of samples. Its multi-axis max@K objective takes the maximum score for each mode across samples and sums those maxima, rewarding a sample only when it raises a mode's maximum. The authors test this credit assignment on a synthetic Gaussian mixture and Stable Diffusion 3.5 Medium with pixel-based color rewards, then apply it to perceived-appearance fairness. On held-out prompts, three automatic evaluators show Fairness Score gains over the base model of 0.22-0.36 for Stable Diffusion 3.5 Medium and 0.07-0.19 for FLUX.1-dev. The abstract reports stronger results than matched RL baselines while maintaining image quality and text alignment.

### Source excerpt

arXiv:2607.14962v3 Announce Type: replace Abstract: Text-to-image (T2I) models produce realistic, prompt-aligned images, but samples for the same prompt often cover only a few visually distinct modes. For person-centric prompts, this lack of diversity can reflect or amplify demographic skew. We formalize representative diversity as target-mode coverage: how well a set of samples covers predefined modes, such as colors or perceived-race categories. We then propose MultiMaxPO, a reinforcement learning post-training method that improves this coverage in diffusion T2I models by optimizing a multi-axis max@K objective. Given one score per mode for each sample, multi-axis max@K takes the maximum over samples for each mode and then sums these maxima. A sample therefore receives credit for a mode only when it raises that mode's maximum, so different samples are rewarded for different modes. We verify this credit assignment on a synthetic Gaussian mixture and on Stable Diffusion 3.5 Medium with pixel-based color rewards, and then apply it to perceived-appearance fairness. On held-out prompts and under three automatic evaluators, MultiMaxPO improves the Fairness Score over the base model by 0.22-0.36 on Stable Diffusion 3.5 Medium and by 0.07-0.19 on FLUX.1-dev, outperforms matched RL baselines, and maintains image quality and text alignment. Code is available at https://github.com/KuOnoda/multi-axis-maxk.

## ODC: Orthogonal Drift Correction for Improved Text-to-Image Semantic Alignment at Inference

DevFeed: [ODC: Orthogonal Drift Correction for Improved Text-to-Image Semantic Alignment at Inference](<https://devfeed.tech/articles/odc-orthogonal-drift-correction-for-improved-text-to-image-semantic-alignment-at-inference-114453.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.00819>)

Author: Kasra Kamali, Basim Azam, Tuan Duong Trinh, Jean Honorio, Naveed Akhtar

Published: 2026-11-03T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [Embeddings](<https://devfeed.tech/topics/embeddings.md>), [Image](<https://devfeed.tech/topics/image.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [adaptive](<https://devfeed.tech/tags/adaptive.md>), [alignment](<https://devfeed.tech/tags/alignment.md>), [clip](<https://devfeed.tech/tags/clip.md>), [code](<https://devfeed.tech/tags/code.md>), [complex](<https://devfeed.tech/tags/complex.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [embedded](<https://devfeed.tech/tags/embedded.md>), [embedding](<https://devfeed.tech/tags/embedding.md>), [embeddings](<https://devfeed.tech/tags/embeddings.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [generation](<https://devfeed.tech/tags/generation.md>), [guidance](<https://devfeed.tech/tags/guidance.md>), [image](<https://devfeed.tech/tags/image.md>), [image-generation](<https://devfeed.tech/tags/image-generation.md>), [images](<https://devfeed.tech/tags/images.md>), [inference](<https://devfeed.tech/tags/inference.md>), [language](<https://devfeed.tech/tags/language.md>), [semantic](<https://devfeed.tech/tags/semantic.md>), [text](<https://devfeed.tech/tags/text.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>)

### AI overview

Orthogonal Drift Correction (ODC) aims to reduce semantic drift in text-to-image generation through inference-time guidance, without retraining or additional user input. It compares an initial image with the prompt in a shared vision-language embedding space, converts the drift component orthogonal to the prompt direction into a vocabulary-based text surrogate, then uses that representation with the original text embedding in an adaptive rank-reduced concept removal module to condition a second generation pass. The abstract reports CLIP score gains of up to 5.4% over baseline on COALBench-300 and PartiPrompts, and a 4.5% improvement in average 3-in-1 score over competing methods on T2I-CompBench's complex subset. These are reported benchmark results; the excerpt does not specify model-by-model performance.

### Source excerpt

arXiv:2611.00819v1 Announce Type: cross Abstract: Text-to-image models have achieved remarkable success in generating high-quality images from textual descriptions. However, they often struggle with "semantic drift," where the generated output fails to precisely align with complex or nuanced text prompts. While recent approaches have attempted to address semantic errors regarding attribute binding or object presence, there remains a gap for a more holistic method that addresses these issues by directly refining the text embeddings of the initial user prompt. In this work, we introduce Orthogonal Drift Correction (ODC), an inference-time guidance method designed to mitigate semantic drift without requiring model retraining or additional user inputs. ODC guides image generation through a two-stage process. In the first stage, it identifies the semantic drift by evaluating the initially generated image against the user prompt in a shared vision-language embedding space. It then isolates the component of this drift vector that is orthogonal to the prompt's direction and translates it back into text via a vocabulary-based surrogate mechanism. In the second stage, it produces refined text conditioning for a second generation pass by feeding both the initial text embedding and the re-embedded drift representation into an adaptive rank-reduced concept removal module. Our experiments demonstrate that ODC consistently improves prompt-image alignment, increasing CLIP scores by up to 5.4% over the baseline on the COALBench-300 and PartiPrompts benchmarks. Furthermore, on the T2I-CompBench complex subset, ODC improves the average 3-in-1 score by 4.5% over the competing methods. As a plug-and-play module, ODC offers a practical and efficient method for improving the reliability of text-to-image models. Code is available at https://github.com/kasrakamali/ODC.

## The Alignment Illusion: How BadReward Corrupts Text-to-Image Generation

DevFeed: [The Alignment Illusion: How BadReward Corrupts Text-to-Image Generation](<https://devfeed.tech/articles/the-alignment-illusion-how-badreward-corrupts-text-to-image-generation-113971.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2506.03234>)

Author: Kaiwen Duan, Hongwei Yao, Yufei Chen, Ziyun Li, Tong Qiao, Zhan Qin, Cong Wang

Published: 2026-11-03T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [rlhf](<https://devfeed.tech/topics/rlhf.md>), [Adversarial attacks](<https://devfeed.tech/topics/adversarial-attacks.md>), [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [ai alignment](<https://devfeed.tech/topics/ai-alignment.md>), [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>), [post-training](<https://devfeed.tech/topics/post-training.md>), [dpo](<https://devfeed.tech/topics/dpo.md>)

Tags: [alignment](<https://devfeed.tech/tags/alignment.md>), [attack](<https://devfeed.tech/tags/attack.md>), [attack-surface](<https://devfeed.tech/tags/attack-surface.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [code](<https://devfeed.tech/tags/code.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [data](<https://devfeed.tech/tags/data.md>), [feature](<https://devfeed.tech/tags/feature.md>), [feedback](<https://devfeed.tech/tags/feedback.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [generative-models](<https://devfeed.tech/tags/generative-models.md>), [human-feedback](<https://devfeed.tech/tags/human-feedback.md>), [image-generation](<https://devfeed.tech/tags/image-generation.md>), [learning](<https://devfeed.tech/tags/learning.md>), [malicious](<https://devfeed.tech/tags/malicious.md>), [manipulation](<https://devfeed.tech/tags/manipulation.md>), [models](<https://devfeed.tech/tags/models.md>), [post-training](<https://devfeed.tech/tags/post-training.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [reward](<https://devfeed.tech/tags/reward.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>)

### AI overview

BadReward poisons the RLHF stage of text-to-image training by manipulating reward signals rather than changing supervised fine-tuning data. It uses feature collisions to create visually benign preference pairs that align with a malicious target concept in reward-feature space. The abstract reports that online RLHF amplifies the corruption through reward-model feedback, while offline preference optimization learns directly from poisoned preferences. Experiments on modern architectures reportedly found high attack success and visual stealth across different RLHF algorithms; the excerpt provides no specific rates or model names.

### Source excerpt

arXiv:2506.03234v2 Announce Type: replace-cross Abstract: Modern text-to-image generative models increasingly rely on Reinforcement Learning from Human Feedback (RLHF) to align generated content with human preferences and safety standards. This post-training stage is critical for optimizing models toward high-fidelity and value-aligned behaviors using human preference data. However, while previous works have extensively investigated poisoning attacks against the supervised fine-tuning (SFT) stage of text-to-image training, the security vulnerabilities introduced by the RLHF process remain unexplored. In this paper, we present the first exploration into the feasibility of poisoning the RLHF process as a novel attack surface for compromising text-to-image models. We introduce BadReward, an indirect clean-label poisoning attack that manipulates reward signals to embed malicious behaviors. Our approach leverages a feature collision mechanism to produce poison preference image pairs that are visually benign, but aligned with a malicious target concept in the reward feature space. Our experiments demonstrate that this manipulation effectively corrupts the reward signal. Online RLHF further amplifies this corruption through its reward-model feedback loop, whereas offline preference optimization learns directly from the poisoned preferences. Finally, we demonstrate the real-world effectiveness of BadReward on modern text-to-image architectures, showing high attack success rates and visual stealthiness across different RLHF algorithms. Our code is available at https://github.com/ZJUICSR/BadReward.

## From Structural Feedback to Prompt Policies: Learning Faithful Text-to-Image Prompt Editors

DevFeed: [From Structural Feedback to Prompt Policies: Learning Faithful Text-to-Image Prompt Editors](<https://devfeed.tech/articles/from-structural-feedback-to-prompt-policies-learning-faithful-text-to-image-prompt-editors-103091.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.36146>)

Author: Wentao Ye, Yali Ye, Zhiqing Xiao, Ru Peng, Xinpeng Ti, Liyao Li, Zhanming Shen, Peng Lu, Jie Zhang, Haobo Wang

Published: 2026-10-28T04:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>), [Prompt Engineering](<https://devfeed.tech/topics/prompt-engineering.md>), [grpo](<https://devfeed.tech/topics/grpo.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>)

Tags: [attention](<https://devfeed.tech/tags/attention.md>), [automatic](<https://devfeed.tech/tags/automatic.md>), [comparisons](<https://devfeed.tech/tags/comparisons.md>), [complexity](<https://devfeed.tech/tags/complexity.md>), [confidence](<https://devfeed.tech/tags/confidence.md>), [control](<https://devfeed.tech/tags/control.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [editor](<https://devfeed.tech/tags/editor.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [feedback](<https://devfeed.tech/tags/feedback.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [generation](<https://devfeed.tech/tags/generation.md>), [global](<https://devfeed.tech/tags/global.md>), [graph](<https://devfeed.tech/tags/graph.md>), [image](<https://devfeed.tech/tags/image.md>), [learning](<https://devfeed.tech/tags/learning.md>), [policies](<https://devfeed.tech/tags/policies.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [text](<https://devfeed.tech/tags/text.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>)

### AI overview

StruPO aims to improve text-to-image results while preserving the requested scene. It conditions a prompt-editing policy on a calibrated scene graph and uses token-aligned deletions, reordering, and insertions to preserve structure while editing. Training accounts for uncertain grounding with smooth reward shaping and for stochastic generation by averaging multiple images per edit before computing group-relative advantages. Across three disjoint 2,000-prompt test sets and four frozen generation backbones, the abstract reports a better faithfulness-aesthetics trade-off and consistent relational-fidelity gains versus several prompt-optimization and generation-control approaches. Automatic evaluators and a 15-annotator study offer complementary evidence; reported human preference margins rise with relation complexity. After graph calibration, the policy generates optimized text in one pass without an iterative critique-rewrite loop.

### Source excerpt

arXiv:2610.36146v1 Announce Type: new Abstract: Prompt optimization should improve the image a user receives without changing the scene the user requested. Yet stylistic rewriting can weaken object attributes and relations even when global text--image similarity remains high. We propose StruPO, a structure-aware prompt editor that uses a calibrated scene graph as a shared representation for policy conditioning and image-grounded supervision. The graph exposes what should be preserved, while deletion, reordering, and insertion provide a token-aligned way to modify the prompt. This turns structural feedback from an external critique into a training signal for a reusable editing policy. Learning from that signal requires handling both uncertain grounding and stochastic image generation. StruPO retains grounding confidence through smooth reward shaping and averages multiple images for each edit before computing group-relative advantages. Experiments span three disjoint 2K-prompt test sets and four frozen generation backbones, with comparisons to learned prompt optimizers, iterative structural feedback, diffusion-model fine-tuning, and inference-time attention control. StruPO improves the faithfulness--aesthetics trade-off, with consistent gains in relational fidelity. Independent automatic evaluators and a 15-annotator study provide complementary evidence, and human preference margins increase with relation complexity. Controlled ablations connect these gains to graph conditioning, local editing, and image-grounded learning. The resulting policy produces reusable optimized text in one pass after graph calibration, without an iterative critique--rewrite loop.

## ACVA-Bench: Beyond Prompt Following in Abstract Concept Visualization Ability for Text-to-Image Models

DevFeed: [ACVA-Bench: Beyond Prompt Following in Abstract Concept Visualization Ability for Text-to-Image Models](<https://devfeed.tech/articles/acva-bench-beyond-prompt-following-in-abstract-concept-visualization-ability-for-text-to-image-models-99935.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.32058>)

Author: Mengyan Wang, Jiyuan An, Yining Liu, YaoWen Zhang, Liner Yang ang Erhong Yang

Published: 2026-10-26T04:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>)

Tags: [arxiv](<https://devfeed.tech/tags/arxiv.md>), [bench](<https://devfeed.tech/tags/bench.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [benchmarking](<https://devfeed.tech/tags/benchmarking.md>), [capacity](<https://devfeed.tech/tags/capacity.md>), [concepts](<https://devfeed.tech/tags/concepts.md>), [concrete](<https://devfeed.tech/tags/concrete.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [deep](<https://devfeed.tech/tags/deep.md>), [dimensions](<https://devfeed.tech/tags/dimensions.md>), [diversity](<https://devfeed.tech/tags/diversity.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [fixed](<https://devfeed.tech/tags/fixed.md>), [foundation](<https://devfeed.tech/tags/foundation.md>), [future](<https://devfeed.tech/tags/future.md>), [image](<https://devfeed.tech/tags/image.md>), [models](<https://devfeed.tech/tags/models.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [text](<https://devfeed.tech/tags/text.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>), [visualization](<https://devfeed.tech/tags/visualization.md>)

### AI overview

ACVA-Bench evaluates whether text-to-image models can represent abstract concepts without being given concrete objects, scenes, or visual metaphors. The category-balanced benchmark measures semantic sufficiency, representational diversity, and conceptual distinctiveness. Because abstract concepts lack a single reference image, its evaluation method breaks concept meaning into independently verifiable visual-semantic questions; the abstract says human evaluation confirmed the method's validity. Tests of mainstream models found reliance on a small set of prototypical visual patterns, with diversity and distinctiveness remaining areas for improvement.

### Source excerpt

arXiv:2610.32058v1 Announce Type: cross Abstract: Text-to-image (T2I) models have achieved remarkable progress in image quality and instruction following, prompting growing research interest in their capacity for deep understanding that goes beyond literal prompt matching. However, without an explicit visual scheme provided, it remains unclear and has not been systematically evaluated whether T2I models can autonomously translate abstract concepts lacking a fixed visual referent, such as "loyalty," into visual representations that are accurate, diverse, and distinguishable from closely related concepts. We define this capability as Abstract Concept Visualization Ability (ACVA) and introduce ACVA-Bench, a category-balanced benchmark grounded in psycholinguistic norms of abstract concepts, which examines models' ACVA along three dimensions semantic sufficiency, representational diversity, and conceptual distinctiveness under a low-intervention setting that withholds concrete objects, scenes, or visual metaphors. Since abstract concepts lack a unique reference image, conventional evaluation metrics that rely on reference-image matching are difficult to apply. To address this, we propose a fine-grained visual-semantic evaluation method constrained by concept semantics, which decomposes concept semantics into independently verifiable visual-semantic questions to assess the semantic representation quality of generated images, with its validity further confirmed by human evaluation. Benchmarking results on mainstream T2I models show that these models rely on a small set of prototypical visual patterns, leaving room for improvement in representational diversity and conceptual distinctiveness for abstract concepts, thereby providing an interpretable and reproducible evaluation foundation for future research.

## Stable Diffusion 3.5: Better Quality and Easier Prompts

DevFeed: [Stable Diffusion 3.5: Better Quality and Easier Prompts](<https://devfeed.tech/articles/stable-diffusion-3-5-better-quality-and-easier-prompts-94588.md>)

Original publisher: [Read original article](<https://www.runpod.io/blog/stable-diffusion-3-5-release-whats-new>)

Author: Brendan McKeag

Published: 2026-10-22T02:42:59Z

Content type: article

Language: en

Sources: [Runpod Blog.](<https://devfeed.tech/sources/runpod-blog.md>)

Topics: [stable-diffusion](<https://devfeed.tech/topics/stable-diffusion.md>), [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [Generative AI](<https://devfeed.tech/topics/generative-ai.md>), [Image](<https://devfeed.tech/topics/image.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [image-generation](<https://devfeed.tech/tags/image-generation.md>), [quality](<https://devfeed.tech/tags/quality.md>), [stable-diffusion](<https://devfeed.tech/tags/stable-diffusion.md>)

### AI overview

The excerpt reports that Stable Diffusion 3.5 Large prioritizes image quality, while Large Turbo favors efficiency; Medium was announced for release the following week for smaller GPUs. In initial testing, the author found the model could produce photorealistic people and animals with minimal prompting and images up to 2 megapixels, with fewer apparent anatomy issues than SD3. Results were not uniformly accurate: a prompt for a black-and-white Argentine tegu produced reptiles that did not look like that species. The excerpt presents SD3.5 as a new base model, so community checkpoints and fine-tunes were still forthcoming. It also describes a Runpod quick-start path and notes a Camenduru serverless image; the author says a Runpod Quick Deploy serverless image was not yet available.

### Source excerpt

Stable Diffusion 3.5 delivers a major quality leap, fixing past flaws while generating photorealistic images from minimal prompts. Learn what's new, how.

## Text-to-Image Generative Modeling Without Language Pretraining

DevFeed: [Text-to-Image Generative Modeling Without Language Pretraining](<https://devfeed.tech/articles/text-to-image-generative-modeling-without-language-pretraining-93680.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.26863>)

Author: Keshigeyan Chandrasegaran, Yue Zhao, Armin W. Thomas, Ehsan Adeli, Jiajun Wu, Juan Carlos Niebles, Li Fei-Fei, Michael Poli

Published: 2026-10-21T04:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Generative AI](<https://devfeed.tech/topics/generative-ai.md>)

Tags: [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [image](<https://devfeed.tech/tags/image.md>), [language](<https://devfeed.tech/tags/language.md>), [pretraining](<https://devfeed.tech/tags/pretraining.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>)

### AI overview

Halcyon shows that strong text-to-image generation can be trained from scratch without internet-scale language pretraining. The early-fusion autoregressive models use GPIC's 100 million image-text pairs: Halcyon-0.52B reports 48.0 gFD versus LlamaGen-XL/8's 63.3, using 18.2 times less pretraining compute; Halcyon-1.6B reports 35.3 versus JiT-T2I-XXL/8++'s 41.0, using 275 times less. Controlled experiments found text-only pretraining did not improve generation. Replacing caption dropout with image-to-text supervision also enabled captioning while maintaining competitive generation quality under the same budget. These findings are reported in the abstract.

### Source excerpt

arXiv:2610.26863v2 Announce Type: new Abstract: Modern text-to-image (T2I) systems use internet-scale language pretraining for text-conditioned image generation. We ask whether language pretraining is necessary for strong T2I generation. We introduce Halcyon, a family of early-fusion multimodal autoregressive models pretrained entirely from scratch on GPIC, a dataset of 100M image-text pairs. Halcyon-0.52B outperforms LlamaGen-XL/8 (48.0 vs. 63.3 gFD) with 18.2$\times$ less pretraining compute. Further, Halcyon-1.6B outperforms JiT-T2I-XXL/8++ (35.3 vs. 41.0 gFD) with 275$\times$ less pretraining compute. Controlled experiments show that text-only pretraining does not improve T2I generation. Halcyon also learns image captioning by replacing caption dropout with I2T supervision, maintaining competitive T2I generation quality under the same pretraining budget. We show that strong T2I systems can be trained from scratch without internet-scale language pretraining. Code and models are available at https://cs.stanford.edu/~keshik/halcyon

## BEM: Learning Bregman Divergences for Concept Erasure in Flow Matching Models

DevFeed: [BEM: Learning Bregman Divergences for Concept Erasure in Flow Matching Models](<https://devfeed.tech/articles/bem-learning-bregman-divergences-for-concept-erasure-in-flow-matching-models-93310.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.24508>)

Author: Mohamed-Hicham Leghettas, Markus P\\"uschel

Published: 2026-10-21T04:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Model Development](<https://devfeed.tech/topics/model-development.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [generalization in machine learning](<https://devfeed.tech/topics/generalization-in-machine-learning.md>), [flux](<https://devfeed.tech/topics/flux.md>)

Tags: [concepts](<https://devfeed.tech/tags/concepts.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [flow-matching](<https://devfeed.tech/tags/flow-matching.md>), [learning](<https://devfeed.tech/tags/learning.md>), [models](<https://devfeed.tech/tags/models.md>), [optimization](<https://devfeed.tech/tags/optimization.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>)

### AI overview

BEM learns a Bregman divergence through bi-level optimization to fine-tune flow-matching text-to-image models for concept erasure. The method targets nudity, violence, real identities and copyrighted characters, aiming to suppress concepts on unseen prompts while preserving unrelated generations. Under GEM's evaluation protocol, BEM removes more of the target concept than GEM in most settings: the reported unsafe rate for bloody gore falls from 53.9% to 14.4%, and nudity FID falls from 8.64 to 6.64. Retention holds for some concepts but degrades for others.

### Source excerpt

arXiv:2610.24508v1 Announce Type: new Abstract: Concept erasure removes harmful, private, or copyrighted concepts from text-to-image models while preserving their other abilities. Leading models such as FLUX and Stable Diffusion 3 are trained with flow matching, and erasure methods for them fine-tune the model toward a guided teacher's velocities under a Euclidean loss. Yet the squared distance is only one member of a larger family: Bregman divergences are exactly the losses under which flow matching recovers the correct velocity field, and each member shapes differently what is erased and what is preserved. We introduce BEM (Bregman erasure matching), which is, to our knowledge, the first to learn this divergence through bi-level optimization, so that the fine-tuned model forgets the concept on unseen prompts while its unrelated generations stay unchanged. In the experiments we target the suppression of nudity, violence, real identities, and copyrighted characters. Under the evaluation protocol of GEM, BEM removes more of the concept than GEM on most settings, cutting the unsafe rate on bloody gore from 53.9% to 14.4% and the FID on nudity from 8.64 to 6.64, while retention holds on some concepts and degrades on others.

## Beyond the Prompt: Recovering Hidden Concepts in Text-to-Image Models

DevFeed: [Beyond the Prompt: Recovering Hidden Concepts in Text-to-Image Models](<https://devfeed.tech/articles/beyond-the-prompt-recovering-hidden-concepts-in-text-to-image-models-93810.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.27473>)

Author: Zhengfeng Zheng, Minghong Fang

Published: 2026-10-21T04:00:00Z

Content type: article

Language: en

Sources: [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [Embeddings](<https://devfeed.tech/topics/embeddings.md>), [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>), [Image](<https://devfeed.tech/topics/image.md>), [Hyperparameter optimization](<https://devfeed.tech/topics/hyperparameter-optimization.md>)

Tags: [algorithm](<https://devfeed.tech/tags/algorithm.md>), [alignment](<https://devfeed.tech/tags/alignment.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [embedding](<https://devfeed.tech/tags/embedding.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>), [tokens](<https://devfeed.tech/tags/tokens.md>)

### AI overview

EmbGA recovers concepts present in generated images but omitted from their prompts by searching in token embedding space. It uses genetic search: crossover interpolates parent embeddings, mutation moves an embedding toward a randomly chosen vocabulary embedding, and each offspring is projected to discrete tokens before scoring. This avoids gradient-based methods' need for backward access to the scoring encoder and the one-inference-call-per-candidate cost of language-model-guided search; unlike discrete evolutionary search, its operations reflect embedding-space proximity. Across two benchmarks and two exploration budgets, EmbGA has the highest alignment at the larger budget, recovers the most grounded tokens on natural images, and runs much faster than language-model-guided search.

### Source excerpt

arXiv:2610.27473v1 Announce Type: new Abstract: Text-to-image models can generate visual concepts that their prompts never mention, and recent work shows that such associations can be implanted so that ordinary prompts trigger them. Auditing this behavior requires recovering the description that a prompt omits. Existing prompt optimizers are not well suited to this task. Gradient-based methods require backward access to the scoring encoder. Discrete evolutionary search treats semantically unrelated tokens as equally close. Language-model-guided search requires one inference call per candidate, which limits exploration. This paper presents EmbGA, an embedding-space genetic algorithm that recovers concepts present in a set of images but absent from the prompt that generated them. EmbGA keeps the population structure of genetic search but applies its operators in the token embedding space. Crossover interpolates between parent embeddings, and mutation moves an embedding toward a randomly drawn vocabulary embedding. Each offspring is then projected back to discrete tokens before scoring, so every evaluated candidate is a valid token sequence. We evaluate EmbGA on two benchmarks against six optimizers under two exploration budgets. EmbGA achieves the highest alignment when the exploration budget is large, recovers the most grounded tokens on natural images, and runs much faster than language-model-guided search.

## Attractive Faces Skew Text-to-Image Reward Models and Leaderboards

DevFeed: [Attractive Faces Skew Text-to-Image Reward Models and Leaderboards](<https://devfeed.tech/articles/attractive-faces-skew-text-to-image-reward-models-and-leaderboards-94388.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.28195>)

Author: Harvey Mannering, Zhiwu Huang, Adam Pr\\"ugel-Bennett

Published: 2026-10-21T04:00:00Z

Content type: news

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [text-to-image](<https://devfeed.tech/topics/text-to-image.md>), [ai-image-generator](<https://devfeed.tech/topics/ai-image-generator.md>), [Image](<https://devfeed.tech/topics/image.md>), [generative-ai-tools](<https://devfeed.tech/topics/generative-ai-tools.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [analysis](<https://devfeed.tech/tags/analysis.md>), [bias](<https://devfeed.tech/tags/bias.md>), [cognitive-bias](<https://devfeed.tech/tags/cognitive-bias.md>), [collection](<https://devfeed.tech/tags/collection.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [data-collection](<https://devfeed.tech/tags/data-collection.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [generation](<https://devfeed.tech/tags/generation.md>), [image](<https://devfeed.tech/tags/image.md>), [image-generation](<https://devfeed.tech/tags/image-generation.md>), [leaderboards](<https://devfeed.tech/tags/leaderboards.md>), [models](<https://devfeed.tech/tags/models.md>), [mps](<https://devfeed.tech/tags/mps.md>), [quality](<https://devfeed.tech/tags/quality.md>), [ranking](<https://devfeed.tech/tags/ranking.md>), [rankings](<https://devfeed.tech/tags/rankings.md>), [regression](<https://devfeed.tech/tags/regression.md>), [reward](<https://devfeed.tech/tags/reward.md>), [text](<https://devfeed.tech/tags/text.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>)

### AI overview

A controlled user study finds that perceived facial attractiveness can outweigh image quality in preference judgments, even when comparing photographs of identical fidelity or attractive AI-generated faces with unattractive real ones. Simulated Elo rankings show this bias can reverse a leaderboard. In tests editing 40 facial attributes, attractiveness was the strongest score correlate for four of five models (PickScore, MPS, HPSv2.1, HPSv3) and the strongest regression predictor. Preference-aligned generation also shifted faces toward younger, more feminine appearances and could override prompts for unattractive faces, limiting user control. The authors explore mitigations in data collection and model post-training; the excerpt does not report their effectiveness.

### Source excerpt

arXiv:2610.28195v1 Announce Type: cross Abstract: Human preference is now central to the evaluation, guidance, finetuning, and ranking of text-to-image models. However, the susceptibility of human preferences to bias in image generation remains underexplored. We identify a key observation: for human faces, perceived attractiveness overrides image quality in preference judgments, even when comparing real photographs of identical fidelity. In a controlled user study, annotators rank more attractive faces as higher quality, even when unattractive real faces are compared with attractive AI-generated ones. We examine the downstream consequences of this bias. Firstly, simulating Elo rankings from user study data shows attractiveness can invert a leaderboard's ranking, elevating synthetic attractive faces over real, unattractive ones. Secondly, the same bias reappears in the human preference models. Editing 40 annotated facial attributes in isolation, we find that attractiveness is the strongest correlate of score for four of five human preference models (PickScore, MPS, HPSv2.1, HPSv3), and the strongest predictor of score in regression analysis. We further show that preference aligned generation shifts faces toward younger, more feminine faces, and can override explicit prompts for unattractive faces, thereby limiting user control. Finally, we explore mitigation strategies, both for data collection and post training on human preference models. These findings highlight the need for preference models that disentangle image quality from human cognitive bias, enabling more reliable evaluation.

## Generate Images in Python With OpenAI's GPT Image API

DevFeed: [Generate Images in Python With OpenAI's GPT Image API](<https://devfeed.tech/articles/generate-images-in-python-with-openai-s-gpt-image-api-83958.md>)

Original publisher: [Read original article](<https://realpython.com/generate-images-openai/>)

Author: Martin Breuss

Published: 2026-10-09T14:00:00Z

Content type: tutorial

Language: en

Sources: [Real Python](<https://devfeed.tech/sources/real-python.md>)

Topics: [Python](<https://devfeed.tech/topics/python.md>), [ChatGPT API](<https://devfeed.tech/topics/chatgpt-api.md>), [Generative AI](<https://devfeed.tech/topics/generative-ai.md>), [text-to-image](<https://devfeed.tech/topics/text-to-image.md>)

Tags: [api](<https://devfeed.tech/tags/api.md>), [image](<https://devfeed.tech/tags/image.md>), [openai](<https://devfeed.tech/tags/openai.md>), [python](<https://devfeed.tech/tags/python.md>)

### AI overview

This tutorial demonstrates how to use OpenAI's GPT Image API from Python to generate and edit images. It covers setup, API authentication, Base64 decoding, saving response data and metadata, choosing image size and quality, and estimating token costs.

### Source excerpt

Learn how to generate images from text prompts with OpenAI's GPT Image API in Python, tune their size and quality, and edit them with a follow-up prompt.

[Next page](<https://devfeed.tech/topics/text-to-image.md?cursor=WyIyMDI2LTEwLTA5VDE0OjAwOjAwKzAwOjAwIiwgIjMzMDk1NmE5LTQ1YjctNDkxMC1iMmE3LTJmZjExNTdmZjkzMyJd>)