T2I Diffusion Models

In this paper, we uncover the hidden potential of Diffusion Transformers (DiTs) to significantly enhance generative tasks. Through an in-depth analysis of the denoising process, we demonstrate that introducing a single learned scaling parameter can significantly improve the performance of DiT blocks. Building on this insight, we propose Calibri, a parameter-efficient approach that optimally calibrates DiT components to elevate generative quality. Calibri frames DiT calibration as a black-box reward optimization problem, which is efficiently solved using an evolutionary algorithm and modifies just ~100 parameters. Experimental results reveal that despite its lightweight design, Calibri consistently improves performance across various text-to-image models. Notably, Calibri also reduces the inference steps required for image generation, all while maintaining high-quality outputs.

Diffusion models have demonstrated high-quality performance in conditional text-to-image generation, particularly with structural cues such as edges, layouts, and depth. However, lighting conditions have received limited attention and remain difficult to control within the generative process. Existing methods handle lighting through a two-stage pipeline that relights images after generation, which is inefficient. Moreover, they rely on fine-tuning with large datasets and heavy computation, limiting their adaptability to new models and tasks. To address this, we propose a novel Training-Free Light-Guided Text-to-Image Diffusion Model via Initial Noise Manipulation (LGTM), which manipulates the initial latent noise of the diffusion process to guide image generation with text prompts and user-specified light directions. Through a channel-wise analysis of the latent space, we find that selectively manipulating latent channels enables fine-grained lighting control without fine-tuning or modifying the pre-trained model. Extensive experiments show that our method surpasses prompt-based baselines in lighting consistency, while preserving image quality and text alignment. This approach introduces new possibilities for dynamic, user-guided light control. Furthermore, it integrates seamlessly with models like ControlNet, demonstrating adaptability across diverse scenarios.

Text-to-image models often struggle to synthesize correct occlusion relationships among multiple objects, especially in densely overlapping regions. Many training-free layout-guided methods enforce 2D spatial constraints but do not explicitly resolve depth-dependent attention competition, which can cause concept mixing and implausible occlusion. To address this problem, we propose DepthArb, a training-free framework that formulates occlusion generation as attention arbitration within a unified denoising trajectory. DepthArb employs two core occlusion-control mechanisms: Attention Arbitration Modulation suppresses background-object attention within foreground support according to relative depth, while Spatial Compactness Control limits attention dispersion to preserve object coherence. Because interference varies during generation, Occlusion Conflict Estimation constructs a shared spatial conflict field to adaptively weight both objectives. Through a unified spatial-text attention interface, DepthArb operates on U-Net cross-attention and the image-to-text component of MMDiT joint attention without model retraining. We further introduce OcclBench, a benchmark with continuous relative-depth specifications and occlusion-specific evaluation metrics. Experiments on OcclBench and public benchmarks show that DepthArb improves several layout and occlusion metrics over the evaluated baselines while maintaining competitive text-image alignment.

Recent research has shown that text-to-image diffusion models are capable of generating high-quality images guided by text prompts. But can they be used to generate or approximate real-world images from the seed noise? This is known as the diffusion inversion problem, which serves as a fundamental building block for bridging diffusion models and real-world scenarios. However, existing diffusion inversion methods often suffer from low reconstruction quality or weak robustness. Two major challenges need to be carefully addressed: (1) the misalignment between the inversion and generation trajectories during the diffusion process, and (2) the mismatch between the diffusion inversion process and the VQ autoencoder (VQAE) reconstruction. To address these challenges, we introduce a latent bias vector at each inversion step, which is learned to reduce the misalignment between inversion and generation trajectories. We refer to this strategy as Latent Bias Optimization (LBO). Furthermore, we perform an approximate joint optimization of the diffusion inversion and VQAE reconstruction processes by learning to adjust the image latent representation, which serves as the connecting interface between them. We refer to this technique as Image Latent Boosting (ILB). Extensive experimental results demonstrate that the proposed method significantly improves the image reconstruction quality of the diffusion model, as well as the performance of downstream tasks, including image editing and rare concept generation.

Recent diffusion-based models achieve photorealism in image inpainting but require many sampling steps, limiting practical use. Few-step text-to-image models offer faster generation, but naively applying them to inpainting yields poor harmonization and artifacts between the background and inpainted region. We trace this cause to random Gaussian noise initialization, which under low function evaluations causes semantic misalignment and reduced fidelity. To overcome this, we propose InverFill, a one-step inversion method tailored for inpainting that injects semantic information from the input masked image into the initial noise, enabling high-fidelity few-step inpainting. Instead of training inpainting models, InverFill leverages few-step text-to-image models in a blended sampling pipeline with semantically aligned noise as input, significantly improving vanilla blended sampling and even matching specialized inpainting models at low NFEs. Moreover, InverFill does not require real-image supervision and only adds minimal inference overhead. Extensive experiments show that InverFill consistently boosts baseline few-step models, improving image quality and text coherence without costly retraining or heavy iterative optimization.

Recent advances in text-to-image diffusion models have substantially improved the quality of image customization, enabling the synthesis of highly realistic images. Despite this progress, achieving fast and efficient personalization remains a key challenge, particularly for real-world applications. Existing approaches primarily accelerate customization for human subjects by injecting identity-specific embeddings into diffusion models, but these strategies do not generalize well to arbitrary object categories, limiting their applicability. To address this limitation, we propose a novel framework that employs a learned network to predict object-specific textual inversion embeddings, which are subsequently integrated into the UNet timesteps of a diffusion model for text-conditional customization. This design enables rapid, zero-shot personalization of a wide range of objects in a single forward pass, offering both flexibility and scalability. Extensive experiments across multiple tasks and settings demonstrate the effectiveness of our approach, highlighting its potential to support fast, versatile, and inclusive image customization. To the best of our knowledge, this work represents the first attempt to achieve such general-purpose, training-free personalization within diffusion models, paving the way for future research in personalized image generation.

Personalized text-to-image generation lets users fine-tune diffusion models into repositories of concept-specific checkpoints, but serving these repositories efficiently is difficult for two reasons: natural-language requests are often ambiguous and can be misrouted to visually similar checkpoints, and standard post-training quantization can distort the fragile representations that encode personalized concepts. We present PersonalQ, a unified framework that connects checkpoint selection and quantization through a shared signal -- the checkpoint's trigger token. Check-in performs intent-aligned selection by combining intent-aware hybrid retrieval with LLM-based reranking over checkpoint context and asks a brief clarification question only when multiple intents remain plausible; it then rewrites the prompt by inserting the selected checkpoint's canonical trigger. Complementing this, Trigger-Aware Quantization (TAQ) applies trigger-aware mixed precision in cross-attention, preserving trigger-conditioned key/value rows (and their attention weights) while aggressively quantizing the remaining pathways for memory-efficient inference. Experiments show that PersonalQ improves intent alignment over retrieval and reranking baselines, while TAQ consistently offers a stronger compression-quality trade-off than prior diffusion PTQ methods, enabling scalable serving of personalized checkpoints without sacrificing fidelity.

Text-to-Image (T2I) diffusion models have demonstrated strong generation ability, but their potential to generate unsafe content raises significant safety concerns. Existing inference-time defense methods typically perform category-agnostic token-level intervention in the text embedding space, which fails to capture malicious semantics distributed across the full token sequence and remains vulnerable to adversarial prompts. In this paper, we propose DTVI, a dual-stage inference-time defense framework for safe T2I generation. Unlike existing methods that intervene on specific token embeddings, our method introduces category-aware sequence-level intervention on the full prompt embedding to better capture distributed malicious semantics, and further attenuates the remaining unsafe influences during the visual generation stage. Experimental results on real-world unsafe prompts, adversarial prompts, and multiple harmful categories show that our method achieves effective and robust defense while preserving reasonable generation quality on benign prompts, obtaining an average Defense Success Rate (DSR) of 94.43% across sexual-category benchmarks and 88.56 across seven unsafe categories, while maintaining generation quality on benign prompts.

Text-to-image (T2I) models such as Stable Diffusion and DALLE remain susceptible to generating harmful or Not-Safe-For-Work (NSFW) content under jailbreak attacks despite deployed safety filters. Existing jailbreak attacks either rely on proxy-loss optimization instead of the true end-to-end objective, or depend on large-scale and costly RL-trained generators. Motivated by these limitations, we propose JANUS , a lightweight framework that formulates jailbreak as optimizing a structured prompt distribution under a black-box, end-to-end reward from the T2I system and its safety filters. JANUS replaces a high-capacity generator with a low-dimensional mixing policy over two semantically anchored prompt distributions, enabling efficient exploration while preserving the target semantics. On modern T2I models, we outperform state-of-the-art jailbreak methods, improving ASR-8 from 25.30% to 43.15% on Stable Diffusion 3.5 Large Turbo with consistently higher CLIP and NSFW scores. JANUS succeeds across both open-source and commercial models. These findings expose structural weaknesses in current T2I safety pipelines and motivate stronger, distribution-aware defenses. Warning: This paper contains model outputs that may be offensive.

Text-to-image diffusion models have achieved remarkable fidelity in synthesizing images from explicit text prompts, yet exhibit a critical deficiency in processing implicit prompts that require deep-level world knowledge, ranging from natural sciences to cultural commonsense, resulting in counter-factual synthesis. This paper traces the root of this limitation to a fundamental dislocation of the underlying knowledge structures, manifesting as a chaotic organization of implicit prompts compared to their explicit counterparts. In this paper, we propose EruDiff, which aims to refactor the knowledge within diffusion models. Specifically, we develop the Diffusion Knowledge Distribution Matching (DK-DM) to register the knowledge distribution of intractable implicit prompts with that of well-defined explicit anchors. Furthermore, to rectify the inherent biases in explicit prompt rendering, we employ the Negative-Only Reinforcement Learning (NO-RL) strategy for fine-grained correction. Rigorous empirical evaluations demonstrate that our method significantly enhances the performance of leading diffusion models, including FLUX and Qwen-Image, across both the scientific knowledge benchmark (i.e., Science-T2I) and the world knowledge benchmark (i.e., WISE), underscoring the effectiveness and generalizability. Our code is available at https://github.com/xiefan-guo/erudiff.

Diffusion Transformers (DiTs) have significantly enhanced text-to-image (T2I) generation quality, enabling high-quality personalized content creation. However, fine-tuning these models requires substantial computational complexity and memory, limiting practical deployment under resource constraints. To tackle these challenges, we propose a memory-efficient fine-tuning framework called DiT-BlockSkip, integrating timestep-aware dynamic patch sampling and block skipping by precomputing residual features. Our dynamic patch sampling strategy adjusts patch sizes based on the diffusion timestep, then resizes the cropped patches to a fixed lower resolution. This approach reduces forward & backward memory usage while allowing the model to capture global structures at higher timesteps and fine-grained details at lower timesteps. The block skipping mechanism selectively fine-tunes essential transformer blocks and precomputes residual features for the skipped blocks, significantly reducing training memory. To identify vital blocks for personalization, we introduce a block selection strategy based on cross-attention masking. Evaluations demonstrate that our approach achieves competitive personalization performance qualitatively and quantitatively, while reducing memory usage substantially, moving toward on-device feasibility (e.g., smartphones, IoT devices) for large-scale diffusion transformers.

Recent advancements in text-to-image synthesis have been largely propelled by diffusion-based models, yet achieving precise alignment between text prompts and generated images remains a persistent challenge. We find that this difficulty arises primarily from the limitations of conventional diffusion loss, which provides only implicit supervision for modeling fine-grained text-image correspondence. In this paper, we introduce Cross-Timestep Self-Calibration (CTCal), founded on the supporting observation that establishing accurate text-image alignment within diffusion models becomes progressively more difficult as the timestep increases. CTCal leverages the reliable text-image alignment (i.e., cross-attention maps) formed at smaller timesteps with less noise to calibrate the representation learning at larger timesteps with more noise, thereby providing explicit supervision during training. We further propose a timestep-aware adaptive weighting to achieve a harmonious integration of CTCal and diffusion loss. CTCal is model-agnostic and can be seamlessly integrated into existing text-to-image diffusion models, encompassing both diffusion-based (e.g., SD 2.1) and flow-based approaches (e.g., SD 3). Extensive experiments on T2I-Compbench++ and GenEval benchmarks demonstrate the effectiveness and generalizability of the proposed CTCal. Our code is available at https://github.com/xiefan-guo/ctcal.

The rapid growth of the text-to-image (T2I) community has fostered a thriving online ecosystem of expert models, which are variants of pretrained diffusion models specialized for diverse generative abilities. Yet, existing model merging methods remain limited in fully leveraging abundant online expert resources and still struggle to meet diverse in-the-wild user needs. We present DiffGraph, a novel agent-driven graph-based model merging framework, which automatically harnesses online experts and flexibly merges them for diverse user needs. Our DiffGraph constructs a scalable graph and organizes ever-expanding online experts within it through node registration and calibration. Then, DiffGraph dynamically activates specific subgraphs based on user needs, enabling flexible combinations of different experts to achieve user-desired generation. Extensive experiments show the efficacy of our method.

Text-to-image models are commercially valuable assets often distributed under restrictive licenses, but such licenses are enforceable only when violations can be detected. Existing methods require pre-deployment watermarking or internal model access, which are unavailable in commercial API deployments. We present Compositional Semantic Fingerprinting (CSF), the first black-box method for attributing fine-tuned text-to-image models to protected lineages using only query access. CSF treats models as semantic category generators and probes them with compositional underspecified prompts that remain rare under fine-tuning. This gives IP owners an asymmetric advantage: new prompt compositions can be generated after deployment, while attackers must anticipate and suppress a much broader space of fingerprints. Across 6 model families (FLUX, Kandinsky, SD1.5/2.1/3.0/XL) and 13 fine-tuned variants, our Bayesian attribution framework enables controlled-risk lineage decisions, with all variants satisfying the dominance criterion.

Text-to-image diffusion models achieve high visual fidelity but surprisingly exhibit systematic failures in numerical control when prompts specify explicit object counts. To address this limitation, we introduce ATHENA, a model-agnostic, test-time adaptive steering framework that improves object count fidelity without modifying model architectures or requiring retraining. ATHENA leverages intermediate representations during sampling to estimate object counts and applies count-aware noise corrections early in the denoising process, steering the generation trajectory before structural errors become difficult to revise. We present three progressively more advanced variants of ATHENA that trade additional computation for improved numerical accuracy, ranging from static prompt-based steering to dynamically adjusted count-aware control. Experiments on established benchmarks and a new visually and semantically complex dataset show that ATHENA consistently improves count fidelity, particularly at higher target counts, while maintaining favorable accuracy-runtime trade-offs across multiple diffusion backbones.

Diffusion models have become the dominant paradigm for image generation and editing, with latent diffusion models shifting denoising to a compact latent space for efficiency and scalability. Recent attempts to leverage pretrained visual representation models as tokenizer priors either align diffusion features to representation features or directly reuse representation encoders as frozen tokenizers. Although such approaches can improve generation metrics, they often suffer from limited reconstruction fidelity due to frozen encoders, which in turn degrades editing quality, as well as overly high-dimensional latents that make diffusion modeling difficult. To address these limitations, We propose Representation-Pivoted AutoEncoder, a representation-based tokenizer that improves both generation and editing. We introduce Representation-Pivot Regularization, a training strategy that enables a representation-initialized encoder to be fine-tuned for reconstruction while preserving the semantic structure of the pretrained representation space, followed by a variational bridge which compress latent space into a compact one for better diffusion modeling. We adopt an objective-decoupled stage-wise training strategy that sequentially optimizes generative tractability and reconstruction-fidelity objectives. Together, these components yield a tokenizer that preserves strong semantics, reconstructs faithfully, and produces latents with reduced diffusion modeling complexity. Experiments demonstrate that RPiAE outperforms other visual tokenizers on text-to-image generation and image editing, while delivering the best reconstruction fidelity among representation-based tokenizers.

Diffusion-based text-to-image (T2I) models have made remarkable progress in generating photorealistic and semantically rich images. However, when the target concepts lie in low-density regions of the training distribution, these models often produce semantically misaligned or structurally inconsistent results. This limitation arises from the long-tailed nature of text-image datasets, where rare concepts or editing instructions are underrepresented. To address this, we introduce Adaptive Auxiliary Prompt Blending (AAPB) - a unified framework that stabilizes the diffusion process in low-density regions. AAPB leverages auxiliary anchor prompts to provide semantic support in rare concept generation and structural support in image editing, ensuring faithful guidance toward the target prompt. Unlike prior heuristic prompt alternation methods, AAPB derives a closed-form adaptive coefficient that optimally balances the influence between the auxiliary anchor and the target prompt at each diffusion step. Grounded in Tweedie's identity, our formulation provides a principled and training-free framework for adaptive prompt blending, ensuring stable and target-faithful generation. We demonstrate the effectiveness of adaptive interpolation over fixed interpolation through controlled experiments and empirically show consistent improvements on the RareBench and FlowEdit datasets, achieving superior semantic accuracy and structural fidelity compared to prior training-free baselines.

Concept unlearning has emerged as a promising direction for reducing the risks of harmful content generation in text-to-image diffusion models by selectively erasing undesirable concepts from a model's parameters. Existing approaches typically rely on keywords to identify the target concept to be unlearned. However, we show that this keyword-based formulation is inherently limited: a visual concept is multi-dimensional, can be expressed in diverse textual forms, and often overlap with related concepts in the latent space, making keyword-only unlearning, which imprecisely indicate the target concept is brittle and prone to over-forgetting. This occurs because a single keyword represents only a narrow point estimate of the concept, failing to cover its full semantic distribution and entangled variations in the latent space. To address this limitation, we propose Diversified Unlearning, a distributional framework that represents a concept through a set of contextually diverse prompts rather than a single keyword. This richer representation enables more precise and robust unlearning. Through extensive experiments across multiple benchmarks and state-of-the-art baselines, we demonstrate that integrating Diversified Unlearning as an add-on component into existing unlearning pipelines consistently achieves stronger erasure, better retention of unrelated concepts, and improved robustness against adversarial recovery attacks.

Precise Text-to-Image (T2I) generation has achieved great success but is hindered by the limited relational reasoning of static text encoders and the error accumulation in open-loop sampling. Without real-time feedback, initial semantic ambiguities during the Ordinary Differential Equation trajectory inevitably escalate into stochastic deviations from spatial constraints. To bridge this gap, we introduce AFS-Search (Agentic Flow Steering and Parallel Rollout Search), a training-free closed-loop framework built upon FLUX.1-dev. AFS-Search incorporates a training-free closed-loop parallel rollout search and flow steering mechanism, which leverages a Vision-Language Model (VLM) as a semantic critic to diagnose intermediate latents and dynamically steer the velocity field via precise spatial grounding. Complementarily, we formulate T2I generation as a sequential decision-making process, exploring multiple trajectories through lookahead simulations and selecting the optimal path based on VLM-guided rewards. Further, we provide AFS-Search-Pro for higher performance and AFS-Search-Fast for quicker generation. Experimental results show that our AFS-Search-Pro greatly boosts the performance of the original FLUX.1-dev, achieving state-of-the-art results across three different benchmarks. Meanwhile, AFS-Search-Fast also significantly enhances performance while maintaining fast generation speed.

Text-to-image models produce images that align well with natural language prompts, but compositional generation has long been a central challenge. Models often struggle to satisfy multiple concepts within a single prompt, frequently omitting some concepts and resulting in partial success. Such failures highlight the difficulty of jointly optimizing multiple concepts during reward optimization, where competing concepts can interfere with one another. To address this limitation, we propose Correlation-Weighted Multi-Reward Optimization (CMO), a framework that leverages the correlation structure among concept rewards to adaptively weight each attribute concept in optimization. By accounting for interactions among concepts, CMO balances competing reward signals and emphasizes concepts that are partially satisfied yet inconsistently generated across samples, improving compositional generation. Specifically, we decompose multi-concept prompts into pre-defined concept groups (e.g., objects, attributes, and relations) and obtain reward signals from dedicated reward models for each concept. We then adaptively reweight these rewards, assigning higher weights to conflicting or hard-to-satisfy concepts using correlation-based difficulty estimation. By focusing optimization on the most challenging concepts within each group, CMO encourages the model to consistently satisfy all requested attributes simultaneously. We apply our approach to train state-of-the-art diffusion models, SD3.5 and FLUX.1-dev, and demonstrate consistent improvements on challenging multi-concept benchmarks, including ConceptMix, GenEval 2, and T2I-CompBench. The code is available at https://github.com/TheDarkKnight-21th/CMO.

Although text-to-image diffusion models exhibit remarkable generative power, concept erasure techniques are essential for their safe deployment to prevent the creation of harmful content. This has fostered a dynamic interplay between the development of erasure defenses and the adversarial probes designed to bypass them, and this co-evolution has progressively enhanced the efficacy of erasure methods. However, this adversarial co-evolution has converged on a narrow, text-centric paradigm that equates erasure with severing the text-to-image mapping, ignoring that the underlying visual knowledge related to undesired concepts still persist. To substantiate this claim, we investigate from a visual perspective, leveraging DDIM inversion to probe whether a generative pathway for the erased concept can still be found. However, identifying such a visual generative pathway is challenging because standard text-guided DDIM inversion is actively resisted by text-centric defenses within the erased model. To address this, we introduce TINA, a novel Text-free INversion Attack, which enforces this visual-only probe by operating under a null-text condition, thereby avoiding existing text-centric defenses. Moreover, TINA integrates an optimization procedure to overcome the accumulating approximation errors that arise when standard inversion operates without its usual textual guidance. Our experiments demonstrate that TINA regenerates erased concepts from models treated with state-of-the-art unlearning. The success of TINA proves that current methods merely obscure concepts, highlighting an urgent need for paradigms that operate directly on internal visual knowledge.

Post-training quantization (PTQ) is a practical path to deploy large diffusion models, but quantization noise can accumulate over the denoising trajectory and degrade generation quality. We propose Q-Drift, a sampler-side correction that aims to preserve the intended sampling marginals through a deterministic drift adjustment motivated by generalized marginal-preserving SDEs. Q-Drift uses the calibrated conditional residual variance of quantization error to determine a correction factor at each step. Our SDXL study shows that calibration with as few as 10 paired full-precision/quantized runs remains effective. The resulting sampler correction is plug-and-play with common samplers, diffusion models, and PTQ methods, while incurring negligible overhead at inference. Across six diverse text-to-image models (spanning DiT and U-Net), three samplers (Euler, flow-matching, DPM-Solver++), and two PTQ methods (SVDQuant, MixDQ), Q-Drift improves FID over the corresponding quantized baseline in all seven settings of our main evaluation, with up to 4.79 FID reduction on PixArt-Sigma (SVDQuant W3A4), while preserving CLIP scores. Code is available at https://github.com/sooyoung-ryu/Q-Drift.

Users often possess a clear visual intent but struggle to articulate it precisely in language. This intention-expression gap makes aligning generated images with latent visual preferences a fundamental challenge in text-to-image diffusion models. Existing methods either require model training, sacrificing flexibility, or rely on textual feedback, imposing a heavy cognitive burden. Although recent training-free methods use click-based binary preference feedback to reduce user effort, they force Foundation Models (FMs) to infer preferences at the semantic level. When faced with multi-dimensional preferences, FMs suffer from inference overload and fail to identify exact preferred feature values under conflicting user signals. Consequently, a flexible framework for multi-dimensional feature alignment remains absent. To address this, we propose a Hierarchical Relevance Feedback-Driven (HRFD) framework. Recognizing that multiple features struggle to converge simultaneously, HRFD organizes them into a three-tier hierarchy and adapts relevance feedback to enforce coarse-to-fine convergence, minimizing cognitive load. To bypass FM inference overload, HRFD decouples the process into independent single-feature preference inference tasks. Furthermore, to overcome FMs' failure in identifying preferred values, HRFD employs statistical inference to quantify the distribution divergence of features between "liked" and "disliked" image sets, achieving robust and transparent preference measurement. Crucially, HRFD operates entirely within the external text space, remaining strictly training-free and model-agnostic. Extensive experiments demonstrate that HRFD effectively captures the user's true visual intent, significantly outperforming baseline approaches.

Optimizing the noise samples of diffusion and flow models is an increasingly popular approach to align these models to target rewards at inference time. However, we observe that these approaches are usually restricted to differentiable or cheap reward models, the formulation of the underlying pretrained generative model, or are memory/compute inefficient. We instead propose a simple trust-region based search algorithm (TRS) which treats the pre-trained generative and reward models as a black-box and only optimizes the source noise. Our approach achieves a good balance between global exploration and local exploitation, and is versatile and easily adaptable to various generative settings and reward models with minimal hyperparameter tuning. We evaluate TRS across text-to-image, molecule and protein design tasks, and obtain significantly improved output samples over the base generative models and other inference-time alignment approaches which optimize the source noise sample, or even the entire reverse-time sampling noise trajectories in the case of diffusion models. Our source code is publicly available.

Diffusion and flow models achieve State-Of-The-Art (SOTA) generative performance, yet many practically important behaviors such as fine-grained prompt fidelity, compositional correctness, and text rendering are weakly specified by score or flow matching pretraining objectives. Reinforcement Learning (RL) fine-tuning with external, black-box rewards is a natural remedy, but diffusion RL is often brittle. Trajectory-based methods incur high memory cost and high-variance gradient estimates; forward-process approaches converge faster but can suffer from distribution drift, and hence reward hacking. In this work, we present Centered Reward Distillation (CRD), a diffusion RL framework derived from KL-regularized reward maximization built on forward-process-based fine-tuning. The key insight is that the intractable normalizing constant cancels under within-prompt centering, yielding a well-posed reward-matching objective. To enable reliable text-to-image fine-tuning, we introduce techniques that explicitly control distribution drift: (i) decoupling the sampler from the moving reference to prevent ratio-signal collapse, (ii) KL anchoring to a CFG-guided pretrained model to control long-run drift and align with the inference-time semantics of the pre-trained model, and (iii) reward-adaptive KL strength to accelerate early learning under large KL regularization while reducing late-stage exploitation of reward-model loopholes. Experiments on text-to-image post-training with GenEval and OCR rewards show that CRD achieves competitive SOTA reward optimization results with fast convergence and reduced reward hacking, as validated on unseen preference metrics.

As modern text-to-image (T2I) models draw closer to synthesizing highly realistic content, the threat of unsafe content generation grows, and it becomes paramount to exercise control. Existing approaches steer these models by applying Euclidean adjustments to text embeddings, redirecting the generation away from unsafe concepts. In this work, we introduce hyperbolic control (HyCon): a novel control mechanism based on parallel transport that leverages semantically aligned hyperbolic representation space to yield more expressive and stable manipulation of concepts. HyCon reuses off-the-shelf generative models and a state-of-the-art hyperbolic text encoder, linked via a lightweight adapter. HyCon achieves state-of-the-art results across four safety benchmarks and four T2I backbones, showing that hyperbolic steering is a practical and flexible approach for more reliable T2I generation.

In this paper, we explore text-guided image editing in the remote sensing domain using generative modeling. We propose \rsedit, a collection of models from U-Net to DiT with various configurations. Specifically, we present the first comprehensive study of conditioning strategies for building image editing models from off-the-shelf text-to-image ones. Our experiments show that \rsedit achieves the best instruction-faithful edits while preserving geospatial structure. We release the code at https://github.com/Bili-Sakura/RSEdit-Preview and checkpoints at https://huggingface.co/collections/BiliSakura/rsedit.

Text-to-image diffusion models excel at generating images from natural language descriptions, yet fail to interpret numerical colors such as hex codes (#FF5733) and RGB values (rgb(255,87,51)). This limitation stems from subword tokenization, which fragments color codes into semantically meaningless tokens that text encoders cannot map to coherent color representations. We present NumColor, that enables precise numerical color control across multiple diffusion architectures. NumColor comprises two components: a Color Token Aggregator that detects color specifications regardless of tokenization, and a ColorBook containing 6,707 learnable embeddings that map colors to embedding space of text encoder in perceptually uniform CIE Lab space. We introduce two auxiliary losses, directional alignment and interpolation consistency, to enforce geometric correspondence between Lab and embedding spaces, enabling smooth color interpolation. To train the ColorBook, we construct NumColor-Data, a synthetic dataset of 500K rendered images with unambiguous color-to-pixel correspondence, eliminating the annotation ambiguity inherent in photographic datasets. Although trained solely on FLUX, NumColor transfers zero-shot to SD3, SD3.5, PixArt-α, and PixArt-Σ without model-specific adaptation. NumColor improves numerical color accuracy by 4-9x across five models, while simultaneously improving color harmony scores by 10-30x on GenColorBench benchmark.

State-of-the-art text-to-image diffusion models can produce impressive visuals but may memorize and reproduce training images, creating copyright and privacy risks. Existing prompt perturbations applied at inference time, such as random token insertion or embedding noise, may lower copying but often harm image-prompt alignment and overall fidelity. To address this, we introduce two complementary methods. First, Region-Aware Prompt Augmentation (RAPTA) uses an object detector to find salient regions and turn them into semantically grounded prompt variants, which are randomly sampled during training to increase diversity, while maintaining semantic alignment. Second, Attention-Driven Multimodal Copy Detection (ADMCD) aggregates local patch, global semantic, and texture cues with a lightweight transformer to produce a fused representation, and applies simple thresholded decision rules to detect copying without training with large annotated datasets. Experiments show that RAPTA reduces overfitting while maintaining high synthesis quality, and that ADMCD reliably detects copying, outperforming single-modal metrics.

Reinforcement learning (RL) has become a standard technique for post-training diffusion-based image synthesis models, as it enables learning from reward signals to explicitly improve desirable aspects such as image quality and prompt alignment. In this paper, we propose an online RL variant that reduces the variance in the model updates by sampling paired trajectories and pulling the flow velocity in the direction of the more favorable image. Unlike existing methods that treat each sampling step as a separate policy action, we consider the entire sampling process as a single action. We experiment with both high-quality vision language models and off-the-shelf quality metrics for rewards, and evaluate the outputs using a broad set of metrics. Our method converges faster and yields higher output quality and prompt alignment than previous approaches.