The paper introduces Semantic Feedback Guidance (SFG) for multimodal diffusion transformers (MMDiTs), showing that image‑to‑text (I2T) feedback often preserves errors in the evolving image. By suppressing this feedback—either via an auxiliary branch (SFG) or an in‑place variant (iSFG)—the authors achieve improved alignment and higher preference‑model scores across multiple image‑generation backbones, including notable gains on FLUX and SD3.5M. The approach also extends to text‑to‑video and image‑to‑video generation, enhancing VBench scores.
arXiv:2606. 16682v3 Announce Type: replace Abstract: When AI agents use language models to evaluate their own outputs in a feedback loop, systematic biases emerge.
arXiv:2606.18974v3 Announce Type: replace Abstract: Unified multimodal models (UMMs) interleave generated ''visual thoughts'' (VTs) with text reasoning to improve spatial tasks. This incurs roughly a...
The paper investigates image tokenizers as the visual language of unified multimodal models by creating a controlled autoregressive testbed that tracks task‑specific validation losses during multimodal continual pretraining across text, image, text‑to‑image, and image‑to‑text predictions. It shows that losses must be analyzed by task, that the loss–performance relationship varies with the token space, and that better reconstruction does not always lead to stronger downstream performance. The study also demonstrates how tokenizer design choices—such as discriminator use, semantic supervision, and vocabulary size—affect joint modeling and downstream results.
The paper investigates how new concepts can be integrated into unified multimodal models (UMMs) by separating generation and understanding objectives through a novel visual entity bound to a single task direction. Experiments show that the effectiveness of cross‑task usability depends on where the concept is injected into the shared computation, with a mid‑stack alignment objective achieving high concept acquisition with minimal loss to overall performance. The study highlights that unified weights alone are insufficient; the two directions must share a semantic format at the entry point for efficient concept integration.
The paper introduces Diffusion On-Policy Context Distillation (D-OPCD), a method that internalizes the benefits of an agentic harness—such as memory, skills, and iterative prompt refinement—into the weights of a diffusion model. By treating the agent‑improved prompt as privileged context, D-OPCD distills harness knowledge into the model, raising direct‑generation scores from 60.52 to 65.09 across four benchmarks. This internalization allows the harness to shed saturated skills and continue evolving, enabling a second round of skill evolution that further improves performance by 1.83 points.
Subject-driven image personalization---generating new images that preserve the identity of one or several reference subjects in novel scenes---is a foundational capability for modern visual content creation. It is currently dominated by generalized methods that fine-tune a pretrained multimodal diffusion transformer (MMDiT) on hundreds of thousands to millions of paired \emph{(reference, composed-target)} examples, where each composed target is a synthesized image of the subject in a novel scene.
arXiv:2610.15227v1 Announce Type: new Abstract: Unified vision-language models (VLMs) perform both text-to-image (T2I) and image-to-text (I2T) generation with one shared Transformer backbone, yet how...
The paper investigates how visual interventions—training‑free edits applied to multimodal large language models—affect hallucinations and prediction quality. By applying a visual‑residual edit across successive decoder layers and matching its RMS strength, the authors identify an Evidence Anchoring Boundary (EAB) where the intervention’s influence drops sharply; effects before the EAB are 13.78–26.10× larger than after it. They further compare counterfactual subtraction across different question types and introduce calibrated selective routing, which only alters predictions when calibration supports an intervention, improving attribute accuracy by 1.56–1.81 percentage points while avoiding the harms of universal subtraction.
arXiv:2606. 17979v1 Announce Type: new Abstract: Existing RL post-training methods for text-to-image generation usually convert the final-image reward into a single scalar advantage and apply it with the same strength to the entire generative trajectory.
arXiv:2610.07887v1 Announce Type: cross Abstract: Unified multimodal models (UMMs) integrate understanding and generation, yet their generative behavior is rarely governed by what they understand abo...
arXiv:2606.04457v2 Announce Type: replace Abstract: Incorporating visual semantic representations as an intermediate step before image generation can reduce the modeling difficulty between text and i...
arXiv:2609. 31450v1 Announce Type: cross Abstract: Medical vision-language model (VLM) post-training is commonly evaluated through answer accuracy.