Reference-Guided Generation, Human Motion Synthesis, and Smarter Video Consistency Take Center Stage

ReferDiff: Reference-Image-Guided Diffusion for Precise Subject-Driven Generation

Subject-driven generation keeps getting sharper — this paper introduces a conditioning approach that binds reference images more tightly to output subjects without fine-tuning. Practically useful for product imagery and character consistency workflows.

MotionCrafter: Text-Driven Human Motion Video Synthesis with Skeleton-Aware Diffusion

Skeleton-aware conditioning lets MotionCrafter generate physically plausible human motion from text prompts alone, closing a key gap between text-to-video and animation pipelines. This is the kind of building block that game and film pre-vis studios will actually deploy.

FlowConsist: Optical Flow Supervision for Temporally Consistent Text-to-Video Diffusion

Optical flow as an explicit training signal for temporal consistency is an elegant fix to the flickering problem that plagues most open-source video diffusion models. If the gains hold up, this approach is likely to get absorbed into the next generation of video foundation models fast.

StyleTokens: Disentangled Style and Content Control in Text-to-Image Models via Learnable Token Embeddings

Separating style from content at the token level gives practitioners finer creative control without retraining the base model — a practical win for anyone doing brand-consistent or editorial image workflows. The learnable token approach also slots neatly into existing LoRA-based pipelines.

VideoScore++: A Holistic Evaluation Benchmark for Text-to-Video Quality Beyond FVD

FVD has been a notoriously poor proxy for perceived video quality, and VideoScore++ offers a multi-dimensional replacement that tracks semantic alignment, motion naturalness, and aesthetic coherence separately. Better benchmarks mean faster iteration — this is the kind of infrastructure work the field badly needs.

DenseDiffuse: Dense Semantic Correspondence for Training-Free Image Editing in Diffusion Models

Training-free editing methods that leverage dense semantic correspondences let you make precise local edits without any fine-tuning overhead — a significant practical advantage for fast iteration. This builds on recent attention-manipulation work but pushes spatial precision noticeably further.