RefDecoder: Reference-Conditioned Video VAE Decoder for Enhanced Visual Generation
RefDecoder addresses an architectural asymmetry in latent diffusion models where denoising networks are heavily conditioned but decoders remain unconditional, causing detail loss and inconsistency. The approach injects high-fidelity reference image signals into the VAE decoding process via reference attention, with a lightweight image encoder mapping reference frames into high-dimensional tokens co-processed at each decoder up-sampling stage. Evaluated on Inter4K, WebVid, and Large Motion benchmarks, RefDecoder achieves up to +2.1dB PSNR over unconditional baselines and improves VBench I2V scores across subject consistency, background consistency, and overall quality. The module is plug-and-play, compatible with existing video generation systems including Wan 2.1 and VideoVAE+ without additional fine-tuning.
Related guides (2)
Related events (8)
AdaCodec: Predictive Visual Coding for Efficient Video MLLMs
AdaCodec introduces a predictive visual code interface for video multimodal large language models that exploits temporal redundancy in video. Instead of encoding every sampled frame as an independent RGB image, it sends full visual tokens only for reference frames with high conditional predictive cost, and encodes inter-frame changes as compact P-tokens. Evaluated against a Qwen3-VL-8B per-frame baseline across eleven benchmarks, AdaCodec at 1/7 the token budget (32k vs 224k tokens) surpasses the baseline on all long-video benchmarks while reducing time-to-first-token from 9.26s to 1.62s.
Remote VAEs for Decoding with Hugging Face Inference Endpoints
Hugging Face introduces Remote VAEs, a feature for Inference Endpoints that offloads the VAE decoding step of diffusion models to a separate remote service. This approach reduces GPU memory pressure on the primary inference host by decoupling the computationally expensive decoding stage. The pattern is relevant for large latent diffusion models where VAE decoding can be a significant memory and compute bottleneck.
Reroute: Training-free recoverable visual token routing for vision-language models
A new arXiv preprint proposes Reroute, a training-free plug-in that replaces the standard rank-and-remove visual token pruning paradigm in VLMs with a recoverable routing mechanism. Instead of permanently discarding low-ranked tokens, Reroute defers them to re-enter the candidate pool at later decoder stages, addressing the problem that token importance shifts across decoder depth. Evaluated on LLaVA-1.5 and Qwen backbones augmented with FastV, PDrop, and Nüwa pruning methods, Reroute improves grounding performance under aggressive token reduction without sacrificing general VQA accuracy. The approach preserves the theoretical compute and KV-cache budget of the underlying pruning method.
ASRD: Training-free anchor-guided revocable decoding for diffusion LLMs improves accuracy and throughput
A new arXiv preprint introduces ASRD (Anchor Supervised Revocable Decoding), a training-free framework for improving decoding quality in diffusion large language models. The method addresses error propagation and local error reinforcement in revocable decoding by separating trusted 'anchor tokens' (identified via temporal consistency) from uncertain candidates, then applying anchor-guided generation and anchor-perturbed verification. Experiments on math and coding benchmarks show up to 6.4% accuracy improvement and 7.2× inference throughput gains over remasking baselines.
RayDer: Scalable Self-Supervised Novel View Synthesis via Unified Feed-Forward Transformer
RayDer is a unified feed-forward transformer that consolidates camera estimation, scene reconstruction, and rendering into a single backbone for self-supervised novel view synthesis (NVS). By treating dynamic content as a nuisance factor absorbed by a minimal dynamic state, it enables stable training on unconstrained real-world video without requiring dynamic-scene reconstruction. The model exhibits clean power-law scaling with both data and compute across multiple model sizes, and achieves zero-shot open-set performance competitive with supervised state-of-the-art methods on multiple benchmarks.
VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion
VideoMLA applies Multi-Head Latent Attention (MLA) to causal video diffusion, replacing per-head keys and values with a shared low-rank content latent and decoupled 3D-RoPE positional key, achieving 92.7% reduction in per-token KV memory. The paper investigates why MLA works despite pretrained video attention not being low-rank (unlike the spectral assumption motivating MLA in LLMs), finding that the MLA bottleneck itself determines effective rank rather than the pretrained spectrum. On VBench, VideoMLA matches short-horizon baselines, achieves best overall score at long horizons, and delivers 1.23x throughput improvement on a single NVIDIA B200 GPU.
VEPO: Vision-anchored token selection improves RL for visual reasoning
A new arXiv paper identifies a failure mode of entropy-based credit assignment in multimodal reinforcement learning: vision-sensitive tokens with naturally low entropy are systematically ignored, causing the mechanism to collapse in visual reasoning tasks. The authors propose VEPO (Vision-Entropy token-selection for Policy Optimization), which couples visual sensitivity with token entropy via a multiplicative scheme to redirect gradient credit toward tokens that are both visually grounded and semantically informative. VEPO outperforms entropy-only baselines by 2.28 points at 7B scale and 3.15 points at 3B scale on visual reasoning benchmarks.
Squeezing Capacity from MLLMs for Subject-driven Image Generation via Dual Layer Aggregation
This paper proposes conditioning diffusion models on Multimodal Large Language Models (MLLMs) that jointly encode text and reference images, augmented with VAE-based identity conditioning to address copy-paste artifacts and identity preservation failures in subject-driven image generation. A Dual Layer Aggregation (DLA) module aggregates multi-level MLLM features, and a multi-stage denoising strategy progressively balances semantic and fine-detail identity signals during inference. Experiments show improved human preference scores on subject-driven generation benchmarks compared to prior approaches that encode text and reference images separately.