Code2Video-Compass is a general benchmark for code video generation, covering three settings: T2V, I2V and V2V, and containing 2,104 tasks across 10 major categories and 24 subcategories.
RLSplat proposes a feed-forward 3D Gaussian Splatting framework with controllable resource-linear budgets, where representation scale, inference latency and peak memory decrease linearly as the Gaussian budget decreases.
WorldStream is an open-ended 3D world model that can extend playable terrain beyond preset maps, respond to open-ended language instructions in a running world, and restore previous edits when revisiting old locations. It uses a hierarchical planner to output structured decision states: Macro fixes the terrain and hydrology structure of a group of tiles, Micro refines it into tile semantics, and deterministic local code then compiles it into terrain, roads, settlements, vegetation and entities. A single world record stores plans, seeds, object identities, mutable states and edits, supporting continuous expansion and interaction under bounded budgets.
StyleStream 2.0 proposes a fast and controllable streaming voice style conversion framework, reducing end-to-end latency from about 1s in version 1.0 to 520ms on consumer-grade GPUs, achieving 2x speedup while maintaining conversion quality.
StreamEval is a unified benchmark for interactive streaming video generation covering five task categories: virtual human dialogue, virtual human performance, world generation, instruction editing and reference editing. It contains 300 scripted test cases lasting 15-300s, and evaluates on a 0-100 scale across four dimensions: task execution, state consistency, visual fidelity and audio-visual coordination.
PhysSkin++ proposes a physics-informed framework that learns a continuous neural skinning field from static geometry, constructs a compact deformation subspace, and generalizes across shapes without per-object network training. Its neural skinning field autoencoder, combined with a physics-informed self-supervised learning strategy, stably jointly optimizes without simulation trajectories, and supports direct animation rigging of meshes, point clouds, and 3D Gaussians as well as real-time reduced-dimension simulation. Experiments validate applications including cross-shape generalization, real-time simulation, and controllable collision-aware animation.
VideoMaMa++ proposes a mask-guided diffusion video matting framework that achieves sliding-window inference through two learnable embeddings, preserve token and refine token, propagates generated mattes across chunk boundaries, and uses a globally shared reference anchor to prevent error accumulation over long sequences, without additional training.
PVBench is a benchmark for quantitatively evaluating the physical plausibility of generated videos, covering 46 dimensions across 7 major domains and containing 600 carefully designed prompts, 30% of which involve embodied agents such as robotic arms, humanoid robots, and quadruped robots.
SEAM is a zero-audio initialization method that builds embeddings for unseen tokens using known subwords, preserves temporal phoneme order through spectral filtering, and aligns new embeddings to the pretrained conditioning manifold using closed-form ridge regression. Across three languages and two TTS architectures, SEAM approaches or matches baselines with up to 4x less training data, reducing sample complexity by 75% compared with random initialization and mean pooling. The work has been accepted to EMNLP 2026 Findings.
The paper proposes StreamingSwap, a reference-guided flow-matching framework for real-time streaming video face swapping. Methodologically, it introduces Joint Rec-ID Loss, which balances explicit identity supervision and implicit reconstruction according to noise level, and reduces spatial tokens and temporal context through Face-RoI Generation and Short-Context Rollout Training.
SWIFT is a self-bootstrapping training method that extrapolates clean regression targets using a frozen weak denoiser, improving stronger diffusion models without adding inference overhead. On ImageNet-256, it surpasses the baseline best FID with about half the total training compute; on RAEv2, it achieves an FID of 1.15 with unguided inference and 0.99 with guidance.
SpecKCast is a training-free method that predicts diffusion model outputs in the frequency domain to skip redundant denoising steps: it applies a 2D DCT to the output, then uses a lightweight Kalman filter to track each coefficient's value and rate of change along band-time coordinates derived from the noise schedule, invoking the transformer only when prediction uncertainty exceeds a threshold.
Zhongguang Tianze's Digital Intelligence Center is shooting with 6 synchronized cameras and producing a 4D continuous multi-view video dataset at the scale of ten thousand hours every two months; its clients are leading domestic model vendors, telecom operators and e-commerce companies, and its semi-annual report already mentions China Telecom and StepFun. The business emphasizes self-collected and self-labeled data, security and compliance, and out-of-the-box usability, and it has developed its own multi-view calibration and acoustic datasets; customized projects account for a high proportion, and payment is made only after delivery and acceptance pass.
An academic team proposed Brain-IT, which uses the Brain Interaction Transformer (BIT) to scan brain voxels via fMRI and reconstruct visual images through a diffusion process, surpassing previous models such as MindTurner, MindEye2, and NeuroVLA in accuracy with only one hour of training.
MT Forcing is a multi-turn image editing training paradigm for unified multimodal models (UMMs), based on autoregressive diffusion, aimed at solving the drift problems of unstable visual quality and inconsistent content after multi-turn editing.
Researchers propose Painting Change Map, a spatiotemporal control mechanism that explicitly tracks normalized structural changes over 1,000 timesteps and can guide both forward painting and reverse "unpainting" processes. The method encodes local visual differences into RGB space to achieve stroke-level fine-grained control, outperforms existing control baselines on benchmarks, and significantly surpasses all baselines under oracle conditions. The code is open-sourced.
HumanSkill-4D proposes a training-free framework that uses a shared multimodal large language model and geometric operators to select local regions of an animatable Gaussian avatar from language queries, and can specify a time period matching a pose or action. The method binds body parts and clothing categories to persistent Gaussian regions via semantic memory, and composes operator programs using driving joint trajectories in motion memory, making selections traceable. The team built three human-annotated benchmarks, on which it outperforms baselines in part segmentation and 3D and 4D referring segmentation, with only a few minutes of setup per avatar.
UniAVC uses a single set of model weights to achieve text-conditioned joint generation, bidirectional cross-modal generation, and instruction-guided editing without task-specific adapters. It first trains each branch on unimodal data to full generation and editing capability, then uses paired data for cross-modal correspondence; a modality-decoupled scheduler independently transforms noise schedules from a shared timestep, and directly supervises the asymmetric configuration with unimodal clean samples. This enables intent-aware editing and post-inference self-refinement, regenerating only the unsatisfactory side.
LayerEditor is a unified framework that couples instruction-guided image editing with reusable RGBA layer generation. Its core is In-Context RGB-Alpha Token Modeling, which jointly represents the editing canvas and ordered RGBA layers in a single token sequence.
Tencent Photon and NUS propose the Orbia video world model evaluation benchmark, using 720 exploration and revisit cases to test 27 models, checking retention of known scenes, persistence of newly generated content, and cross-view 3D consistency. SolarWM-H3 ranks first overall with 76.32, followed by Lyra2 and EVOKE in second and third place; content persistence in long video generation is generally lower than in short videos, and the average score for forward-backward trajectories in the world model group drops from 66.21 to 40.96.
Gen4P proposes a unified framework that can turn a pretrained video generator into a feed-forward 4D perceiver; without iterative denoising, a single encoder forward pass can process clean video latents and predict the required 4D attributes. Its on-demand querying supports flexible spatial sampling, continuous-time queries provide temporal flexibility, and a pretrained video-depth LoRA can be plugged in as a frozen geometric prior. Gen4P achieves SOTA or competitive performance on 3D point tracking, video depth, and camera pose estimation tasks.
SUPRA-Tex proposes a single-image 3D generation method that achieves super-alignment with the reference image by decoupling visible material estimation, surface alignment, and observation-guided 3D completion. The method first uses an MLLM-based image editing model to predict lighting-separated PBR textures under the reference viewpoint, then projects them onto the visible surface and encodes them as native observation latent variables, which an interleaved flow Transformer fixes as read-only anchors to generate complete material latent variables.
OmniVerifier-C1 proposes a decomposed hierarchical visual verifier that breaks complex prompts or environment rules into fine-grained sub-questions and scores each one, replacing traditional binary verifiers that only output True/False.
CoWorld Reasoner is an agent framework that uses a video world model (VWM) as a unified visual reasoning engine, exposing various visual operations through tools and letting the VLM autonomously decide whether to imagine, select operations and interpret the generated visual evidence.
The EVA-Cap framework decomposes audio-visual captions into atomic-level audio-visual events, converts unstructured text supervision into event-centric alignment, and trains on the self-built dataset EVA-Data through SFT and GRPO. The GRPO stage optimizes around three objectives: global event completeness, intra-event attribution, and inter-event synchronization. On the event-centric evaluation of ChronusAV, it improves by an average of 9.0%, and on WorldSense by 3.3%. The paper has been accepted by NeurIPS 2026, and the code will be open-sourced.
SCOPE is a GRPO-based reinforcement learning objective that lets a prompt rewriting model output semantic partitions while generating enhanced prompts, pools log-probability changes according to semantic spans, and shares loss coefficients for tokens within a span. On Qwen-Image, SD3, and FLUX.1-dev, SCOPE improves by an average of 9.48% on GeneVal and 7.62% on T2I-CompBench compared with the original GRPO, without additional span rewards or image rendering.
Velox is a training-free attention router that uses frustum tile correspondences in streaming geometry for region-level context selection, solving the problem that KV caches in chunked video diffusion models slow down as rollouts grow. At the 48th chunk, Velox accelerates attention by 4.36× and denoising by 2.43×, and on the WBench navigation test set across all 158 cases, its average score difference from dense LingBot-Fast-Cam remains within 0.15.
To address the high rollout optimization cost caused by perturbed reverse denoising trajectories in diffusion model reinforcement learning, RPO turns attention routing into stochastic exploration actions, samples key blocks according to native attention weights at several denoising levels, and uses a dual-rollout estimator to provide advantage signals. On text-to-image and text-to-video tasks, RPO improves generation quality while training efficiency is nearly 4× higher than Flow-GRPO-type methods, and at inference time it restores dense attention and retains the original deterministic sampler.
Speculative Forcing treats intermediate clean estimates in autoregressive video generation as drafts, predicting which tokens can be accepted early and frozen as final output, thereby skipping subsequent denoising steps. On LongLive, it achieves the quality of a four-step generator with about 1.7 denoising steps, a speedup of over 1.5x, reaching 41.7 and 37.8 fps; on Lingbot-World 2.0, it maintains quality with about 2.5 steps, a 1.35x speedup, reaching 24.5 fps.
MOSS-Image is an open-source image generation system that pretrains a 5B single-stream diffusion transformer (DiT) from scratch, supporting bilingual Chinese-English long-text rendering and precise layout control, and is equipped with an intent-aware agentic prompt enhancement module. On the long-text rendering benchmark LongLongTextBench, it achieves an average text coverage of 96.3%, ranking first among open-weight models; in the end-to-end system comparison on RealLayoutArena, it ranks second. The team will open-source the model, training recipe, and evaluation framework.
Researchers propose LILIE (Language-Informed and Lossless Image Encoder), an invertible image transform that, through joint training with a vision-language model, enables the model to recover image descriptions from noised features and thereby construct a language-informed loss for text-to-image generation.
PPMem is a video world model that uses planar primitives as persistent 3D memory. By adding a plane prediction head and a Depth Token Fusion (DTF) module to the video diffusion transformer, each autoregressive chunk generates both RGB video and planar geometry.
MTEdit builds a multi-turn image editing dataset through a pipeline centered on packed trajectory canvas (PTC) generation, containing 500K trajectories, up to 8 turns, averaging 5.38 turns, to reduce the visual quality drift caused by turn-by-turn rollouts.
The research proposes the MIST benchmark, which separates "whether the scene is set" from "whether instruction changes are executed," finding that the bottleneck of video world models lies in unexecuted instruction changes rather than forgetting. The accompanying method M3-Agent recursively rewrites input instruction details, raising the MIST reveal score by more than three times, improving the T score on WRBench by 26.1% and the U score on MBench-T by 52.7%.
MMLab@HKUST of HKUST, Tencent Video, and the University of Hong Kong propose Adaptive Reward Routing, which, on the DiffusionNFT framework, recomputes multi-reward routing for audio-video diffusion models in real time during training rather than fixing it once from the base model.
AReA is a training-free attention mechanism that lets the receptive field evolve from coarse to fine over the denoising process: early on it performs cross-scale coarse attention on downsampled key/value to capture global structure, and later it performs local attention within a gradually shrinking window to refine details. Experiments show it outperforms existing training-free baselines across target resolutions and different models, while being faster and more efficient. Currently only the flux1-krea-dev code is open-sourced, with Qwen-Image, Ideogram 4, and Flux2 support "coming soon."
The research proposes the Knowledge Transfer Probing (KTP) protocol, which injects new knowledge into Unified Multimodal Models (UMMs) through image generation training, then tests its retrievability via pure text question-answering.
Addressing multi-event combinations, multi-layer audio mixing, and asymmetric cross-modal relationships in joint audio-video generation models, researchers proposed the event-centered evaluation framework AVC-Reward, which decouples audio and video events at the event level and performs cross-modal cross-validation, aggregating into a coarse-to-fine three-level consistency metric.
OmniXRef is a unified multimodal multi-reference video editing framework that supports combinations of text, image, and video references. Its OmniXRef-Model uses a single-stream diffusion transformer to jointly process the source video and MLLM multimodal context.
ACEdit proposes a training scheme that keeps image editing models stable across multiple consecutive editing turns. Its core is introducing Self-Forcing into image editing: during training the editor is forced to process its own previous output, while a teacher editor provides distillation signals; the teacher is always anchored to the original clean image and receives composite editing instructions, forming asymmetric conditioning. Experiments show significantly reduced quality degradation over long multi-turn sequences while maintaining editing quality and instruction following.