Skip to content
Date and source

Dates follow the news timeline in Beijing time and include both endpoints.

Oct 15Thu
  1. Hugging Face Daily Papers22

    Dusting the Phantom: allocation strategies for decaying content in dynamic Gaussian representations

    To address the problems of fast-moving small objects being easily lost and details in highly dynamic regions being hard to recover in dynamic Gaussian representations, the research proposes two strategies, Seed Injection and Temporal Biased Relocation, which use temporal appearance changes and reconstruction error to locate under-reconstructed regions, and use time-resolved optimization signals to guide Gaussian relocation, respectively. Without changing the underlying representation and while maintaining a fixed Gaussian budget, the two improve reconstruction quality on multiple dynamic scene datasets.

  2. Hugging Face Daily Papers22

    Mine, Sample, Render: test-time diversity alignment for generative models

    Researchers propose MINE-SAMPLE-RENDER (MSR), a test-time method that mines interpretable categorical patterns from a corpus of real samples and samples them statistically, then uses the generative model only as a conditional renderer to output the sampled configurations, thereby improving generation diversity. On tasks including weather report generation, user simulation, web design, text-to-image, and agent-generated 3D assets, MSR improved alignment with the corpus distribution and diversity of outputs.

  3. 我爱计算机视觉 · 公开公众号 RSS37

    NeurIPS 2026 | OmniMem enables long video generation to directly retrieve the complete history

    Northeastern University and Adobe Research have proposed OmniMem, achieving full-history long video generation for the first time and allowing the model to retrieve the memories it currently needs in real time. The method keeps full-resolution historical KVs on the CPU, recalls them to the GPU on demand, and customizes Triton and CUDA kernels around block-level retrieval.

Oct 14Wed
  1. AITNT · AI创作资讯60

    Agnes open-sources 28.7-billion-parameter audio-video generation model Agnes-Video-2.0-MoE

    Agnes open-sources the audio-video generation model Agnes-Video-2.0-MoE, based on DaVinci-MagiHuman with approximately 15.3 billion parameters; it expands some computation modules in the middle 32 of 40 network layers each into three experts, bringing total parameters to approximately 28.7 billion, while the number of activated parameters per pass remains comparable to before the modification.

  2. Hugging Face Daily Papers25

    StreamErase: toward ultra-fast streaming video erasure

    StreamErase achieves streaming video erasure with sub-second latency and up to 26 FPS on a single GPU, enabling real-time interactive removal of target objects from video. The method first progressively transfers the erasure capability of a bidirectional base model to an efficient unidirectional streaming eraser, and employs sliding-window causal attention with attention sinks and KV cache rolling to achieve constant-speed generation under a fixed context window.

  3. Hugging Face Daily Papers29

    WorldCast: a multi-step world state prediction benchmark, GPT-Image-2.5 scores 72.8

    Researchers propose the WorldCast benchmark, containing 320 hand-designed cases across 8 categories, to evaluate multi-step world state prediction tasks, testing 12 representative models in total. GPT-Image-2.5 System scores 72.8, showing preliminary real-world visual state prediction ability; Nano Banana Pro scores only 42.8, and the best open-source model scores 20.1. The experiments also find that strong generative models can effectively leverage the capabilities of their underlying language models.

  4. Hugging Face Daily Papers22

    RepForcing: representation-space initialization for autoregressive video distillation

    RepForcing proposes replacing intermediate teacher trajectory regression with representation-space supervision to learn causal initialization for few-step autoregressive video generators, reusing synthetic videos generated by bidirectional models without extra causal ODE trajectory generation. Under four-step sampling, after DMD the method raises VBench Total from 84.04 to 84.66 (+0.62) and VisionReward from 6.326 to 10.074 (+59.25%).

  5. Hugging Face Daily Papers26

    XGEN releases DAO and JING: a generative world simulation framework

    The XGEN team proposes a generative world simulation framework that combines the computable world engine DAO with the first-person interactive experience model JING to enable continuous human-machine interaction. DAO maintains character, object, and spatial states and their relationships, advances according to interaction rules and world time while preserving the consequences of past actions, and then derives observer-specific conditions for JING to generate the corresponding audiovisual experience. Users can explore the environment, manipulate objects, and converse with characters through viewpoint control and text instructions, and their actions simultaneously change the shared world state and subsequent experiences.

  6. Hugging Face Daily Papers27

    Orbia: can generated video form a consistent 3D world?

    Researchers introduce Orbia, the first benchmark to systematically evaluate the 3D consistency of interactive world models, assessing along three dimensions: input preservation, persistence of generated content, and 3D self-consistency. The benchmark brings together public datasets, self-collected scenes, and simulations, building 720 indoor and outdoor instances and six exploration-revisit protocols, covering 27 text-, action-, and camera-conditioned models.

  7. Hugging Face Daily Papers22

    Self-conditioning enables deep sequential reasoning in Flow Matching

    The study finds that the trajectories of a native flow matching sampler can be simulated by a constant number of parallel model evaluations, and that increasing sampling steps cannot improve the expressive power of a fixed-depth backbone model. Self-conditioning (using the previous step's estimate as a condition) breaks this bottleneck: across multiple visual reasoning tasks, flow matching with self-conditioning achieves higher accuracy and greater reasoning depth, and performance improves as sampling steps increase; theoretically, a fixed-depth backbone can reach the expressive power of a space-bounded Turing machine.

  8. Hugging Face Daily Papers22

    Speculative Sampling in interactive video generation

    Addressing the reliance of interactive video generation on small bidirectional chunks to reduce latency, researchers propose Speculative Sampling, which extends the modeling window by using noise latents as padding and caches predictions for reuse when conditions are unchanged, thereby combining the image quality of a large bidirectional window with the low interaction latency of small chunks. Latency analysis shows the method eliminates the lower bound on waiting time caused by chunk boundaries, making the bidirectional window size a decision variable that can jointly optimize image quality and sampling time, and enabling large bidirectional models to be served in interactive streaming systems without complex streaming post-training.

  9. Hugging Face Daily Papers24

    Intern Lumina U2: a multi-codebook diffusion language model for full visual understanding and image generation

    Intern Lumina U2 is the first diffusion language model to represent multi-codebook visual tokens in a fully discrete space, where each visual position is represented by a set of codes from complementary codebooks; at the input, embeddings are concatenated by position, and at the output, a lightweight multi-codebook autoregressive head predicts the codes position by position, while the diffusion backbone processes all positions in parallel.

Oct 13Tue
  1. 机器之心 · 公开公众号 RSS36

    Teams from the University of Oxford and others propose the interactive video world model Oneira, which uses a World State Table to record object states

    Teams from the University of Oxford, Monash University, and the University of Bristol propose the interactive video world model Oneira, which uses a World State Table maintained by a Coding Agent to record scene objects and their states, separating 'world state' from 'frame generation' for processing.

  2. 机器之心SOTA模型 · 公开公众号 RSS22

    Zhipu GLM-5.3-Highspeed Gray-Release Exposure, Microsoft Releases Microsoft-Decision-1 Decision Scoring Model

    Zhipu GLM-5.3-Highspeed has appeared in gray-release feedback on the Zcode community, with some 3.15.1 subscribers randomly receiving this high-speed generation version; the general release scope and API pricing have not yet been confirmed. Microsoft has released Microsoft-Decision-1, which is post-trained based on Qwen3.5-9B and completes binary judgment, candidate selection, and scale scoring in a single forward pass. It is available for public preview on Foundry and hosted access on OpenRouter.

  3. Hugging Face Daily Papers12

    Scaling online optimal transport minibatches to improve flow matching training stability

    Researchers propose a scalable approximation of online entropic optimal transport that substantially increases the pairing batch size between source samples and data in flow matching training while retaining the benefits of optimal transport. They theoretically prove that the method strictly reduces full stochastic gradient variance and integration error in both the unregularized and entropy-regularized cases, and that the improvement grows as the online coupling batch size increases. Experiments on single-cell gene interpolation flows and image flow training verify that larger online coupling consistently improves training stability and final generation performance.

Oct 12Mon
  1. 机器之心Pro · 搜狐公开专栏33

    Text-to-Image "World Consistency" Benchmark TerraVis Is Here: Accepted at NeurIPS 2026, Scoring 18 Categories Item by Item

    TerraVis, accepted by the NeurIPS 2026 Evaluations and Datasets Track, proposes a "world consistency" evaluation framework for text-to-image generation, covering 18 categories of questions across three layers: objects, interactions, and scenes. It not only gives scores but also points out specific problems, visual evidence, and severity.

  2. Hugging Face Daily Papers24

    SyncReward: A Human-Aligned Synchronization Reward Model for Audio-Video Generation

    The research team proposes SyncReward, a reward model for evaluating and optimizing synchronization in audio-video generation, and releases the SyncReward-Data-80K dataset with 80K professional synchronization ratings and the SyncReward-Bench benchmark of 1K clips. The model achieves the highest agreement with human ratings in evaluations and can improve generation synchronization through inference-time best-of-N selection and reinforcement learning post-training. Data, benchmark, model and code will be open-sourced.

Oct 11Sun
  1. 王若风 · 开源创作项目分享62

    YuE2 released: write the ABC score first, then render the audio—taking the white-box music generation route

    YuE2 changes the approach to music generation: it first produces a readable and editable ABC notation score, then renders it into a 48 kHz stereo finished track with vocals and accompaniment complete. The code uses Apache 2.0, and the model weights are CC BY-NC 4.0; individual creators and musicians can freely use the weights to produce songs and monetize them, while commercial integration requires separate licensing.

Oct 7Wed
  1. GitHub:AI视频开源项目发现40

    Self Gradient Forcing Plus: Decoupled gradient flow enables autoregressive video generation

    Teams from Tsinghua University and JD.com's Joy Future Academy have open-sourced Self Gradient Forcing Plus (SGF+), which resolves gradient conflicts in SGF by separating context writing and denoising parameters, enabling autoregressive video generation of up to 24 hours using only a 5-second training window.

Oct 6Tue
Oct 3Sat
Sep 18Fri
  1. GitHub:AI视频开源项目发现70

    Alibaba Open-Sources SparkDiffusion: A Video Generation Acceleration Framework for Wan 2.1/2.2

    Alibaba open-sourced SparkDiffusion, an acceleration framework for DiT video generation models that combines sparse low-rank attention RoLa, few-step distillation CrossDistill, and self-developed high-performance operators, achieving over 200x end-to-end inference speedup compared with dense multi-step baselines.

    Why it matters: Alibaba open-sourced SparkDiffusion, using sparse attention and few-step distillation to speed up Wan 2.1/2.2 inference by more than 200x, and released multiple weight tiers.

Sep 16Wed
  1. Epidemic Sound28

    Epidemic Sound and Oxford VGG launch the audio retrieval framework AuroLA, now live for sound effect search

    Epidemic Sound and the University of Oxford's VGG have launched the audio-text retrieval framework AuroLA, based on the open-source Qwen2.5-Omni-7B, using a multimodal large model as both the fast retrieval and reranking backbone, surpassing the previous state of the art on all tested benchmarks with roughly 1% of the training data of comparable systems. The framework is already in production, powering Epidemic Sound's sound effect search, and a manually annotated validation benchmark of 1,297 sound effects has been open-sourced under the CC BY-NC-ND 4.0 license.

Jan 16Fri
Oct 23Wed
  1. Hugging Face Blog22

    CinePile 2.0: Building a stronger long-form video QA dataset with adversarial refinement

    CinePile 2.0 is released, an upgraded version of the long-form video QA dataset CinePile, using a newly proposed adversarial refinement method to improve data quality. The original CinePile was released in May 2024, containing approximately 300,000 training samples and 5,000 test samples. Humans outperform the best commercial vision models by 25% and open-source models by 65% on this benchmark. The team also open-sourced the implementation of the adversarial refinement method, which can be used to enhance existing datasets.

Jun 20Thu
Feb 15Thu
  1. OpenAI News82

    OpenAI Research: Video Generation Models as World Simulators

    OpenAI explores large-scale training of generative models on video data, jointly training text-conditioned diffusion models on videos and images of varying durations, resolutions, and aspect ratios, using a transformer architecture that operates on spatiotemporal patches of video and image latent codes. Its largest model, Sora, can generate one-minute high-fidelity videos. Results indicate that scaling video generation models is a promising path toward building general-purpose physical world simulators.

    Why it matters: OpenAI publicly reveals Sora's technical roadmap for the first time, explaining how video generation models can be scaled into physical world simulators.

Sep 8Fri
  1. Hugging Face Blog62

    Hugging Face brings T2I-Adapter controllable generation support to SDXL

    The Hugging Face Diffusers team, in collaboration with the T2I-Adapter authors, provides T2I-Adapter support for Stable Diffusion XL and releases checkpoints for conditions such as sketch, canny, lineart, depth, and openpose.

    Why it matters: Comparing ControlNet's parameter count and inference overhead, it presents a T2I-Adapter controllable generation solution on SDXL and reusable training scripts.

Jul 27Thu
  1. Hugging Face Blog62

    Apple and Hugging Face run Stable Diffusion XL on Mac with Core ML mixed-bit quantization

    Apple and Hugging Face have ported the Stable Diffusion XL base model to Core ML, usable on Apple Silicon Macs running the macOS 14 public beta; it currently uses the ORIGINAL attention implementation, and the refiner stage has not yet been ported.

    Why it matters: Apple and Hugging Face jointly present a Core ML mixed-bit quantization solution for SDXL on Mac, with measured latency for various models and compression recipes.

Dec 15Wed
Jun 17Wed
  1. OpenAI News20

    OpenAI's Image GPT: Generating Images with Transformers

    OpenAI found that a large Transformer model trained on language can also generate coherent image completions and samples when trained instead on sequences of pixels. The research also established a correlation between sample quality and image classification accuracy, showing that its best generative model contains features under an unsupervised setting that are comparable to those of top convolutional networks.

Apr 23Tue