Skip to content
Date and source

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

Nov 17Tue
  1. 中广互联 · 短剧24

    The Empathy Communication Mechanism of Cultural Tourism Micro-Dramas: Let Audiences First 'Enter the Story' Then 'Set Off'

    Cultural tourism micro-dramas achieve the transformation from empathy effects at the psychological-attitude level to cultural identity and communities of action through the construction of an empathy communication field, the triggering and deepening of audience empathy via narrative content, the generation of empathy feedback, and the transformation of empathy effects. This mechanism relies on platform technologies such as vertical 9:16 presentation, anchor components, and algorithmic recommendations, as well as narrative strategies such as reproducing collective memory through cultural tourism topics and conveying values through multiple character types, providing references for the production and broadcasting of cultural tourism micro-dramas.

Nov 13Fri
  1. Hugging Face Daily Papers35

    UniSpark: Exploring Hybrid Attention for Multimodal Understanding and Generation

    UniSpark-9B is proposed, a multimodal model with approximately 9B parameters, combining full attention with GatedDeltaNet linear attention, supporting visual question answering, text-to-image generation, instruction editing, and interleaved image-text prediction. It scores 2470.3 on MME-S and 0.83 on GenEval. The training data includes 159.7M generation samples, 213.4M understanding samples, and 624.8K interleaved sequences.

  2. Hugging Face Daily Papers24

    JVT: Just Video Transformer Makes Pixel-Space Video Diffusion Simple

    JVT (Just Video Transformer) proposes a simple large-patch DiT-style Transformer for pixel-space video diffusion, without pretrained autoencoders or U-Net-style skip connections. Its generation quality is comparable to latent diffusion baselines in both class-conditional and unconditional settings, and it supports joint generation of pixel-aligned modalities. Through camera-conditioned RGB-disparity generation, JVT improves camera following and obtains explicit scene geometry without requiring an additional reconstruction model.

Nov 12Thu
  1. Hugging Face Daily Papers37

    How PixVid lifts latent-space video diffusion into pixel space

    PixVid adapts a pretrained latent-space video model to generate directly frame-by-frame in pixel space, replacing the VAE decoder with a convolutional head. Fine-tuned from Wan2.1-14B, the method approaches the latent-space baseline with about 1.3M training clips and generates 2× higher resolution with the same number of tokens per frame. It maintains alignment for depth-conditioned editing under fast motion and can respond to each action in the next frame.

Nov 5Thu
  1. Hugging Face Daily Papers28

    NitroFusion2: Introducing Adversarial Loss into the Text Stream of Multimodal Diffusion Transformers

    NitroFusion2 reduces few-step generation of adversarial diffusion distillation from the usual 4-8 steps to 1-2 steps with comparable quality. It leverages the MMDiT text stream as a manipulable adversarial representation via Text-Stream Adversarial Distillation (TSAD) and Aspect-Cue Discrimination (ACD). One-step generation produces clearly rendered text and rich details, while two steps remain competitive on evaluation benchmarks.

Nov 2Mon
  1. Hugging Face Daily Papers29

    MoviEdit: Making Long-Video Multi-Shot Editing Simple

    MoviEdit is an end-to-end framework for scalable long-video multi-shot editing, simultaneously addressing data, modeling, and evaluation challenges. It generates 87K paired edited videos through an automatic data generation pipeline without manual annotation, and employs a cascaded bidirectional autoregressive design to establish a globally consistent editing draft on spatiotemporal compressed representations, then refines it via autoregressive propagation. MoviEdit-Bench benchmark tests show that MoviEdit significantly outperforms existing open-source models and matches industry-leading models in text-guided and image-guided editing.

Oct 31Sat
Oct 30Fri
  1. Hugging Face Daily Papers29

    SpecSlot: Read, Allocate, Route — Role-Coupled Tokenization for Pixel-Space Generative Transformers

    SpecSlot is a pixel-space rectified flow Transformer that decouples output detail from global attention sequence length via role-coupled tokenization. On ImageNet-512, SpecSlot improves FID by 2.84 over the baseline while reducing FLOPs by 3.9%; in text-to-image generation, the DPG-Bench score exceeds JiT-B/8 by 1.06, with forward cost reduced by 40.6%.

Oct 26Mon
  1. 机器之心Pro · 搜狐公开专栏70

    Microsoft proposes Code Video Model: controlling video generation with executable code

    The Microsoft team proposes Code Video Model, in which a Coding Agent converts a prompt into a Three.js program, unifying geometric layout, object motion, camera trajectories, and event timing into an executable scene representation; the preview video rendered by the program provides spatiotemporal references for the video model.

    Why it matters: The Microsoft team uses executable code instead of prompts to control video generation; training-free staged condition scheduling allows shot design to be inspected and modified before generation.

  2. Hugging Face Daily Papers25

    OmniShowcase: Omnidirectional Subject-Consistent Video Generation from Multi-view Images

    OmniShowcase proposes a method for generating omnidirectional subject-consistent videos from multi-view images, achieving subject consistency under 360-degree surrounding viewpoints through automatic data synthesis and geometric conditioning. Experiments show it achieves state-of-the-art in subject identity and 3D consistency while maintaining competitive video quality and motion. The research also constructs the Showcase-4K dataset (12 camera motions, 88 trajectory configurations) and the ShowcaseBench evaluation set.

Oct 25Sun
  1. 我爱计算机视觉 · 公开公众号 RSS36

    Open-o3-Video-v2 released: a video agent that thinks while watching and actively retrieves spatiotemporal evidence

    A team from Peking University, Fudan University, and Nanyang Technological University proposed Open-o3-Video-v2 based on Open-o3-Video, turning video reasoning into a multi-round agent process of "discover information gaps → actively crop clips → locate spatiotemporal evidence → incorporate into reasoning."

Oct 24Sat
  1. Tech Xplore42

    CU Boulder team proposes DeepSFU: detecting AI face swaps in video conferences without decompressing video

    A CU Boulder team has proposed DeepSFU, an AI face-swap detection system for video conferencing that directly reads inter-frame coding information in compressed video without decompressing and reconstructing frames, consuming 143 times less compute than existing models. Tested on 1,000 online videos containing AI face swaps, the system achieved near-perfect accuracy while only needing to inspect 1 frame out of every 100. The research team hopes to extend it to large-scale video conferencing scenarios.

Oct 23Fri
  1. Hugging Face Daily Papers31

    FilmFlow: an agent system for 3D virtual cinematography

    FilmFlow is an agent framework that converts natural language instructions or reference videos into explicit camera trajectories in dynamic 3D environments. Its Director agent generates executable camera plans by combining atomic operations that simultaneously control translation, rotation and composition, while three specialized agents—Interpreter, Scout and Critic—are respectively responsible for extracting scene intent, selecting the initial viewpoint and evaluating rendered results for iterative optimization.

  2. Hugging Face Daily Papers39

    AnyGameRenderer: cross-game entity-consistent generative rendering and an entity-level benchmark

    AnyGameRenderer is a video-to-video generative renderer that can turn placeholder visuals such as gray-box and colored geometric shapes produced by AI agents into high-quality footage with characters, motion details, and effects, while keeping the layout, camera, and the position and timing of each entity unchanged.

  3. Hugging Face Daily Papers31

    Verus-Insight: explainable artifact detection for AI-generated video

    Verus-Insight breaks down artifact detection for AI-generated video into three stages: inspection planning, evidence verification, and evidence aggregation, and uses human-anchored full-pipeline supervision and progressive capability training to improve fine-grained artifact recognition and video-level detection. The team also released the Verus-Bench benchmark with fine-grained human artifact annotations. Experiments show that the model has a clear improvement in recovering human-perceivable artifacts without increasing false artifact assertions.

  4. AITNT · AI创作资讯60

    CUHK MMLab and others propose DuoMatching to improve the image quality of few-step video generation

    A team from CUHK MMLab, ByteDance and HKUST proposed DuoMatching, which lets a video generator accept both joint-distribution supervision and frame-level marginal-distribution supervision, introducing image priors during training without adding inference computation. Under the per-frame 2-step setting, the VBench total score rose from 80.67 to 83.53, the overall preference rate against all four baselines exceeded 80%, and the dynamic degree changed slightly from 94.44 to 93.61. The paper is public on arXiv, and the code and model weights are released together.

  5. Hugging Face Daily Papers30

    Video as a lens: a survey unifying video understanding and generation through world modeling

    A 102-page survey proposes world modeling as a unified perspective, organizing research on video foundation models and video world models into four operations: read, write, share and interact. Read infers world properties from observed video; write synthesizes video under specified or learned world conditions; share integrates understanding and generation with common modeling mechanisms; interact predicts and updates world states under actions and evolving observations.

  6. Hugging Face Daily Papers31

    FAPL: making MLLM latent variables truly drive cinematic video generation

    Researchers propose Factor-Aware Predictive Learning (FAPL), which, through grounded training of a latent variable generator with a frozen MLLM and connector, enables the Video DiT to truly utilize MLLM latent variables rather than bypassing the native conditioning pathway. The method combines factor-span replacement to achieve factor-aware control over characters, scenes, actions and camera, and outperforms baselines on all three cinematic video generation tasks: sparse input to video, factor-aware editing and multi-shot generation.

Oct 22Thu
  1. Hugging Face Daily Papers42

    RealMe: Generating immersive volumetric video with minimal capture

    RealMe proposes a unified framework centered on a high-resolution video diffusion model that can enhance sparse, low-resolution multi-view captures into 2K, 4D-consistent video for volumetric video reconstruction. The model encodes auxiliary data as latents and combines them with layout-aware positional embeddings, adapts factorized spatiotemporal attention to multi-view capture layouts, and introduces a frequency loss in latent space to narrow the gap between generated video and real captures.

  2. Hugging Face Daily Papers22

    ALLOT: A training-free block sparse attention allocation method for video diffusion transformers

    Researchers proposed a training-free sparse allocation method, ALLOT, which treats the combination of each layer and each denoising step as a unit, estimates its contribution to the final video error through unit probing and trajectory calibration, and jointly optimizes the density of each unit under a measured latency budget.

  3. Hugging Face Daily Papers38

    RubatoCodec: Adaptive speech tokenization and streaming decoding with a budget

    RubatoCodec is a variable-frame-rate speech codec based on X-Codec-2, achieving adaptive tokenization through utterance-level event budgets and streaming decoding. Its Adaptive Firing mechanism uses codec features and waveform spectral changes to divide variable-length events without needing a learned boundary predictor. At low event rates, reconstruction fidelity and intelligibility outperform fixed-frame-rate coding, intelligibility remains best on English and Japanese benchmarks, and in matched TTS comparisons WER is lower while UTMOS and UTMOSv2 are higher.

  4. Hugging Face Daily Papers38

    DirCast: Esports Live Directing and Commentary with a Single Streaming Video Language Model

    DirCast uses a single streaming video language model to simultaneously perform directing and commentary for live esports broadcasts: it scans all player camera feeds at low resolution every second, a hold-or-switch directing module decides whether to cut the camera, and then it reads the selected camera feed at high resolution before speaking commentary.

Oct 21Wed
  1. Hugging Face Daily Papers47

    Joints, Camera, Action!: jointly generating 3D motion and video for arbitrary skeletons

    Researchers propose a model that can directly jointly generate video and 3D motion for arbitrary skeleton rigs, and estimate the camera to register the motion into the 3D scene. It connects a 3D motion model supporting arbitrary skeletons with a pretrained video model through a small number of cross-attention layers, is trained on about 200K training pairs, and in evaluations drives multiple skeletons by text more faithfully than video-first pipelines, while supporting video-to-skeleton motion, skeleton motion-to-video, and keyframe control.

  2. Hugging Face Daily Papers31

    OverSeer: improving multi-agent world model consistency with a fixed auxiliary camera viewpoint

    OverSeer is a multi-agent world model that, by introducing a spatially fixed auxiliary camera viewpoint beyond the agents' first-person views as a proxy for a world-centric representation, helps the model maintain temporal and cross-agent consistency. The auxiliary view encodes only changes relative to a static reference image via delta-VAE, significantly reducing the tokens and computation required for processing. Experiments on Minecraft, Doom, and real multi-camera captures show improvements in both world and agent consistency.

  3. Hugging Face Daily Papers38

    HRoPE: homography relative position encoding

    HRoPE proposes surface-aware relative position encoding, representing each patch's local geometry as a tilted 3D plane and injecting the plane-induced homography as a relative transformation into attention, to correct perspective distortion and align patch features across views.

  4. fal 官方博客28

    EAGLE-4 technical deep dive: Fireworks and fal release speculative decoding benchmark dataset

    Fireworks and fal jointly released a technical analysis of EAGLE-4 and Prompt Enhancement SpecBench, a speculative decoding benchmark dataset for creative tasks. The dataset contains 400 samples, covering 100 each of text-to-video, image-to-video, text-to-image and image-to-image. On Qwen 3.8 27B, EAGLE-4 uses sliding window attention and a low-rank candidate head to reduce draft cost, and the verification tree dynamically adjusts width and depth based on GPU latency measurements.

  5. Hugging Face Daily Papers38

    Any 3D scene in just 1K Tokens: a 3D-native representation for large-scale scene generation

    Researchers propose generating 3D scenes directly in a high-dimensional implicit 3D latent space derived from 2D foundation models such as DINOv2, SigLIP2, and DA3, and using a diffusion Transformer to perform diffusion modeling within that space, avoiding the representation redundancy and limited spatial consistency brought by 2D multi-view or video diffusion. Experiments show that this method outperforms 2D approaches in efficiency and spatial consistency, and validate its scalability as representation and model capacity increase. The paper has been accepted to NeurIPS 2026, and the code has been open-sourced.

Oct 20Tue
  1. Hugging Face Daily Papers34

    ActionRoPE: achieving spatially consistent control of flat-camera game world models with action coordinates

    ActionRoPE proposes a parameter-free action interface that converts action-driven camera displacement into world-coordinate rotational positions, so that tokens at the same world position maintain the same positional address over time, and pairs it with Coordinate-Conditioned Noise Scheduling (CNS) to allocate denoising schedules according to known versus new regions.

  2. Hugging Face Daily Papers45

    Manga2Animate: An agent framework for faithful manga animation using script and storyboard planning

    Manga2Animate is a multi-agent framework that breaks manga animation down into two intermediate representations: an animation script and a visual storyboard. Each representation is generated by a creation agent and reviewed by a paired review agent against the original manga storyboard. The framework uses representation-specific selection and a two-layer feedback mechanism to progressively refine both, then synthesizes the final animation. Experiments and user studies show that its output is more faithful to the original manga and has better shot-to-shot consistency than existing state-of-the-art methods.

  3. Hugging Face Daily Papers45

    Stateful Harness: A world model framework that separates rule execution from video rendering

    Researchers propose the Stateful Harness framework, which separates world state evolution from video rendering: rules maintain the world state, a frozen video model renders each event, and the generated clips are then used to update the state. The team built WorldState-Bench, containing 50 pairs of counterfactual twin samples; with rule prompts alone, the best of 11 video backends was only 32% correct; MiniMax-H3 rose to 64% when the ending was explicitly specified, and reached 92% with the full framework connected.

  4. Hugging Face Daily Papers31

    InfiniDream: Building open-ended photorealistic 3D worlds with procedural generation

    InfiniDream is a VLM-driven framework for open-ended procedural 3D world generation. It reuses existing procedural algorithms through symbolically constrained asset generation, producing editable, seed-diversified photorealistic assets. Its end-to-end text-to-world pipeline combines semantic scene planning with Deterministic Frustum-Guided Grid Placement to achieve precise object placement, procedural filling, and animation in large-scale 3D worlds.

  5. Hugging Face Daily Papers34

    RoAR-VAE: Improving latent diffusability through robust prediction

    RoAR-VAE improves the latent diffusability of visual tokenizers by predicting semantic features and reconstructing clean images from corrupted latents. Under the same training budget, its average performance on generation benchmarks is 2.19 points higher than the strongest baseline tokenizer, and its small-text reconstruction accuracy is 16.93 percentage points higher than FLUX.1-VAE.

  6. Hugging Face Daily Papers40

    DeOverlay: A benchmark and source-anchored flow model for video overlay removal

    DeOverlay is a source-anchored flow model that starts from source latents and reconstructs clean video with fixed source guidance, training a shared velocity network through region-weighted flow matching to achieve continuous correction from intermediate states. The team also released DVTE-Bench, containing 1,631 video pairs across 8 categories of text overlays, with OCR-filtered clean targets and frame-aligned pixel masks.

  7. Hugging Face Daily Papers48

    Fréchet Forcing: Training long-horizon streaming world models with short videos

    Researchers propose Fréchet Forcing, which fine-tunes a world model and its multi-step sampler through three constraints—generation anchoring, context fidelity, and realism matching—without needing a teacher model, online discriminator, or long videos. MiniWorld-1B, pretrained on 253-frame clips, after fine-tuning for up to 36 GPU hours on 381-frame clips, can roll out stably for more than 4,000 frames and consistently follows action instructions in robotic manipulation and camera-control exploration.

  8. 我爱计算机视觉 · 公开公众号 RSS50

    WorldCast tests 12 visual generation models: can they draw a chair, and can they also "fake" a chair?

    WorldCast built a benchmark covering 8 task categories, 320 handcrafted cases, and 1,352 visual states to be generated, evaluating 12 image models, unified multimodal models, and video generation systems. GPT-Image-2.5 paired with GPT-6 Astra scored 72.8, surpassing the participating video generation systems; when switched to outputting all steps at once as a grid image, the score dropped to 50.1.

  9. Luma22

    Luma's new method for sequence parallelism of multimodal masks

    Exact sequence parallelism for packed multimodal masks. It selects ring attention, sparse all-to-all, or Ulysses layer by layer, with numerics identical to a single GPU, and outperforms the FlashMask + Ulysses hybrid scheme on 47 of 48 cases on up to 128 H100s. https://lumalabs.ai/news/packed-multimodal-attention