Skip to content
Date and source

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

Oct 20Tue
  1. 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.

  2. 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.

  3. 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.

  4. IT之家68

    Ant Ling opens source Ming-3.0 speech series models, covering speech synthesis, audio understanding and dialogue, and speech recognition

    Ant Ling has open-sourced the Ming-3.0 speech series, including three series—speech synthesis, audio understanding and dialogue, and speech recognition—with five models in total, which developers can select by task or combine into workflows. Ming-3.0-tts's Voice Design supports describing the target timbre in text, and can also set the timbre with reference audio; director mode can specify how lines are performed, emotional changes, stressed words, and pacing.

    Why it matters: Ming-3.0 splits speech synthesis, audio understanding, and recognition into five models, among which director mode can describe the performance style of lines in text.

  5. CGPress72

    Mirelo releases SFX 2.0 video-to-sound-effect model

    Berlin-based AI research team Mirelo released the video-to-sound-effect model SFX 2.0, which can modify a video's audio without regenerating the visuals, preserve dialogue, remove unspecified music, and generate sound effects, Foley and ambience synchronized with action and consistent across cuts.

    Why it matters: The video-to-sound-effect model has been updated to redo the audio track without changing the visuals, and has been integrated into editing software and third-party platforms.

  6. The AI Pixel58

    Krea releases Krea 3 image editing model, native 2K single-image edits in under 8 seconds

    Krea released the image editing model Krea 3, with most edits completed within 8 seconds at native 2K, at $0.06 per image, or $0.03 at 1K. The official comparison cites Nano Banana Pro at a median of 33 seconds and $0.135 per image, and GPT Image 2 at 90 seconds and $0.22. The model supports up to 6 reference images per edit and can generate images from marker sketches, mood boards and line art.

  7. 机器之心SOTA模型 · 公开公众号 RSS22

    DeepSeek updates V4.1-Flash-1017 and launches Search/Fetch API; StartLux releases decision model v1.1

    DeepSeek updated V4.1-Flash-1017 simultaneously on its official website, App and API on October 17; the API model name remains deepseek-flash, with unchanged invocation method and pricing, and its Terminal-Bench 4.0 score under the Claude Code framework rose from 31.2 to 42.3.

  8. Slop Reel · AI 视频制作简报62

    ByteDance's OmniHuman 2.0 goes live, bringing digital human video down to $4.20 per minute

    BytePlus launched OmniHuman 2.0 on October 15. A single image plus one audio track can generate a 720p performance with lip sync, with text prompts controlling gestures and camera movement, and multiple people allowed in the same frame. The price dropped from $0.12 per second to $0.07 per second, i.e. $4.20 per minute and $21 for a five-minute long take. The product is sold through ByteDance's enterprise platform BytePlus, and budgets need to account for how many takes are required to get one usable minute of footage.

  9. 智东西40

    Three OpenAI models with Aurora codenames appear in API and Arena, possibly GPT-Image-3

    Three entries—aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha—briefly appeared in the OpenAI API model list, and image models with the same names on Arena have already generated cat images with the red and green versions. Reviewer Lumina speculated this could be GPT-Image-3, saying the red version generates faster but with slightly lower image quality, the green version is slower with better results, and the blue version has not yet been tested. OpenAI has not disclosed the models' identities or release plans.

  10. Reddit comfyui · 创作与工作流分享73

    OpenMOSS Releases MOSS-Image Image Model, Now Supported in ComfyUI

    OpenMOSS released MOSS-Image, a 5B DiT model using the Qwen3.5-9B text encoder and FLUX.2 VAE, available in Base and SFT versions under the Apache 2.0 license.

    Why it matters: MOSS-Image is open-sourced under Apache 2.0 and supports structured layout control, allowing evaluation of text rendering capability against existing open-source image models.

Oct 19Mon
  1. AITNT · AI创作资讯78

    Kling 4.0 fully launched: 30-second native direct output, 15 reference assets and 10 keyframes tested

    Kling 4.0 is fully launched, with a single native generation of up to 30 seconds, supporting long takes and continuous narrative. The new model strengthens stability in high-speed motion and complex camera movement. The all-purpose reference feature allows a single task to combine up to 15 image, video, subject and voice assets, including up to 10 images, up to 5 videos and up to 7 subjects, and supports up to 10 keyframes to control character states and plot direction.

    Why it matters: Kling 4.0 extends single generation to 30 seconds and supports 15 reference assets, allowing creators to judge the usability of long takes and action scenes accordingly.

  2. Reddit StableDiffusion · 创作者与工作流分享22

    6 open-source models compared on the same prompt: Krea2, Qwen2.1, Flux3 Klein, Flux 2 Klein, ZIT and Moss

    A user compared the outputs of six open-source models, Krea2, Qwen2.1, Flux3 Klein, Flux 2 Klein, ZIT and Moss, using the same prompt. Krea2 needed a text-refusal bypass LoRA to generate the missing text and make Cousin Itt appear; Qwen 2.1 seemed to interpret Lurch's gray hair as gray skin. The prompt was a Wednesday Addams-themed realistic photography scene.

Oct 18Sun
  1. Reddit StableDiffusion · 创作者与工作流分享20

    MOSS-image can generate images without LoRA

    MOSS-image can generate images directly from prompts without LoRA, and is compatible with 1girl-type prompts and JSON-style formatting. The example prompt generated an overhead photo of a female minotaur lying on her back on grass, wearing a white tank top with an orange rounded-font logo reading "Muuhters," labeled as shot on a Canon R5 DSLR.

  2. Hugging Face Daily Papers34

    Painting Change Map: using spatiotemporal change maps to guide painting-process video generation

    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.

Oct 17Sat
  1. RenderBob · 图像视频工作流72

    Kuaishou Kling 4.0 released on October 15, with HDR, continuation, and API still pending

    Kuaishou says Kling 4.0 was released on October 15, with a single generation of up to 30 seconds, Omni Reference supporting up to 15 assets and up to 10 keyframes, and integration with Kling MCP and command-line tools.

    Why it matters: Kling 4.0 is live, but HDR, two-minute continuation, and the API are still marked as coming soon; the article also lists the parameter differences between the Flash tier and the full version.

  2. FastVideo 视频模型 · 模型发布60

    FastVideo releases FastH3-QAD-v2 video-audio model

    FastVideo releases FastH3-QAD-v2, an 8-step video and audio model trained with NVFP4 weights and quantization-aware distillation (QAD), runnable on a single Blackwell GPU (RTX 5090, RTX PRO 6000) or DGX Spark.

    Why it matters: The 8-step video and audio model lands on a single Blackwell GPU with NVFP4 quantized weights, and error data versus the BF16 teacher model is provided.

  3. Hugging Face Daily Papers31

    HumanSkill-4D: training-free language queries can animate Gaussian avatars, with spatial and temporal grounding

    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.

  4. Hugging Face Daily Papers38

    UniAVC: unified audio-video creation via progressive training and asymmetric scheduling

    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.

Oct 16Fri
  1. 可灵 Kling58

    Kling 4.0 fully opens on kling.ai, focused on upgraded motion and creative control

    Kling 4.0 is now fully available on kling.ai, with the company saying it brings more refined motion, more precise creative control and professional-grade audiovisual output. At Advertising Week New York 2026, Kling AI shared the stage with WPP and Adobe to discuss how AI video is moving from creative experimentation to real production, emphasizing that the next stage is not just generation quality but connecting models, tools and workflows so creativity can scale with consistency and controllability.

  2. 新智元 · 公开公众号 RSS71

    Shengshu Technology Releases Vidu Q4 Preview, with Launch Pricing as Low as 0.09 Yuan/Second

    Shengshu Technology released Vidu Q4 Preview, focused on character performance, camera movement, and complex visual effects, strengthening character acting and dynamic camera work in 2D and 3D anime scenes, with launch pricing starting as low as 0.09 yuan per second.

    Why it matters: Vidu Q4 Preview bundles character performance, camera movement, and multi-reference capabilities and offers launch pricing starting at 0.09 yuan per second, which can be used to judge whether test-shoot costs have entered an affordable range.

  3. Hugging Face Daily Papers24

    SUPRA-Tex: 3D generation based on super-aligned textures

    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.

  4. 可灵 Kling70

    Kling releases 4.0 video generation model

    Kling's official account released 4.0 with the caption "The art of 4.0" and the hashtags #Kling4 and #KlingAI. The quoted @2x_labs content said Kling 4.0 has been officially released, with the slogan "Define the next frame with imagination."

    Why it matters: Kling's official account has signaled the release of 4.0, so creators can watch for changes in the new version's video generation capabilities.

  5. comfyorg · YouTube 创作视频38

    Inside the behind-the-scenes conversation between Black Forest Labs and ComfyUI on FLUX 3 Image

    Black Forest Labs research scientist Yumeng Li and ComfyUI creative technology expert Paul Hansen held a live discussion on the R&D decisions and trade-offs of FLUX 3 Image. The model supports composing with bounding boxes and editing images box by box, with unmodified areas remaining unchanged, and can perform multiple targeted edits simultaneously. It also supports multi-reference generation with up to 10 images, natively outputs 2K and 4K, and comes with a ComfyUI competition with a prize pool worth $25,000.