Skip to content
Date and source

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

Oct 19Mon
  1. Hugging Face Daily Papers36

    PhysSkin++: A real-time general-purpose physics animation framework based on self-supervised neural skinning

    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.

  2. Hugging Face Daily Papers45

    VideoMaMa++: Temporally consistent video matting with preserve and refine tokens

    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.

  3. Hugging Face Daily Papers24

    SEAM: A spectral embedding alignment and merging method for low-resource TTS

    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.

  4. Hugging Face Daily Papers50

    StreamingSwap: A face-swapping framework for real-time video streaming released

    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.

  5. Hugging Face Daily Papers25

    SWIFT: self-guided weak-to-strong diffusion model training via target extrapolation

    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.

  6. Hugging Face Daily Papers33

    SpecKCast: accelerating diffusion model inference with spectral Kalman prediction

    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.

Oct 18Sun
  1. Reddit comfyui · 创作与工作流分享60

    ComfyUI tutorial: Testing and fixing workflows with FLUX 3 Base and Turbo

    The author tested FLUX.3's image generation and image editing in ComfyUI, comparing the Base version (CFG 3, 25 steps) with the Turbo version (CFG 1, 16 steps). In terms of generation, the model performs well on textures, artistic style diversity, wildlife and natural environments, but is weaker on hands, fingers, feet, realistic skin and complex scenes; image editing is instead its strength, with the Turbo version performing well on style transfer, outfit changes, background replacement, weather changes and object removal.

  2. Reddit comfyui · 创作与工作流分享8

    ComfyUI user asks for help: repeated textures appear on skin after 3k steps of LoRA training

    A ComfyUI user asked for help on Reddit: their trained LoRA keeps generating repeated textures on skin after 3k steps, and this only occurs when the LoRA is enabled, ruling out workflow issues. The user found that the dataset images themselves contain similar textures, just less obvious, and asked whether it can be fixed without redoing the dataset.

  3. Monster社区 · 中文ComfyUI实测与工作流38

    Getting started with the Flux.3 5.6B all-in-one package: 4-step iterative text-to-image and single-image, multi-image editing

    The Flux.3 5.6B all-in-one package tutorial demonstrates three uses of the same package: text-to-image, single-image modification and multi-image modification, with the demo interface set to 4 iteration steps. For text-to-image, fill in the prompt first, then check the size and iteration parameters before generating; for single-image modification, upload a reference image and specify the parts to change; for multi-image modification, place both a character reference image and a clothing image and explain the purpose of each. To enter the ComfyUI node interface, access custom port 8187; the actual address shown when the package starts up prevails.

  4. note · seisei_ai_1st · 图像视频创作分享40

    Kling 4.0, Seedance 2.5, Wan 3.0 compared under identical conditions: AI video generation hands-on test

    Using 4 photos and the same prompt on SJinn to generate 9:16 vertical 30-second videos under identical conditions, Kling 4.0, Seedance 2.5 and Wan 3.0 were compared. Wan 3.0 had the best detail and image quality, and the best hand and character stability; Kling 4.0 had the most natural motion; Seedance 2.5 lost out on image quality because it is only 480p. All three models succeeded on the first generation with zero retakes, but grabbing food such as "picking up saury" remains a common difficulty.

  5. TelosCompass · note AI视频创作实践22

    DomoAI video generation money-saving guide: confirm model and duration first, then use Relax mode

    DomoAI's Relax mode can generate videos without consuming credits when three conditions are met: plan, model and duration, but under the same plan some models and durations only support Fast mode. Users should check the pricing page in the order plan → model → duration, and confirm that the "Relax モード" switch is on; they cannot judge merely by the credit number disappearing from the generate button. Relax is suitable for trial work when you can wait until the usage time limit and have margin for one redo; use Fast when in a hurry.

  6. note · kongo_jun · 图像视频创作分享38

    Trying MOSS-Image in ComfyUI: OpenMOSS's 5B T2I model

    The 5B text-to-image model MOSS-Image, open-sourced by the OpenMOSS team, can now run in ComfyUI and offers Base and SFT versions. In testing, generation took 14 seconds and used 15GB of VRAM on an RTX 4090, requiring the qwen3.5_9b text encoder and flux2-vae, and using the BBOX node (Nodes 2.0's Vue node design must be enabled). The author rated its art style and text generation as good.

  7. Stuubzzz · 视频工作流实测66

    Hands-on test of Qwen-Image 2.1 Turbo in ComfyUI: 8-step scheduling and 300 renders

    The author tested Qwen-Image 2.1 Turbo with int8 files on an RTX 5090: native 2K (4 MP) images averaged 11.8 seconds, versus 36.9 seconds for the 30-step base model, about 3.1 times faster.

    Why it matters: The author used 300 renders to test the differences between Turbo and the base model, and provides reusable 8-step sigma and resolution settings.

  8. note · kawakijourney_ai · 图像视频创作分享30

    Drawing 3D female characters with Midjourney: 5 types of fantasy character design sheets and standing illustration prompts released for free

    Using Midjourney to generate 3D fantasy female characters for games, covering 5 types: cute mage, adult magus, witch, saint, and battle mage, each with two English prompts—one for a character design sheet and one for a standing illustration—10 in total, with Japanese translations and design notes. The design sheets use a horizontal 16:9 layout showing front, side, back, and angled full-body views plus expression variations; the standing illustrations use a vertical 9:16 single-character centered full-body composition. Both share the same appearance description to keep the character consistent.

  9. note · fordtaurus116 · 图像视频创作分享36

    Vidu Q4 Preview hands-on: making Japanese character animation with GPT Image 2.5 × Pollo AI

    The creator tested the video generation model Vidu Q4 Preview on Pollo AI, generating original character illustrations with GPT Image 2.5, then produced a 5-second Halloween short and a 16-second self-introduction animation, with natural Japanese dialogue intonation and lip sync. The model has a low cost for short animation production, and the credits required vary with video length and resolution.

  10. note · promptdesign · 图像视频创作分享22

    New AI beauty portrait prompt collection ⑮: 100 templates ready to copy into ChatGPT and Gemini

    This prompt collection contains 100 JSON-format image generation prompts on the theme of "AI beauty × spectacular travel scenery," ready to copy into tools such as ChatGPT and Gemini. In addition to the base version, each theme includes 4 variations: 9:16, 1280×670, background change, and atmosphere change, along with negative_prompt and aspect_ratio fields.

  11. note · myloose · 图像视频创作分享10

    [AI image generation research] Wanting AI to draw monolidded men ⑩ "Just don't draw it!!!" The final episode

    After three days of intensive use of Astra to generate monolidded male portraits, the author hit usage limits and had their account locked for three days. They ultimately gave up on solving the double-eyelid problem within ChatGPT and instead had ChatGPT quickly develop a mobile PWA web app to manually erase the double-eyelid lines in illustrations. The app will be released in the next installment, and the author says the root problem has not been solved.

  12. note · singulatelier · 图像视频创作分享22

    How SINGULATELIER semi-automates daily image posts with its exclusive AI character HAZ

    SINGULATELIER created the exclusive AI character HAZ (Eigetsu) as a unified brand visual identity, running through note posts on various themes such as music, video, and merchandise. The author leaves details such as clothing, scenes, and composition to AI employee Chris, taking charge only of direction and final judgment, and uses the GPT PLUS scheduled task feature to produce images and copy at a fixed time every day.

  13. note · zou_aidouga · 图像视频创作分享42

    25 AI video camera movement prompts from Google's official guide: 13 camera angles + 12 camera movements, with a ChatGPT-generated comparison cheat sheet

    Based on the 13 Camera angles and 12 Camera movements in Google Cloud's video generation prompt guide (updated October 15), the author used ChatGPT to generate comparison images of the same woman, the same alley, and the same sunset, making a cheat sheet of 25 camera movement names.

  14. note · nobinlog · 图像视频创作分享26

    Generating fighting-game illustrations with Anima in ComfyUI-MangaScreen

    Generating fighting-game illustrations with the Anima model in ComfyUI-MangaScreen; the author thinks the Qwen-image version is decent in quality, but images with delicate coloring gradients adapt poorly, and illustrious is probably even less suitable, while flat color schemes and line-art styles produce results more easily. Because the Qwen-image version was hard to get right in a short time, they switched to Manga-Lite to finish, and noted that hair and eyes tend to get thick and require repeated parameter tuning, so it doesn't suit every art style.

  15. Reddit StableDiffusion · 创作者与工作流分享15

    Moss Image character LoRA training: which methods work for you?

    A creator used Ostris Toolkit and Fizgig to test training methods that had previously worked on Krea2, ZIB, IDG4, and SDXL, but when used for Moss Image character training the similarity was only about 80%. Their attempts with Academia SD and LOKR methods performed even worse; believing their dataset was already sufficient, they asked the community for Moss Image character LoRA training experience.

  16. Reddit StableDiffusion · 创作者与工作流分享15

    FastH3 loses its speedup after connecting a LoRA in ComfyUI—how to fix it?

    A user using FastH3 in the default ComfyUI template workflow found that after connecting the LoRA's model output to Basic Guider and Basic Scheduler, the Model Attention Backend and Model Sparse Attention nodes became disconnected, causing FastH3 to lose its speedup. The issue comes from a help request post in the Reddit StableDiffusion community.

  17. Reddit StableDiffusion · 创作者与工作流分享12

    LoRa training help request: character LoRa trained for 10 epochs looks completely unlike the target in ComfyUI

    A user trained a character LoRa on onetrainer with 10 epochs, local batch 2, and repeats 10, then imported it into ComfyUI and generated with the prompt mychar_camille, but the results look completely unlike the training images. The post gives no solution and only asks how to adjust training to get a usable character LoRa.

  18. 诀.com · 创作开源项目拆解22

    Midjourney beginner's guide: the full workflow from cultivating taste and building a P to generation and editing

    This Midjourney tutorial uses a natural-light female portrait as an example to demonstrate the complete web workflow from scratch to final image: first browse works in Explore to cultivate taste, then select images in Personalize to build a switchable aesthetic profile (P), then write prompts and parameters in Create to generate images, pick a base image, and use Vary or Edit to change clothing, repaint with masks, and extend the canvas to add blank space.

  19. Reddit comfyui · 创作与工作流分享22

    Building a video pipeline with human review in ComfyUI: how to increase usable output per hour?

    An e-commerce content team used AI-assisted development to build an internal ComfyUI pipeline: product images plus approved scripts are generated via parallel GPUs, then upscaled, transcribed and subtitled, and finally manually reviewed and archived. The team raised three operational questions: how to use lightweight checks to distinguish speech generation errors from transcription errors, how to incorporate human acceptance rate and review time into throughput measurement, and how to turn rejection reasons into improvements to scripts or the generation stage.