Skip to content
Date and source

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

Oct 22Thu
  1. note · kazu_fuwa · 图像视频创作分享73

    Making a 10-minute short drama and vertical short videos with MiniMax H3: how to fix broken hands, abnormal blinking and six legs

    The author used the MiniMax H3 Ref2VA model on local ComfyUI to produce a 10-minute comedy short drama, 7 types of vertical short videos and a 60-shot cat character video, at about 2 minutes per shot, or about 3.5 minutes after switching to a 768p 8-step LoRA.

    Why it matters: The author ran MiniMax H3 on local ComfyUI to make multi-shot character videos, explaining one by one the causes of and fixes for broken hands, blinking and extra limbs, which can be transferred to your own production pipeline.

  2. 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.

  3. Reddit comfyui · 创作与工作流分享22

    User compares LTX 2.5 and Wan2.2 video generation speed on an RTX 5090 laptop

    A user compared LTX 2.5 and Wan2.2 generating 5-second videos on an HP Omen 16 laptop (RTX 5090 with 24GB VRAM): LTX 2.5 took only 0.54 minutes, while Wan2.2 took 45 minutes. The user found that when running Wan2.2, the GPU used only 14GB of VRAM, with the rest offloaded to 32GB of system memory (28GB actually used); both setups used out-of-the-box default settings.

  4. Reddit comfyui · 创作与工作流分享22

    Building a video generation workflow in ComfyUI: how to make more realistic AI videos

    A user built a complete video generation pipeline in ComfyUI, from YuE2 generating songs, Z-Image Turbo producing character images, and Qwen-Image-Edit 2511 making keyframes, to Wan2.2-S2V lip-sync, two-stage rendering with Wan2.2-I2V, and FILM frame interpolation to 48fps. However, the generated results often show distortions such as three arms in hands and multi-person scenes, and the user is seeking suggestions for more realistic tools and open-source models.

  5. Reddit comfyui · 创作与工作流分享38

    Comparison test of FastH3 Omni and Ref2VA + 8steps Turbo on a 3090

    A user compared FastH3 Omni and Ref2VA + 8steps Turbo on a single 3090: all four shots were native 768×1344 at 24fps, with identical prompts, seeds, and identity and clothing references, and all backgrounds generated from text prompts. Omni averaged 8 minutes 26 seconds per render, while Turbo + VSA took 8 minutes 01 seconds, both covering 158-frame generation and showing the first six seconds of each segment. The user believes Turbo + VSA has better image quality and is slightly faster.

  6. Reddit comfyui · 创作与工作流分享22

    Seeking recommendations for character expression and manga effect workflows: ComfyUI solutions beyond PixAI Studio and Tsubaki 3

    An AI anime studio, after using PixAI Studio and Tsubaki 3 to complete expression sheets, character emotional consistency, manga shading effects, and color-to-black-and-white manga storyboard conversion, is now looking for reusable ComfyUI workflows. Its self-built solution relies on character LoRA, IPAdapter, and ControlNet, but maintaining character consistency across outputs is very costly, and in production it prefers ready-made workflows.

  7. 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.

  8. note · imamirai_editors · 图像视频创作分享36

    ChatGPT image generation: Aizawa freely releases prompts for five scenes: morning, noon, evening, night, and late night

    Aizawa posted five scene illustrations on X from the morning of October 21 to late night on the 22nd, and freely distributed corresponding ChatGPT image generation prompts, covering opening curtains in the early morning, reading a letter during the day, napping on the sofa in the evening, wrapping in a blanket at night, and clipping a moon-shaped bookmark late at night. The prompts require attaching a reference image of an adult character aged 21 or older that the user has the right to use, to maintain facial features, hairstyle, and eye color, and uniformly specify a delicate and clean anime style and horizontal composition.

  9. Stable Diffusion Art 教程78

    Implementing MiniMax H3 multi-keyframe video and audio generation with a ComfyUI workflow

    Stable Diffusion Art released a tutorial on how to use MiniMax H3's Add Guide node in ComfyUI to lock multiple keyframe images, video clips, and audio to specific frame positions on a 24fps timeline, generating a video with stereo dialogue, sound effects, and music in one pass.

    Why it matters: The tutorial provides specific node settings for multi-keyframe locking and audio-visual continuation, and fixes the timestamp truncation issue in the official template.

  10. Reddit comfyui · 创作与工作流分享38

    Fibo 2 edit, Klein 9b, and Qwen 2.1 image editing comparison: who is stronger at 1 megapixel

    A user built a comparison page testing the image editing capabilities of Fibo 2 edit, Klein 9b, and Qwen 2.1 at 1 megapixel with the same seed and int8 single-pass generation, concluding that Qwen 2.1 remains the best for their use case, and Fibo could be a good choice if it gains community support.

  11. Luma News48

    Ray 3.2 vs Seedance 2.5: Which Luma model is better suited for your project?

    Luma's Ray 3.2 and Seedance 2.5 focus on frame-level control and synchronized audio generation, respectively, allowing users to choose based on project needs. Ray 3.2 supports HDR output, EXR export, and up to 16 keyframes, making it suitable for professional post-production; Seedance 2.5 can generate 30-second clips with synchronized sound effects, music, and lip sync in one go. Both are available through Luma and can be switched within the same workspace.

  12. Reddit StableDiffusion · 创作者与工作流分享38

    Comparing two reference-image editing methods for Moss Image Edit

    User Sixhaunt trained a rank-16 LoRA (6,000 steps) with AI toolkit and compared two reference-image input methods for Moss Image Edit: the direct reference-image method tends to mix inputs together without a LoRA, while Ostris's editing method approximates image overlay without a LoRA and gives better editing results with a LoRA. The Ostris method renders about 45s per image, the simple reference method about 15s, and training takes about 3 times as long.

  13. Reddit comfyui · 创作与工作流分享58

    The author shares using the MiniMax 360 surround workflow to improve facial consistency

    The author shared a 360-degree orbit MiniMax workflow in ComfyUI, calling it the definitive method they found for multi-angle face consistency. The approach is to first prepare a very solid frontal face image, then have the AI do a 360-degree orbit and capture frames at each important angle; the author claims face consistency reaches 100%, outperforming the search for real photos of real people at different angles.

  14. note · kasmintea · 图像视频创作分享18

    Generating AI illustrations with ChatGPT: creating a sense of air with forest light and shadow

    The author used ChatGPT to generate AI illustrations, asking for emphasis on particle feel, opacity and perspective, referencing the light effect of car headlights in the rain; ChatGPT provided multiple options at once and explained the common elements. The final choice was a forest scene: foreground and background vegetation naturally distributed, light intensity and direction supplementing the layers, dark-area depth enhancing the sense of presence, and leaves with low noise blending into the background without stealing focus from the subject. The author believes AI illustrations are better described in terms of Lighting and Writing to convey atmosphere rather than stated plainly.

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

    Running LTX-2.5 on Ryzen AI Max 395+ with ComfyUI to generate a 5-minute Slimer video, how to solve visual inconsistency

    A user installed ComfyUI on a Ryzen AI Max 395+ with 64GB of memory, using LTX-2.5 to generate 10-second clips at 720p, 24fps, aiming to assemble an approximately 5-minute Ghostbusters Slimer video, but each of the 10 results had problems—first growing legs, then voices, then green or white buttocks appearing. He plans to stitch about 300 segments, with each render taking about 10 minutes, and is seeking a more stable model or prompt solution, with plans to later upgrade to 1080p and then upscale to 4K.

  16. Reddit comfyui · 创作与工作流分享18

    Workflows perform worse after upgrading ComfyUI to 4.2, user asks for rollback method

    A ComfyUI user upgraded from version 3.x to 4.2, and multiple previously well-functioning workflows developed blurry images, broken details such as eyes and lips, and backgrounds and textures blurring together, with LoRA and parameters unchanged. The user uses krea2 and found that LoRA strength above 1.0 seems to cause problems, but lowering it still couldn't restore the previous clear results, and is asking about the cause and how to roll back to 3.x.

Oct 21Wed
  1. TelosCompass · note AI视频创作实践40

    Using ChatGPT to write storyboards from text: first determine the presentation order and framing, then generate images

    The article proposes an approach of first writing the storyboard in text and generating images only after confirmation: break the planned scenes into shots, write out the presentation order and the overall/medium/close-up framing for each, then have ChatGPT output a storyboard plan according to a unified project. Using a fictional umbrella shop 20-second, 4-shot video as an example, it requires the AI not to generate images before confirmation, and to revise line by line from four points: narrative coherence, visual consistency, shot function, and production volume.

  2. Higgsfield 官方博客38

    A hands-on comparison of the top ten AI video generators in 2026: Higgsfield, Runway, Kling, and more

    Higgsfield ranks first in the 2026 comparison of the top ten AI video generators, integrating 35+ models and tools for character consistency, ad production, and agent capabilities into one subscription. Runway focuses on editing and VFX workflows with Gen-4.5, the official Kling app offers direct access to Kling 3.0/4.0, and Google Flow offers Veo 3.1 starting at $4.99. All platform prices and quotas were verified in October 2026.

  3. Reddit comfyui · 创作与工作流分享13

    MiniMax ref2video places a character into a 3D video

    The user used the default MiniMax ref2video template in ComfyUI to place a character into a 3D video; 8 seconds at 900x400 resolution took 40 minutes on an RTX 5090. The user is looking for a turbo LoRA configuration that can match the base template's ability to follow the reference video, but the turbo LoRAs they have tried cannot follow the input video as accurately as the base template.

  4. Reddit comfyui · 创作与工作流分享41

    MiniMax H3 I2V secondary character's eyes slightly deform, how to keep the face stable without a "fixed" look?

    The user generated a 9:16 gym promo short with MiniMax H3 I2V in ComfyUI (minimax_h3_fl2v_turbo LoRA, 8 steps, 5 s @ 24 fps, 1.5 MP upscaled via RTX Video Super Resolution to 1080×1920); the main character's face is stable, but the background coach (about 90–100 px wide) has eyes that drift frame by frame.

  5. Generative AI · 创作教程与项目分享40

    SeaArt vs PixAI anime AI image generation comparison: which to choose in 2026?

    With the same set of anime prompts, anime-focused PixAI generates using its in-house model Tsubaki.3, while SeaArt requires picking an anime model among mixed features like images, video, and face swap. PixAI's Tsubaki.3 was trained on 50 million images selected from about 500 million anime images, and supports editing images with natural language, reusing characters via @image1, and generating multi-panel comic pages.

  6. 袋鼠帝AI客栈 · 公众号公开订阅58

    Kangaroo Di Open-Sources 3 WorkBuddy Design Skills Covering Posters, 3D Blind Boxes, and Brand Promo Videos

    Kangaroo Di has open-sourced 3 design skills adapted for WorkBuddy and Miora: a Halloween poster premium color scheme, a 3D blind box designer toy concept image, and a brand promo video planning workflow; the first two are included in the design-skills collection and the third in the director-skills collection.

  7. 沃垠AI · 公众号公开订阅62

    Writing Code to Make Videos with MiniMax M3.1 Flash, with Full MV Prompts

    The author, Leng Yi, used MiniMax M3.1 Flash to produce a 2-minute video and an MV by writing code to render them; the former took 35 minutes and cost about 0.3 yuan. The workflow is to have the model write deterministic rendering functions, use Three.js, Canvas, CSS animations, and so on to write the visuals as a web page controlled by a time parameter t, then use a headless browser to capture frame-by-frame screenshots and hand them to FFmpeg to encode into an MP4.