Skip to content
Date and source

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

Nov 18Wed
  1. note · kongo_jun · 图像视频创作分享62

    Trying the 2step LoRA of MiniMax-H3-Image and P-H3-Image in ComfyUI

    The author configured and tested the 2step LoRA of MiniMax-H3-Image and P-H3-Image in ComfyUI; generation times were 16 seconds and 19 seconds respectively, with VRAM at 21GB for both. The author feels that for T2I, P-H3-Image is more to their taste, while for image editing, MiniMax-H3-Image is more to their taste, and notes that the speed-up LoRA seems useful in high-resolution generation.

Nov 11Wed
  1. Reddit StableDiffusion · 创作者与工作流分享75

    PAN 2 released: open-source 29B video generation model

    PAN 2 released, an open-source 29B latent diffusion model supporting video generation up to 20 seconds, 720p, 24fps, under Apache 2.0 license. The publisher also released training code and provides integrations with Diffusers, FastVideo, and vLLM. The model enables an LLM to generate dense text states/scripts, timing events and actions, with camera nested under events but decoupled from actions.

    Why it matters: Open-source 29B video generation model and training code released, allowing creators to evaluate its 20-second 720p generation capability and LLM-driven script-level controllability.

Nov 6Fri
  1. Reddit StableDiffusion · 创作者与工作流分享64

    P-JiT 21K's ComfyUI custom node released, 2B pixel-space text-to-image model can be used directly.

    The author released a ComfyUI custom node for P-JiT 21K, enabling this 2B-parameter pixel-space text-to-image model to run directly in ComfyUI. The node requires no pip install, provides bf16 (4.4GB) and int8 (2.3GB) weights, int8 is about 1.8x faster, and can run with a minimum of 4GB VRAM. It includes text-to-image, prompt upsampling, and refinement workflows; for refinement, denoise 0.1-0.2 is recommended.

Nov 4Wed
Oct 28Wed
  1. Reddit comfyui · 创作与工作流分享68

    PrunaAI releases a 6-step LoRA adapter for Qwen-Image-2.1, without requiring an upsampling step.

    PrunaAI releases a 6-step LoRA adapter for Qwen-Image-2.1, compressing image generation from 40 steps to 6 steps. The 6-steps-upsampled version removes the separate prompt upsampling step, saving 41 to 80 seconds depending on prompt style. The LoRA is loaded directly onto Qwen-Image-2.1, and the pipeline remains unchanged.

    Why it matters: The new adapter compresses image generation from 40 steps to 6 steps and removes the upsampling step, allowing creators to assess whether to replace their existing generation workflow.

Oct 26Mon
  1. AITNT · AI创作资讯68

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

    The Microsoft team proposes the 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, with preview rendering taking less than a minute.

    Same story, the pick is 《Microsoft proposes Code Video Model: controlling video generation with executable code》

Oct 25Sun
Oct 24Sat
  1. Reddit StableDiffusion · 创作者与工作流分享38

    Qwen-Image 2.1 VAE 8px grid fix fine-tune released

    To address the 8px grid issue in the Qwen-Image 2.1 VAE, the community released a Texture-Fix fine-tuned version that trains only up_blocks.2/.3, suppresses the 8px peak while improving sharpness, and lets subsequent convolutional layers prevent the pattern from entering RGB high frequencies. The model is now on Hugging Face, with both original and ComfyUI versions available.

Oct 23Fri
  1. 机器之心SOTA模型 · 公开公众号 RSS38

    Canva releases the Craft image generation model; Shanghai AI Lab open-sources the ShuSheng-A1.5 research agent

    Canva released the Craft image generation model on October 22. It uses a two-stage process—composition planning first, then pixel rendering—supports 2K output, more than 30 aspect ratios, and in-image text in 55 languages, and is available through Canva Dream Lab and a public API.

  2. MindStudio Blog · 创作分享72

    Fish Audio releases the Drama 3 speech model: controlling performance with parenthetical stage directions

    Fish Audio released the text-to-speech model Drama 3, allowing users to write stage directions in square brackets within scripts to adjust tone, pacing, and performance. It also supports word-level control, inline sound effects, and multi-speaker scripts. The model is released as open weights with 36B parameters on Hugging Face (fishaudio/drama-3). According to user reports, full-precision operation requires approximately 46GB of VRAM, while the community 4-bit quantized version requires approximately 32GB.

    Why it matters: Fish Audio open-sources a 36B speech model, using bracketed stage directions to control performance, and provides benchmark differences compared to ElevenLabs along with self-hosting VRAM requirements.

  3. MindStudio Blog · 创作分享75

    LTX 2.5 Open Weights Released: 22B Video Model Available for Self-Download and Operation

    Lightricks released LTX 2.5 open weights on Hugging Face. This is a 22B-parameter diffusion model that can be downloaded and run on your own GPU. It also provides a full dev version and a distilled accelerated version, along with bf16, int8, and nvfp4 quantization formats.

    Why it matters: Lightricks open-sourced a 22B video model along with audio and multi-shot capabilities. The article breaks down the components of the download package and the self-hosting requirements.

Oct 22Thu
  1. Reddit StableDiffusion · 创作者与工作流分享58

    MOSS-Image-Edit LoRA released, compatible with Base and SFT and accompanied by ComfyUI nodes

    The editing LoRA for MOSS Image has been released, compatible with both the Base and SFT versions, with weights hosted on Hugging Face. The author also provides links to ComfyUI custom nodes and workflows, and says they implemented a native version with caching to speed up inference; Kijai has submitted a PR to ComfyUI, so native support may follow.

  2. 量子位 · 搜狐公开专栏75

    Moss Intelligence's open-source MOSS-Image surpasses 1.3 million overseas downloads in 4 days

    Moss Intelligence's open-source 5B-parameter text-to-image model MOSS-Image gained 1.3 million downloads in overseas communities within 4 days of launch, received 1.4k Stars on GitHub, and topped the Hugging Face text-to-image trending list.

    Why it matters: The 5B-parameter open-source model gained 1.3 million downloads in 4 days of launch; the community has already produced accelerated and quantized versions and LoRAs, allowing observation of its actual performance in text rendering and layout control.

Oct 21Wed
  1. Reddit StableDiffusion · 创作者与工作流分享58

    MOSS-Image visual text encoding ComfyUI node and workflow released

    The author developed a custom ComfyUI node for MOSS-Image that feeds reference images directly into the Qwen3.5-9B vision tower and keeps image embeddings in the text conditioning, so images participate in generation as part of the prompt rather than being input as VAE latents. The node and workflow are open-sourced on GitHub, used together with the MOSS-Image-SFT-8-steps-dmd LoRA on Hugging Face.

  2. comfyorg · YouTube 创作视频72

    ComfyUI supports Hao Ai lab's FastH3 OmniRef model

    ComfyUI announced support for Hao Ai lab's FastH3 OmniRef, a model aimed at fast, high-quality video and audio generation. OmniRef enables support for multimodal reference materials, so creators and teams no longer rely solely on text prompts to control generated content; the model can run locally, allowing teams to generate and retain materials on their own hardware without relying on hosted APIs. The official announcement also provides a cloud entry point and a Hugging Face open-weights link.

    Why it matters: ComfyUI integrates FastH3 OmniRef, supporting multimodal reference materials and local operation, so creators can assess building their own video production pipeline accordingly.

  3. Reddit comfyui · 创作与工作流分享78

    Hao AI Lab open-sources FastH3 OmniRef, with ComfyUI support on day one

    The FastVideo team at Hao AI Lab has open-sourced FastH3 OmniRef, which generates video with synchronized audio from a prompt plus reference images, short video clips, or sound. The model needs only 8 steps instead of 50, is up to 11.9 times faster than the base MiniMax-H3, has entered the ComfyUI 0.42.2 workflow templates, and has open weights and training code.

    Why it matters: FastH3 OmniRef is released as open weights, produces video in 8 steps, and supports day-one ComfyUI integration, so creators can use it to evaluate local reference-to-video pipelines.

Oct 20Tue
  1. comfyorg · YouTube 创作视频72

    ComfyUI supports the MOSS-Image image model

    ComfyUI announces support for OpenMOSS's open-source image model MOSS-Image, a 5B-scale model that can render clear, legible text and accepts structural layout prompts. The official workflow has been released on Hugging Face, aimed at image generation scenarios requiring layout control such as posters, advertising, and localized creative work.

    Why it matters: ComfyUI integrates the open-source 5B image model MOSS-Image, whose text rendering and structural layout control offer direct reference value for poster and advertising image generation.

  2. SuSu_酥酥60

    Ant Ling opens sources the Ming-3.0 speech series of 5 models

    Ant Ling has open-sourced the Ming-3.0 speech series, comprising 5 models in total, covering speech synthesis, audio understanding and dialogue, and speech recognition, under the MIT license and already available on Hugging Face and ModelScope. TTS supports describing timbre in text and can control emotion, stress, and rhythm through director mode; ASR supports dialects, hotwords, and context awareness. The author provides getting-started steps for downloading weights, cloning the TTS repository, and running inference per the README, among which the 0.5B version is the lightest.

  3. 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.

  4. 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 18Sun
  1. 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.

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

    A less-restricted LoRA trained on Qwen-Image-2.1 that supports all-around poses

    A creator released a rank-64 LoRA fine-tuned on Qwen-Image-2.1 that can generate almost any pose, covering prompts and compositions the base model struggles with. The LoRA works for both text-to-image and image editing without a separate adapter; tested parameters are 25 steps, CFG 4.0, LoRA strength 1.0. It is still v0.1 with no formal benchmark yet, and the weights and inference code are public.

  3. Reddit comfyui · 创作与工作流分享12

    Is there a ComfyUI extension that can download and install models and LoRAs directly from Hugging Face or Civitai?

    A user posted in the ComfyUI subreddit asking whether there is an extension in ComfyUI Manager that avoids terminal operations and lets you directly paste Hugging Face, Civitai or GitHub links to download models, LoRAs and checkpoints and automatically place them in the corresponding directories. The user also hopes for support for entering API tokens for authentication, and for managing and deleting downloaded models within the UI, achieving one-click model download and management.

Oct 17Sat
  1. 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.