free-short-video v7.2.2 released: Docker multi-architecture images and data persistence notes
free-short-video v7.2.2 is released; the Docker image supports linux/amd64 and linux/arm64 and has been pushed to GHCR and Docker Hub.
free-short-video v7.2.2 is released; the Docker image supports linux/amd64 and linux/arm64 and has been pushed to GHCR and Docker Hub.
AIArtEngine released version v8.4.3, with a full changelog covering v8.4.2 to v8.4.3. The original text did not disclose specific feature changes or parameter details.
ComfyUI-Spectrum-MOSS updated to v0.8.1, rewriting the ER-SDE bridging logic and adding an euler/er-sde sampler settings guide; a user measured about 20% speedup on a 3060 12Gb.
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.
Video Automation v0.2.14 has been released, fixing security issues with certificate dependencies and desktop builds. This version was completed by @1076184145 in commit #21; see v0.2.13...v0.2.14 for the full changelog.
The open-source creative tool released v1.6.9.1, optimizing the canvas empty-node ratio toggle animation, floating layer drag positioning, and connection hit detection, and making quick creation results display stably by resource identity to avoid video reloads and image flickering. In the canvas Agent, Pi can perform one limited continuation run for unfinished plans, and video polling continues querying the original task when the connection is interrupted. Added Cheaper Inference Chat and MaxForAI video declarative protocol plugins; the database schema is upgraded to 59.
AIArtEngine released version v8.4.2, with a full changelog covering the changes from v8.4.1 to v8.4.2. The original text did not disclose specific features, parameters, or performance data.
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.
SmartSub v3.11.0 upgrades "Advanced Options" to "More Settings," split into "Common" and "Expert" pages, with expert items such as decoding and fine-tuning of sentence segmentation moved to the Expert page. The font selector is grouped by recently used, common fonts, and all fonts, supports searching by Chinese name and pinyin, and the synthesis preview will flag missing characters and recommend available fonts.
DramaClip v0.2.2 is released, adding three toolbox capabilities: video copy extraction, speech-to-text, and text-to-speech, and includes a built-in HTTP gateway with local loopback only plus token authentication for external Agent calls.
The user trained a detail and texture LoRA "RAW Upscale MOSS" for MOSS-Image using LoRA Dataset Studio, trained on MOSS-Image-Base for 2000 steps, rank 16 / alpha 16, 1024 resolution, finally exported as FP16 SafeTensors, and has been made freely available for download on Civitai.
ComfyUI workflow templates released v0.11.94, installable via pip install comfyui-workflow-templates==0.11.94. This version adds three new templates: Mirelo SFX, VEED, and HeyGen Voice 1.0 Turbo.
The Vathography author released the ComfyUI-Marigold-V2-Tiled-Depth-Vathography node, which can estimate depth maps from high-resolution photos. The node is free, its output is commercially usable, and an example workflow is included. The author says the node should run on an 8 GB GPU plus about 24 GB of system memory, pending confirmation, and is soliciting hands-on feedback from different GPUs.
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.
Stability Matrix v2.16.5 adds Fizgig, a LoRA training studio for Flux 2 Klein 9B, Krea 2, MiniMax H3, and Qwen Image 2.1, integrating dataset preparation, training, LoRA repair, and extraction.
ai-video-editor is a browser-based, local-first open-source video editor that combines a multi-track timeline with on-device AI inference, completing editing, subtitles, and matting locally. The tool integrates Whisper small for audio-to-text with waveform-aligned subtitles, uses MODNet and YOLOS for portrait matting and subject avoidance, and includes MI-GAN for removing watermarks and clutter plus NanoVSR for 4× upscaling.
free-short-video v7.2.1 is released, offering linux/amd64 and linux/arm64 multi-architecture images, already pushed to GHCR and Docker Hub.
ComfyUI Workflow Templates released v0.11.93, now available via PyPI and installable with pip install comfyui-workflow-templates==0.11.93. This version's version number update was completed by @comfyui-wiki in #1370; see v0.11.92...v0.11.93 for the full changes.
Developer Qianshuixia (qianshuixia) has open-sourced shrimpreel, a SaaS site template for "generating viral-template follow-shot AI videos from a single photo," where each trending keyword can be replicated into an independent site that goes live after changing the configuration and assets.
mlx-audio released v0.6.1, fixing the prompt contract for Ming dense and supporting Ming-3.0-tts-0.5B, while adding the Gradium Phonon TTS model. This version also fixes server-side STT uploads to use writable temporary files, and fixes a crash in irodori-tts where speaker_kv_scale rolled back under independent CFG and dual-conditioning settings.
The open-source tool Pictide can automatically combine tag-based prompts and send them to ComfyUI; as users browse the results, they give feedback via likes or dislikes, and a local recommender learns their preferences to guide subsequent prompt combinations. Prompt generation, preference learning, and recommendation all run locally, and images and feedback are also stored locally. It is currently best suited to models that respond well to Danbooru-style tags. Users can also anchor a favorite element (such as "forest") and continue exploring different combinations around it.
AI Canvas v0.10.12 optimizes the large-canvas save logic: when returning to the launch page or closing the app, it prioritizes saving the main content, and opening the project library no longer repeatedly saves the full canvas. The launch page adds a general assistant that supports Q&A, writing, translation, and analysis of files and reference images, and can add external Streamable HTTP MCP services with per-tool authorization. ComfyUI now uses a native connection to receive real-time progress, distinguishing queuing, model loading, and actual execution; for macOS, this version provides only an Apple Silicon installer.
The open-source short drama generation tool agnes-video-generator released v7.2.0, adding an external material mode alongside automatic and manual modes; at each checkpoint it generates creative prompts that can be pasted directly into Dreamina (即梦), and the video checkpoint prompt is a shot-by-shot script numbered by scene.
The author disclosed a set of parameters for training a MOSS-Image character LoRA locally on 8 GB of VRAM: the default setting trains for 800 steps, extended to 1200–1500 steps for large datasets or difficult concepts, with a learning rate of 0.0004, rank/alpha 8/8, and 512² resolution.
Why it matters: The author provides the complete training parameters and step range for a MOSS-Image character LoRA, and explains why 8000 steps would overfit, serving as a starting reference for local training.
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.
DramaFlow Studio has released version v0.1.15, with the full changelog covering v0.1.14 to v0.1.15. The original text does not disclose specific feature changes, parameters, or availability details.
ComfyUI workflow templates has released v0.11.91, installable via pip install comfyui-workflow-templates==0.11.91. This version marks the new partner node templates as local-only.
OpenScreen released v2.1.2-rc.1; compared with v2.1.1, the change is that installs driven via OpenScreen MCP now count toward active statistics. This version is a release candidate; the full changelog is available in the comparison from v2.1.1 to v2.1.2-rc.1.
AIArtEngine released v8.0.0, with the full changelog covering changes from v7.8.0 to v8.0.0. The original text did not disclose specific features, parameters or availability details.
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.
mlx-audio released v0.6.0, adding support for the Fish Audio Drama 3 speech model, Phonon-2 speech recognition and MiMo-V2.5-ASR STT, plus Granite Speech 4.1 Plus. This version improves Breeze-TTS and VibeVoice generation performance, and fixes Qwen3-ASR streaming timestamps, Qwen3-TTS prefill, and the fallback when Metal is unavailable.
ComeCut v0.2.1 released its first Windows desktop version, only about 30MB in size, with fast startup and low resource usage. It adds a multilingual interface, automatic dubbing (supporting Chinese, English, Korean and Japanese, requiring the corresponding TTS), and auto-save when exiting the editing editor, and fixes the issue of audio being lost after separation, while displaying vocals and background audio in different colors.
ComfyUI v0.42.0 was released on October 19, officially supporting OpenMOSS's 5B image model MOSS-Image, and adding Qwen 3.x series performance improvements Part 2.
The author got TRELLIS.2 running normally in ComfyUI on an AMD RX 9060 XT 16 GB (gfx1200) under Windows 11 using ROCm, and put the fixes and workflow in the GitHub repository ovaoz/trellis2-amd-windows.
InvokeAI 7.0.0 alpha 3 has been released, adding role-based board sharing, support for the Qwen-Image-2.1 image model, and the SageAttention acceleration backend for NVIDIA GPUs. This version speeds up full-image generation at 2048² resolution on an RTX 4090 by 15–35% and video generation by 12–35%. Note that v7 performs a one-way upgrade of the v6 database, and the official recommendation is to install and test with a brand-new root directory.
The author has released the open-source project freavy, a free local-first node canvas running on ComfyUI for AI image and video generation, with assets and rendering done on your own GPU, requiring no subscription and uploading no content.
The open-source AI mod NPCTalk brings script-free real-time dialogue to GTA 5: players can talk to NPCs via microphone, and NPCs respond with AI-cloned original voices. The system runs on NVIDIA Nemotron ASR, Qwen 3.8 27B (Cerebras platform), Fish Audio Drama 3, and TypeSafe Jev working together, with push-to-talk latency of about 1.15 to 2.2 seconds.
The developer released the open-source TypeScript library tscomfy, which reads the ComfyUI server's /object_info to generate types for installed nodes (including custom node packs); workflows are written as code, and wiring errors, misspelled checkpoints, or missing inputs are reported in the editor, with autocompletion for models and samplers.
LichtFeld Studio 0.5.4.1 is released, supporting dragging splats from the scene graph into folders or directly into Blender as PLY/SPZ, with the up axis selectable in preferences. This version adds training session resume, batch video frame extraction at the source frame rate, and numeric position, rotation, and scale fields for crop boxes and ellipsoids. It also fixes image loading when no GPU JPEG decoder is available, Python console freezes, and VRAM exhaustion when merging large compositions.
The developer released the open-source Android client Comfy Workbench v0.3.0, which lets users import and run ComfyUI workflows on their phones. Users only need to start ComfyUI with --listen 0.0.0.0 and enter http://:8188 on the same Wi-Fi to connect, with no additional software required; remote access is optional.