Fizgig 9.2 was released on November 8. The time to train MOSS-Image LoRA on an RTX 5090 (94 images, batch 1, Muon, 70 epochs) was shortened from 49 minutes to about 20 minutes, and memory usage dropped from 21.7 GB to 7.1 GB.
GLITCH Studio is recruiting 1-2 technical beta testers to test its ComfyUI-based local AI filmmaking Windows application. The application supports workflows such as WAN, MiniMax H3, Krea 2, LTX 2.5, and FLUX, and has completed clean machine installation tests for FLUX image editing and LTX 2.5 video rendering.
MiniMax H3 Novel Pipeline is preparing to release version 2.5.1, refactoring the ComfyUI node structure and improving the character profile editing process. Updates include H3 reference budget (9 images, 3 audio files), 4-15 second scene limits, spatial continuity JSON Schema validation, and OpenAI/LM Studio backend configuration. Node logic has been separated into src/nodes/ and src/core/ for easier future maintenance and expansion.
Developer BittiAI built slopfab, a native command-line tool that lets coding agents directly generate media files, supporting models such as MiniMax H3, Krea 2, MOSS, Fibo 2.0, YuE2, and MiniMax Music3, covering video, image, and music generation. The tool can run locally on Windows/Linux, supports CUDA/Vulkan, requires no cloud or account, and can be executed remotely via SSH.
MiniMax H3 Novel Pipeline is preparing to release v2.5.1, refactoring Provider configuration and adding an OpenAI backend. Character profile integration supports editable v2 profiles, with H3 reference budget limited to 9 images, 3 audio clips, and 4–15 second scenes. The spatial continuity JSON Schema has been bundled and is used for runtime validation.
Timeline Studio v1.0.10 has been released, adding six editable beat templates and WebMCP integration. The templates include a three-panel beat montage, rotating cards, a seven-panel intro, and more. For playing cards, you can assign separate images or videos to each of the four suits. WebMCP allows agents to discover templates, bind assets, preview changes, and apply or revert them. It also updates video proxy editing, split-screen alignment, and exporting selected clips individually.
ComfyUI Comic Creator v1.69.1 update fixes a crash when generating with linked OCR and MOSS-Image tabs. It also resolves VRAM insufficiency caused by the OCR Chat TE function, adds bbox drawing selection and Ctrl+Z undo deletion, and updates the manual to the latest version.
drama-skills v0.11.2 streamlined and deduplicated the full set of skill instructions (12 public creation and maintainer skills, totaling 154 Markdown files), reducing the total line count from 20,428 lines to 8,569 lines. Entry format boundaries were clarified, specifying second-level heading category terms, the list of required fields, the prohibition of --- separators inside blockquotes, and the requirement to paste the text-to-video 'static visual anchor' verbatim. No migration is required; see CHANGELOG for the complete changes.
ComfyUI Comic Creator v1.68.1 and VRM Pose Studio v0.2.0 updates strengthen integration; VRM Pose Studio now displays areas outside the preview output size as semi-transparent.
SnapRef v2a released as a stable version of the Anima reference adapter, adding per-reference weights and six workflows; evaluation is roughly on par with v1a. Installation requires placing snapref_v2a.safetensors into models/loras/ and using the SnapRef node instead of LoraLoader; the FP8 file is 195 MB, and under a 64-case two-seed evaluation, appearance label recall is 0.8722 vs 0.8727.
Why it matters: SnapRef v2a adds per-reference weights and six workflows, with evaluation on par with v1a; creators can adjust character consistency as needed based on this.
IV Studio 1.0 released, free and open-source ComfyUI film studio adds Linux/Mac support, and includes one-click installation of missing nodes, local character LoRA training, and 154 MCP tools.
Why it matters: Version 1.0 adds Linux/Mac support, one-click installation of missing nodes, and local character LoRA training; 154 MCP tools make the ComfyUI film workflow more complete.
Remiqora Video releases a free and open-source video studio based on ComfyUI. Enter a sentence or a song and it generates a complete video with sound on a local PC. Supports 15 seconds to 10 minutes, landscape/portrait/square aspect ratios. Songs can have lyrics automatically recognized and be cut to the beat. Characters and scenes remain consistent across shots. Individual scenes can be re-shot separately, and bad shots are automatically detected and re-shot.
Why it matters: Free, open-source, and runs locally. It wraps a ComfyUI workflow into a complete application that produces a video from a single sentence. Supports single-scene re-shooting and character consistency, and can evaluate local video production efficiency.
The author has developed a FilmDateStamp custom node for ComfyUI that simulates the optical burn-in and emulsion halo effects of real film date stamps, replacing simple digital text stickers. The node includes 4 hardware-accurate fonts and 12 era presets, and can be installed via ComfyUI-Manager with no heavy dependencies.
JJYB AI VideoAutoCut v3.4.5 released, fixing mixed-cut duration precision and task result isolation issues. Full pytest 62 items passed, real FFmpeg normal/beat-point each 120 segments measured 602.033 seconds (target 602.040 seconds). Source package released; after upgrade, re-run the mixed cut.
drama-skills v0.11.1 rewrites the technique explanations and positive/negative examples for source material analysis, story development, script writing, visual design, storyboarding, and prompts into plain language that ordinary creators can directly follow, and fixes two issues in the source material chapter index and episode map template. This update adds two directions for adapting novel psychological narration into short dramas, criteria for satisfying moments in non-face-slapping genres, and clarifies that light comedy interruptions, deadpan freezes, and breath-holding stillness are not transition empty shots that should be deleted.
Kling 4.0 launches Keyframes feature, supporting up to 10 keyframes to guide scene unfolding, generating from storyboard to video in one go, now available for trial on Magnific.
Why it matters: The new feature supports up to 10 keyframes to guide scene generation, allowing creators to more precisely control the narrative pacing of the video.
TBG has integrated the region-guided RevealLayer layered decomposition method into ComfyUI and released it as a custom node. Unlike Qwen-Image-Layered's automatic layering, RevealLayer uses bounding boxes to guide regions and estimate occluded content. The author provides three workflows: manual, semantic, and SAM. The author recommends refining layer by layer after extraction; single elements can be upscaled to 4K, and larger sizes can be tiled and upscaled using TBG ETUR.
A C++ developer has released a stable Beta version of ComfyUI's native desktop client, using a C++ core to bypass the browser and Litegraph architecture, supporting smooth rendering of thousands of nodes, and extending frontend nodes such as reroutes and bookmarks. The source code and Windows precompiled binaries have been made public, accompanied by a demo video.
Taiwanese tech company imCore Corp. launched the AI film production platform Core Studio at the Taiwan Creative Content Fest in Taipei, equipped with 24 professional AI agents as a virtual production team.
AI Canvas v0.10.29 released. The script workbench adds intelligent collaboration, story relationship graph, and original text writing, supporting multi-model team collaboration. Original text and episode scripts can be saved as TXT within the project. The story relationship graph supports filtering by episode and type, linking official assets, and viewing original text evidence; intelligent scheduling supports multi-model team role division, dependency handoff, independent review, and budget control.
Dials encodes a pair of prompts into an approximately 14KB direction control file that can adjust the generation direction on MiniMax H3 and Pika H3 Flex like a slider, without requiring training, reference images, or additional rendering.
Why it matters: The approximately 14KB prompt direction file can adjust the motion or style of MiniMax H3 and Pika H3 Flex without training, complementing LoRA.
ComfyUI new node Tab Capture can view or capture real-time video, audio, and images from other browser tabs as workflow reference material. It supports single-frame or batch capture at a set fps, can include tab audio, and outputs the actual fps to ensure correct playback speed. Chrome/Edge works best; ComfyUI needs to run on localhost or HTTPS; DRM sites like Netflix capture as a black screen.
Xiaoying Agent v1.8.2 released, adding text thinking level configuration, supporting different parameter formats for OpenAI, Anthropic, Gemini, DeepSeek, Qwen, etc. Fixed issues such as Anthropic 400 error and truncated streaming responses; database schema upgraded from 89 to 91. Short drama production drafts can be resumed for editing, and login and registration are merged into a single form.
JJYB_AI_VideoAutoCut v3.4.3 update: Fixed the mixed-cut 'highlight segment merge failure'. The quick-cut merge has been refactored into 'per-segment normalization + concat stream-copy concatenation'. Each segment is individually normalized to unified parameters; segments without an audio track automatically have a silent audio track added. On failure, the specific ffmpeg error summary is passed through. End-to-end testing successfully merged 4 heterogeneous segments; 32 pytest cases passed. Licensed for personal learning use only; commercial use prohibited.
Blackmagic Design releases DaVinci Resolve 21.2, adding support for Apple MacBook Pro PROLED XDR touchscreen, introducing Gen 6 color science LUT, and claiming rendering speedup up to 5x on M5 Pro/M5 Max.
free-short-video v7.2.5 released, providing multi-architecture Docker images and npm installation. The new version adds fast retry mode, retry count and logging, real-time polling status, and fixes the issue where video tasks get stuck waiting after restart. Compatibility remains unchanged; existing configurations and APIs require no changes.
Minimax H3 Temporal Spatial node update, supporting high-resolution video-to-video workflows; the author tested successfully at 2.5 MP on an RTX 4070 Ti with 12GB VRAM. The node splits generation into blocks and temporal windows, maintaining animation transfer with no offset and no ghosting, and can also be used for second-stage upscaling, but the author does not recommend it for full generation in the first stage.
InfiniteCreation v1.0.8 update: added custom pipeline orchestration for the Skill used by each node, and supports import/export sharing. Also enhanced subtitle recognition, optimized generation of character clothing/age assets in materials, and fixed some bugs in storyboard review and projects. The Windows x64 installer InfiniteCreation-1.0.8-win-x64-setup.exe is available for download.
Tencent's AI content creation platform WorkRally has completed an upgrade, moving the Agent creation entrance to the homepage, launching features such as AI script generation and diagnosis, smart editing, and copyright deposit, and introducing three vertical models for the video industry.
Why it matters: Tencent has consolidated script, generation, editing, and copyright deposit into a single platform, allowing individual creators to evaluate how full-process tools optimize production costs.
drama-skills v0.11.0 released: benchmark decomposition can be restored into word-for-word dialogue scripts, unified character scenes across episodes, and shot-by-shot storyboards, and adds subtitle, audio, and shot segmentation skills. Tested on the first 12 episodes of a 3D animated drama, the dialogue character error rate was 0.34%; issues with same room, same voice, and shot transitions all have corresponding solutions.
Why it matters: The new version restores benchmark decomposition into generatable projects and provides practical approaches for same room, same voice, and shot transitions, which creators can adapt to their own production workflows.
Juliu Studio v0.1.22 was released on November 6, containing 2 commits to the main branch, providing 3 resource files, see the v0.1.21...v0.1.22 changelog for updates.
Open-source creation tool open-ai-canvas v1.8.1 released; database Schema upgraded from 87 to 89. Video models add a reference-video billing switch (off by default); when billing by second or token, reference video duration/usage is counted; short drama production supports selecting video models that require reference images. Before upgrading, back up the database, and update Web and Backend together.
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.
comfyui-workflow-templates v0.11.103 released to PyPI, installable via pip install comfyui-workflow-templates==0.11.103. New partner templates added: GPT Image 2.6, Gemini Omni 1.2 Flash, Pika H3 Flex, and Pruna upscale.
World Labs releases Atlas real-time research preview, demonstrating via Frameboy streaming frame generation from cloud GPU to browser at up to 15fps. This version highly optimizes Atlas while maintaining novel view synthesis capability; it is also integrated into Viewfinder, allowing users to build scenes frame by frame and then explore in real time.
Why it matters: Real-time streaming generation reduces exploration cost to near zero, enabling creators to assess the potential of interactive world models to replace existing workflows.
V-Ray 7 update 5 released, adding AI assistant and MCP server, which can search for or generate 3D assets from a single reference image, and match scene lighting and atmosphere.
Google Gemini Omni 1.2 Flash is now available in Runway. The model supports creating and editing videos up to 15 seconds long, with more natural dialogue, stronger shot-to-shot consistency, the ability to refine scenes without restarting, and upscaling to 4K.