Open-weight video & audio models
Video and audio models you can download and run yourself instead of renting through an app: LTX, Wan, MiniMax H3, HunyuanVideo-Foley, ACE-Step and friends. Every claimed speed comes from the project's own README unless a Lab post says otherwise.
Lab posts (49)
- Apple FastFrames: open research video models that generate 480p frames in under 50 ms
Apple researchers released FastFrames, two open-weight text-to-video models (1.3B and 14B) built to generate frames fast enough to respond live to keyboard or drag input. They're research-only and output 480p, so this isn't a production tool yet, but it's a clear look at where real-time previs and interactive world models are heading.
- LTX pits Alpha Gen against Bria and CorridorKey, Bria RMBG 3.0 lands in ComfyUI, HyperFrames starts a YouTube channel, and Creatify launches Max
Matting was the theme of today's official posts. LTX ran its Alpha Gen against two other tools on the same footage, and ComfyUI added Bria's new background remover. HyperFrames launched a weekly YouTube show, Creatify introduced an always-on ad agent, and Higgsfield's $1M contest closes Nov 26.
- Markerboard draws whiteboard explainers to your voiceover, plus a Blender face-capture stack, ComfyUI blockout tips and Kevin Ngo's JavaScript drawing prompt
A brand-new open skill turns a topic into a marker-on-whiteboard explainer timed word by word to the narration. Also today: a Claude-scripted Blender face-capture stack, ComfyUI's advice on how detailed a blockout should be, the prompt behind Kevin Ngo's viral JavaScript drawing, and LTX's write-up of its SDR-to-HDR research.
- Three open-source video tools picking up steam: Mago Studio, Rigby2D and a MiniMax H3 'short-drama workbench' for ComfyUI
None of these is a launch-day headline, but all three are open source, being actively updated and doing real production jobs: a local-first AI video editor that renders through ComfyUI, a free 2D rigging and animation tool with audio lip-sync and an MCP server, and a ComfyUI pack for long, segmented MiniMax H3 videos.
- MiniMax H3-Image: open-weights 16 MP image model from the H3 family, with 16 references and day-one ComfyUI support
MiniMax released H3-Image on Nov 10: one open-weights checkpoint for text-to-image and editing, built on the MiniMax H3 video model without its audio layers. MiniMax says it makes about 16-megapixel images in roughly 10 seconds and takes up to 16 reference images. It is on the MiniMax API, MiniMax Code and Hailuo AI, and ComfyUI added native support the same morning. For video makers it is a keyframe and reference-still tool that shares H3's look.
- ComfyUI v0.45.0 for video makers: RelayVSR video upscaling, H3-Image, Seedream 5.5 and Breeze TTS 2 nodes
ComfyUI tagged v0.45.0 on Nov 10. For video work, the headline is native RelayVSR, a 2x/4x video upscaler that runs a big model only on keyframes and a small decoder on everything in between. The release also adds MiniMax H3-Image, Seedream 5.5 and Breeze TTS 2 nodes, plus the Kling 4.0 Flash, Seedance 2.5 4K and Gemini Omni 1.2 Flash partner-node updates from last week.
- #OpenSourceHiggsfield winners: Panorama Stickers, Blocksfield, Imagination Studio and Paint a Plate, plus 10 open builds worth a look
Comfy announced the #OpenSourceHiggsfield winners on Nov 10, as scheduled: more than 300 builds from 45+ countries, all open source. Best Overall went to nomadoor's ComfyUI Panorama Stickers. Below is what each winner does, plus the honorable mentions that matter most if you make video.
- LTX puts its VFX tools on the API: Alpha Gen mattes, SDR to HDR, Reframe and 8K Upscale, with ComfyUI partner nodes
On Nov 10 LTX made four VFX jobs available as API endpoints: Alpha Gen (a matte from ordinary footage), SDR to HDR, Reframe (extend the scene to a new aspect ratio instead of cropping) and Upscale to 1080p, 4K or 8K. They are also on fal and Griptape, and ComfyUI merged partner nodes for all four the same afternoon.
- HyperFlow H3 Flash: Video Rebirth's 14B, 6-step MiniMax-H3 that makes video and sound on one consumer GPU
Video Rebirth compressed the ~31B MiniMax-H3 transformer to 13.7B, cut sampling to 4 to 8 steps without CFG, and released it with ComfyUI templates, GGUF quants down to 9.1 GB and, as of Nov 9, official single-file INT8 checkpoints for ComfyUI. It makes 5 to 15 second clips with a synced stereo soundtrack. Weights carry the MiniMax H3 community licence, which needs MiniMax's permission in the US, EU, UK and South Korea.
- Freefield: an open-weights Higgsfield-style studio on ComfyUI (120 apps, Modal or your GPU)
Freefield rebuilds a Higgsfield-like image/video/VFX/audio studio as ComfyUI workflows with open weights only: each app is a folder with an app.json and an API-format graph. The README lists about 120 apps, lab GPU runs with time and peak VRAM, and a path to run on Modal or on your own ComfyUI. Around 973 GitHub stars as of this morning.
- Viggle Remake: open H3 LoRAs that re-shoot a clip (swap or edit) plus an agent MCP
On Nov 4 Viggle shipped Remake: a video transformation agent with open weights. Per Viggle, two MiniMax-H3 LoRAs (swap and edit) keep the original motion, camera, timing and soundtrack while you recast people, swap objects, or change background, weather, light or style. Hosted app from about $0.96 per 10 s at 720p (their figure); ComfyUI and MCP paths for local/agent runs.
- MoneyPrinterTurbo is trending again: topic in, finished short out
The 2024 short-video generator by harry0703 is back on GitHub's daily Trending lists (all languages and Python) on Nov 4, a day after @RoundtableSpace featured it. It's MIT-licensed, sits near 139K stars, and has quietly grown from a stock-footage slideshow maker into a front end for Seedance, MiniMax H3 and a dozen voice engines.
- ComfyUI lands in ChatGPT, and v0.44 adds ProRes saving and MiniMax H3 edit nodes
On Nov 4, Comfy announced that Comfy Cloud is now an official ChatGPT plugin, so you can run ComfyUI image, video, audio and 3D generation from a chat without installing anything or owning a GPU. The v0.44.0 core release that shipped the night before adds ProRes output and a batch of video nodes.
- #OpenSourceHiggsfield closes: five last-hour ComfyUI builds
Entries for ComfyUI's #OpenSourceHiggsfield challenge closed at 9 AM PT on Nov 3, and the hashtag filled with last-minute submissions. Winners are due Nov 10. These five stood out for working video makers: a 3D previs tool, a kinetic-type engine, a MiniMax H3 film studio, a song-sectioned music-video tool and a take-based shot manager.
- P-H3-Image: Pruna turns MiniMax H3 into an 8-step image model
Pruna released P-H3-Image on Nov 3: an open-weights text-to-image and editing model distilled from the MiniMax H3 video model. ComfyUI supports it natively the same day. For video makers, it's a way to make key frames and reference stills that share H3's look before you animate them.
- #OpenSourceHiggsfield runs to Nov 3, winners Nov 10: four LTX builds to know
Comfy pushed the #OpenSourceHiggsfield deadline to Nov 3 at 9 AM PT and its challenge page now lists winners on Nov 10, not Nov 6. Over the weekend LTX picked out four entries built on LTX-2.5. Here's what changed and what each build does.
- Pika H3 Flex: open six-step adapters for MiniMax H3, plus audio stems
Pika released two open adapters that run MiniMax H3 in six sampling steps, with ComfyUI templates the same day, and put 1080p to 4K, audio stems and a prompt rewriter behind its API. Here's what's open, what it needs, and the licence catch for US users.
- Body Double: recast the performer in a take with open weights in ComfyUI
One of the last #OpenSourceHiggsfield entries, and the one that drew the loudest reaction: Body Double swaps the person in a real take for a new character while the motion, camera, set and original audio stay. Here's how it works, what it needs, and who it's for.
- #OpenSourceHiggsfield final day: six open studios to try
Comfy's open rebuild challenge closes Friday, Oct 30 at 10 AM MDT, and the last 36 hours brought the most complete entries yet. Here are the six worth opening, what each needs to run, and the licences to check.
- Fish Audio Drama 3 final weights top open TTS arenas
Fish Audio pushed upgraded Drama 3 weights to Hugging Face this morning, and Artificial Analysis now ranks it the top open-weights model on both of its speech arenas. For narration and multi-voice scenes, it's the open TTS to try first.
- NVIDIA's Open-H3-Regenerate-2K: a three-step 2K pass for MiniMax H3
NVIDIA researchers released an open refiner that turns a cheap 768p MiniMax H3 draft into a 2K clip in three steps, and within a day it had ComfyUI nodes and slimmed-down LoRAs. Here's what it does, what it costs to run, and where it falls short.
- Higgsfield sends Comfy a cease and desist; the challenge stays up
On the eve of the deadline, Higgsfield's lawyers asked Comfy to drop the Higgsfield name from its open-rebuild challenge. Comfy's CEO posted the letter and said no. Here's what each side says and what it means if you entered.
- anymove: creature mocap from one two-view AI video
Patch's new Apache-2.0 repo turns a single AI clip into a GLB animation for a dragon, a spider, a wolf or anything else you can rig. Rehan's launch post passed 1,300 likes and 1,900 bookmarks in a day, and the curators picked it up overnight.
- Mickmumpitz's one-model H3 short film and Singularity V2
Mickmumpitz made an anime short using MiniMax H3 for everything (character sheets, turnarounds, voices, sound effects and the shots) and gave away the ComfyUI workflows. The same day, tutorials landed for Singularity V2, a community H3 fine-tune aimed at action.
- Code2Video-AA ranks 27 models on writing videos in code: Claude Opus 5.5 leads, GPT-6.1 Sol is the value pick
Artificial Analysis now runs an independent version of HeyGen's Code2Video Bench: 114 product and brand briefs, each turned into a HyperFrames composition and judged blind. It's the first public scoreboard for the question agent-video teams keep asking: which model should write the code?
- OPENjutsu: Comfy open-sources a character-swap app on MiniMax H3, and a fully local fork follows within hours
Comfy published an MIT-licensed, open take on the 'swap the person in a clip' trick that Higgsfield's Genjutsu made popular. It runs on MiniMax H3 and SAM 3.1. A community fork already runs it entirely on a 16 GB card.
- Henge: an open studio for H3 character swap, motion transfer and head swap, plus Higgsfield open-sources its 'Inside AI' short
One day after Comfy's OPENjutsu, LoRA author Alisson Anjos released Henge, an MIT app that drives his own MiniMax H3 LoRAs for swaps, motion transfer and sharpening, with an MCP server for your agent. The same morning, Higgsfield published every prompt and asset behind its new short film.
- HyperFlow Infinity: an open agent loop that keeps generating video until you stop it
Video Rebirth open-sourced a harness that makes MiniMax H3 write, watch and direct its own next shot in a loop. Their demo is a single 60-minute run. The catch is a multi-H200 GPU node and a LoRA licence that excludes the US, EU and UK without MiniMax's sign-off.
- PopTok: a fully local pipeline from one prompt to a lip-synced music video (ACE-Step, LTX-2.5, InfiniteTalk)
Give it a song and lyrics, or only a prompt. Simulated songwriter and director 'workshops' plan the piece, then local models sing it, draw it, animate it as one continuous take and lip-sync the sung lines. You need serious hardware.
- MiniMax H3's community speed stack: 4-step LoRAs, sparse attention, and FreeVideo on an 8 GB card
MiniMax added 11 community projects to its official H3 index on Oct 25: distillation LoRAs, sparse attention, reference compression, a hires-fix cascade and new places to run the model, from TPUs to Macs. The one most people will try first is FreeVideo, which per its README runs H3 in 8 GB of VRAM.
- Greybox first, render second: Blender previs is becoming the steering wheel for Seedance, Kling and H3
Several creators and tools landed on the same workflow this week: block the shot in grey geometry, let that set the camera and timing, then have a video model paint the look. One version runs on open weights (SolarRenderer on H3), one is a Blender sidebar (Scenario) and one needs no GPU (CozyClay).
- Tencent Hy-Nomas: an open interactive video world model you steer as it renders
An Apache-2.0 model that generates video one segment at a time and takes a new text prompt or camera path between segments. It ships with Cinematic and Robotics LoRAs and a config for a single RTX 4090.
- Kimi K3.1 Pro: an open-weight model pitched at Blender, CAD and Three.js scenes
Moonshot's new 2.8T-parameter MoE (104B active, 1M context) is marketed for writing and running Blender and CAD scripts and for animated Three.js scenes. Weights are promised by Nov 4.
- MDMZ's MiniMax H3 character-swap workflow for ComfyUI: SAM 3 motion guide, auto-prompting, and where it breaks
One ComfyUI graph that puts a new character into your clip with the original lighting, motion and lip sync, or flips it so your character stays in their own scene and copies the motion. Needs 40+ GB of models and a strong GPU.
- Pruna P-Video-2-Pro-Avatar: one portrait plus audio or a script becomes a talking video, built on MiniMax H3
Pruna turned MiniMax H3 into a talking-portrait endpoint. You give it one image plus your own audio or a text script, and it returns lip-synced video from 1 second to 3 minutes. It costs $0.02/s at 480p during a launch discount that runs to Nov 5, and ComfyUI says nodes are coming.
- LTX Desktop 1.3.1: green-screen-free cutouts, drawn motion paths, 4K and a planning assistant in the free local app
Lightricks' open-source desktop app now cuts a subject out of a clip with no green screen, lets you draw up to three motion paths, follows Pose, Depth or Canny guide videos, renders 1440p and 4K, and has an assistant that plans and queues jobs.
- Local AI transitions between real footage, inside DaVinci Resolve: an honest test of LTX 2.5 and MiniMax H3 Omni vs Seedance 2.5
Kevin Vandermarliere grabs start and end frames from his Resolve timeline, generates car-to-face-to-POV transitions in ComfyUI on a 24 GB RTX 3090, and shows the misses. Then he hides the leftover morphing with speed ramps and Fusion blur.
- LTX-2 1.5.0: 4K renders in about half the time, plus keyframes you can pin or merely guide
Lightricks merged a batch of performance work from doubleAI's Maxwell into the open LTX-2 pipelines. On two B200s, a five-second 4K clip drops from 250 to 134 seconds. The same release lets you mark each image either to replace a frame exactly or just to guide it, and folds keyframe interpolation into the main two-stage pipeline.
- FastH3 Omni goes open: reference-to-video with sound on your own GPU, and the case for letting Opus cut cheap clips
Hao AI Lab open-sourced FastH3 Omni, an 8-step MiniMax H3 that takes up to 12 image, video and audio references and returns 5–15s clips with synced stereo sound. ComfyUI supported it on day one.
- Epic's Unreal Engine plugin for Claude Code is trending. The video angle is Sequencer previs
Hundreds of Unreal Editor tools in 30+ toolsets, including Sequencer, Niagara, materials and Control Rigs, exposed to Claude Code over MCP. It isn't new, but it's #1 on GitHub's Shell trending today, and it opens up agent-built previs and blockouts.
- ComfyArt: a local Lovart-style infinite canvas on your ComfyUI
tldraw + React infinite canvas as a ComfyUI custom node. No account, no credits, no cloud. Optional design agent over MCP. Image plus short video via bundled MiniMax H3 workflows.
- Official ComfyUI: Blender blockout to Seedance 2.5
Same idea as LTX Layout-to-Render — 3D controls the camera, AI finishes the shot — but the Seedance 2.5 path, posted by the official ComfyUI account.
- #OpenSourceHiggsfield: Comfy's $25K challenge and the first open rebuilds
Comfy put up $25K for open-source rebuilds of Higgsfield-style apps. Within a day: a camera-move studio, open-weights VFX presets, a script-to-Resolve film studio, and an MCP-only video editor.
- MiniMax H3 gets fast locally: Sol-H3 Studio, HyperFlow 8-step, Ref2VA-VSA
NVlabs' Sol-H3 claims 35.5s for a 5-second H3 clip with audio on one RTX 5090 (vs 989s stock). Ref2VA-VSA claims ~72s for reference-to-video on a 4090 in ~13.5 GB.
- Fish Audio Drama 3 in ComfyUI: directed multi-speaker voice, then H3 talking video
35B open-weight TTS with inline cues ([sighs], [whispers]) and up to 16 speakers, running in ComfyUI from 8 GB VRAM by keeping MoE experts in system RAM. Then hand the take to MiniMax H3 for the picture.
- Block the shot in Blender, render it in LTX: the Layout-to-Render LoRA
Give it a grey clay playblast for the camera and a single styled frame for the look. It returns a finished shot that keeps your camera path and object placement.
- The top GitHub video repos right now, ranked by stars (Oct 2026)
Twenty repos for AI video, agent video skills and ComfyUI workflows, pulled from the GitHub search API on Oct 14 2026.
- Wan 3.0 in ComfyUI: reference tags for images, video and audio
Official templates for text-to-video, image-to-video and reference-to-video, with clips up to 30 seconds and synced audio.
- An open-source audio stack for video: music, foley, lip-sync, timing
Five repos that cover most of the sound side of an AI video pipeline without a subscription.
Skills, repos & workflows in the directory (63)
- FastFrames Repo
Apple ML Research open-weight autoregressive two-step video diffusion models, released Nov 2026: FastFrames-S (1.3B) and FastFrames-L (14B, trained from Wan2.1-T2V-14B) generate 832×480 video at interactive latency (per Apple, ~36 ms/frame on one B200 for S), plus a 16× faster LingBot 1.3B build for live camera control. Linux + NVIDIA/CUDA 12.8; research-only license (no commercial use).
- MCP for Blender MCP
Siddharth Ahuja's MIT community MCP server + Blender addon (formerly blender-mcp): agents read the scene, run Blender Python, pull Poly Haven/Sketchfab/Poly Pizza assets, generate meshes via Meshy, Tripo, Rodin or Hunyuan3D, and (2.2.8+) generate World Labs Marble worlds as Gaussian splats in Blender 5.3+. Claude Code and Codex plugins; optional paid Premium. Socket has no auth: keep it on localhost.
- Mago Studio Repo
AGPL-3.0 local-first desktop AI video editor (Windows/macOS/Linux): script-to-shots Director with continuity between shots, frame-accurate Transform (masked edits, relight, character replacement), versioning and review. Renders through a bundled portable ComfyUI with MiniMax H3 + LoRAs (Windows: NVIDIA 8 GB VRAM, ~63 GB disk) or optional paid Mago Cloud. MiniMax H3 licence applies; use swaps only with consent.
- AICG3D Short-Drama Workbench ComfyUI
GPL-3.0 ComfyUI pack (Chinese UI) for long, segmented MiniMax H3 videos and local short dramas: segment parsing, second-pass upscale on the joined cut to avoid seams, exposure lock, per-character voice locking across segments; workflows in the repo. MiniMax H3 licence applies.
- Paint a Plate ComfyUI
AInVFX's Apache-2.0 ComfyUI node pack for VFX plates: paint on the first frame to erase, add or relight and LTX 2.5 carries it through the shot; paint smoke/fire/fog on blank plates via the ainvfx-fluid IC-LoRA; EXR-in, EXR-out review and export. Open weights only (FLUX 3 Klein, LTX 2.5, SAM 3.1); 32 GB+ VRAM recommended. #OpenSourceHiggsfield Open Weights Only winner.
- Reshoot 4D ComfyUI
itsmetamike's AGPL-3.0 ComfyUI node pack: capture a video of a person in 4D, set a new virtual camera path in a node-panel editor, retake and finish with LTX 2.3, plus lens/DOF/motion-blur passes. Local, open weights only. #OpenSourceHiggsfield honorable mention.
- OneShot MV ComfyUI
VisualFrisson's MIT ComfyUI nodes + studio app: a song plus one line of setting becomes a lip-synced music video (beat/section analysis, invented cast via Krea 2, LTX-2.3 singing clips, SyncNet-checked retakes, auto-edit with drum-driven FX). Single Artist mode only with a consenting artist's photo. #OpenSourceHiggsfield honorable mention.
- MiniMax-H3-Image ComfyUI
MiniMax's open-weights 33B text-to-image and editing model from the H3 family: ~16 MP output, up to 16 reference images, 6 steps, native in ComfyUI v0.45.0 plus SGLang/vLLM-Omni/diffusers; also on the MiniMax API and Hailuo AI. Licence excludes US/EU/UK/KR without a MiniMax licence and requires a 'Powered by MiniMax-H3-Image' notice.
- HyperFlow H3 Flash Repo
Video Rebirth's 13.7B few-step (4-8 steps, no CFG) distill of MiniMax-H3 that generates 5-15 s video with synced stereo audio; text, first/last-frame and up-to-12-reference modes, ComfyUI templates, GGUF from 9.1 GB and official ComfyUI INT8 checkpoints. Code Apache-2.0; weights under the MiniMax H3 community licence (US/EU/UK/KR need MiniMax approval).
- Viggle Remake Repo
Open MiniMax-H3 LoRAs (swap + edit) from Viggle that re-shoot a clip while keeping motion, camera, timing and soundtrack; hosted app, ComfyUI nodes/int8/GGUF, and an MCP so agents can remake multi-shot films. MiniMax H3 community licence on weights; consent-only use.
- Freefield Repo
MIT open-weights Higgsfield-style studio by neuregex on ComfyUI: ~120 image/video/VFX/audio apps as app.json + API workflows (LTX-2.5, MiniMax H3, FLUX 3, etc.), lab runs with time/VRAM, Modal or local GPU. No partner/API nodes.
- H3 Character Sheet Designer ComfyUI
ComfyUI node by ukr8b3g-cmyk that uses MiniMax H3 image references to turn one character image into a full turnaround sheet: up to 13 views (head, full body, outfit front and back, hands, feet), up to three extra detail references, and a 7-view layout preset. Comes with a sample workflow; download the models separately.
- P-H3-Image ComfyUI
Pruna's 8-step text-to-image and multi-reference editing model distilled from MiniMax H3; native ComfyUI support, 4-bit set ~20 GB / ~21 GB VRAM at 8 MP per Pruna. MiniMax H3 Community License (excludes US/EU/UK/KR without a MiniMax licence).
- Comfy Previs Repo
MIT browser previs tool by sbobyn (#OpenSourceHiggsfield entry): three.js 3D blockouts, cast reference cards, gamepad/iPhone camera, then storyboard stills and AI video (LTX 2.5, Seedance 2.5, Kling O1, Wan VACE) via Comfy Cloud or local ComfyUI, with per-take cost estimates and a Claude MCP connector.
- MOTNGRFX ComfyUI
Apache-2.0 ComfyUI node pack + browser app by okkazoo: open-weights kinetic-type motion graphics where Gemma 4 plans, LTX 2.5 shoots with subject mattes so type passes in front of and behind the subject, and ACE-Step scores; 12/16/24 GB graphs.
- IV Studio Repo
MIT desktop film studio by Matthew Hollis for MiniMax H3 on your own ComfyUI or Runpod: idea to script, storyboard, shots and finished short (#OpenSourceHiggsfield entry).
- Seam ComfyUI
MIT transition studio by @jdnichollsc: two shots (images, clips or a .glb) in, an LTX-2.5 in-between out via ComfyUI; sequences, CLI and a Claude Code plugin (skill + MCP).
- Chimera Repo
MIT app by @_aCalle that turns a 3D model and an action prompt into editable skeletal animation: LTX-2.5 makes the motion video, Chimera fits the model's own rig and exports GLB/FBX/BVH.
- VFX 4 All ComfyUI
MIT app with 45 cinematic effect presets: Qwen Image Edit 2511 builds a scene around your photo, LTX-2.5 animates the effect with sound; open weights, local ComfyUI or Comfy Cloud.
- Earthfall ComfyUI
GPL-3.0 app by @jamie_jk: landmark plus portrait in, an 8-10 s dive from orbit to that person out, with FLUX 3 Klein, LTX 2.5 and a Qwen3.5 checker; 24 GB NVIDIA GPU.
- Massless Repo
MIT browser studio (live at massless.chucksi.com) that designs a shot with camera, lens and film choices, renders a still with open-weight FLUX.1 schnell or SDXL, then flies 28 camera moves (dolly zoom, crash zoom, arcs, bullet time) through it in WebGL with a storyboard, Kokoro voiceover and image-to-3D. No accounts; work stays in your browser.
- Pika H3 Flex Repo
Pika's open six-step adapters for MiniMax H3 (FL2VA: text/image/first-last frame to video with audio; Ref2VA: up to 12 image, video and audio references). Local 480p/768p, 5-15 s, 24 fps, 32 kHz stereo; tested on H100 80GB, ComfyUI templates. Adapters under the MiniMax H3 Community License, which excludes the US, EU, UK and South Korea; 1080p-4K and audio stems are hosted API only.
- Open-H3-Regenerate-2K Repo
NVIDIA's open refiner for MiniMax H3: draft at 768p, upsample the latent, and a LoRA regenerates it at 2560×1440 (up to 4K) in three Euler steps, with a sharpening LoRA and a fast H3 VAE. Authors claim a 5 s 2K clip in 22.45 s on one B200, about 5× the official 2K API. Re-draws detail like img2img, so faces and text can shift. Kijai's rank-16 to rank-384 LoRAs cut the 12 GB file to 0.15 to 6 GB. MiniMax H3 Community License plus mixed component licences.
- video-to-game Skill
lovis' agent skill (Claude Code, Codex, Cursor, Gemini) that turns a 10 to 15 s MiniMax H3 Max gameplay clip into a playable single-file Three.js browser game: it studies the clip frame by frame, generates every image, 3D model, sound and music track with fal, writes the game, then plays and films it as proof. The README lists 31 games for $202 in total (about $6 each), with a live gallery. MIT.
- OpenField Repo
wexloops' MIT local Higgsfield-style studio on ComfyUI and open weights: Director shot controls, Scene (local LLM plans and cuts 2-6 shots), Stage (greybox and .glb blocking rendered via the LTX-2.5 layout IC-LoRA), Mirror phone-motion transfer, Reshoot, seam-scored loops, HDR/EXR finishing, 31 MCP tools. NVIDIA 24 GB for video.
- Kitty Open Studio MCP
MIT self-hosted AI video studio on your own ComfyUI: 33 curated open-weights workflows as cards, a timeline editor, and MCP so Claude or ChatGPT can direct a music video. Runs on Runpod (pod or pay-per-second serverless) or your own GPU.
- Cheapsfield Repo
Apache-2.0 Windows studio for product and architecture visuals on ComfyUI + Comfy SDK, core nodes only, no paid APIs: sketch to packshot or render, relighting, staging, 1080p video with sound via LTX 2.5. NVIDIA 12 GB min, 16 GB recommended.
- Rapid-H3 ComfyUI
Apache-2.0 ComfyUI node pack: MiniMax H3 video+audio 3-5x faster on one NVIDIA DGX Spark (about 117 s vs 590-630 s stock for a 5.9 s clip, per the author) via INT8 sm_121 kernels and a 6-step draft-and-refine recipe. H3 weights keep their own licence.
- anymove Skill
Patch's Apache-2.0 creature mocap from one AI video: MiniMax H3 Max films a rigged creature from front and side in one frame (~$0.26 per 5 s clip on fal), a DINOv2 joint reader trained per creature finds joints in both views, and the rig is fitted and exported as a GLB. Wings, tails, eight legs; six Claude Code skills. Needs an NVIDIA GPU (24 GB to train a reader) and Blender 5.x. Big gestures fit best; eight-legged walks and lying down don't, per the README.
- Minimax-h3 Singularity V2 ComfyUI
WarmBloodAban's community fine-tune of MiniMax H3 for ComfyUI. Per the model card, V2 (INT8, ~34 GB) adds more action and varied motion, closer prompt following, an audio-reference fix and stronger photoreal/3D CG/cinematic CG styles; pruned builds by Abiray. Some users prefer V1.3 for calm shots. Card says Apache-2.0; H3's community licence also applies.
- MatteDesk ComfyUI
Lech Gruszewski's local Windows app on ComfyUI for VFX alpha mattes: SAM 3.1 core, CoTracker3, Sapiens2 edge band, edge engines (MatAnyone2, VideoMaMa, ViTMatte, BiRefNet_HR/Lucida), optional LTX-2.5 Alpha-Gen (24 GB VRAM), Nuke/After Effects handoff and an agent API + CLI. A Doctor installs missing nodes and models. NVIDIA GPU. GPL-3.0; some weights (e.g. ViTMatte) non-commercial.
- Henge ComfyUI
Alisson Anjos's MIT studio on ComfyUI + MiniMax H3 with trained per-mode LoRAs: character swap, motion transfer (guide latents or TARA), head swap and Sharpness (to 2K/4K). SAM 3.1 multi-character tracking, shot-by-shot long clips, a built-in assistant and an MCP server. Local needs 32 GB VRAM / 64 GB RAM and ~125 GB of models; or RunPod pod/Serverless. Weight and LoRA licenses apply; consent required for real people.
- OPENjutsu ComfyUI
Comfy Org's MIT character-swap app: replace a character in an existing clip while keeping motion, camera and timing, built on the akatz-labs H3 workflow with MiniMax H3 + SAM 3.1. Per the README it deploys to the Comfy Developer Platform (96 GB RTX PRO 6000, ~2 min per warm 5 s 480p part, deployment credits), 5-15 s parts at 480/768, optional GPT Image 2.5 first frame; also usable as a Comfy Hub app. Use only footage and likenesses you have rights and consent for.
- openjutsu (local fork) ComfyUI
Devozikjr's MIT fully local take on OPENjutsu: MiniMax H3 + SAM 3.1 + Qwen3-VL 8B. Simple, advanced (1-4 characters), object swap, swap-from-picture via SolarRenderer, face and colour fix passes; author lists RTX 4060 Ti 16 GB + 64 GB RAM as tested. H3 Community License applies; consent required for real people.
- FreeVideo ComfyUI
FlashML's Apache-2.0 launcher + ComfyUI workspace for local MiniMax H3 (OpenVDN 8-step VDN-H3). README claims 8 GB VRAM / 16 GB RAM via weight streaming; int8 on GeForce/RTX 30, FP8 on workstation cards; Windows/Linux, Mac preview; text, first/last frame, image/video/audio refs, community LoRAs, agent control and timeline (v0.4.0). H3 Community License terms apply to weights.
- SolarRenderer Repo
Structure video (greybox, segmentation, skeletons, depth, SMPL mesh, even a pencil storyboard) → finished clip on MiniMax H3 with two LoRAs in 4 steps; optional first frame + prompt set the look; streaming block-causal mode; open benchmark and free Hugging Face demo.
- DMAD (4-step H3) Repo
Four-step MiniMax H3 text-to-audio-video distillation LoRAs with Diffusers and native ComfyUI weights, example workflows and training code. T2VA only (not Omni references); don't stack with other speed adapters.
- HyperFlow Infinity Repo
Video Rebirth's agent harness for open-ended long video on MiniMax H3 + the HyperFlow 8-step LoRA: H3 renders a 5 s segment, a Qwen VL watches it and updates a scene-graph memory, a director step writes the next beat, and the loop continues from the last latents (60-minute single-run demo at 1344×768). Pause, branch from any segment, resume. Needs ≥4 NVIDIA GPUs (tested on H200 only). Code Apache-2.0; LoRA weights under the MiniMax H3 Community License (excludes US/EU/UK/South Korea without MiniMax authorization).
- PopTok Repo
Local prompt-or-song → music video pipeline: LLM songwriter and director 'workshops', ACE-Step 1.5 song takes ranked by Whisper, MOSS/Qwen-Image-Edit stills, LTX-2.5 + LTX-2.3 Transition LoRA as one continuous take, cuts snapped to downbeats, InfiniteTalk lip-sync on sung lines. Spreads jobs over DGX Spark, RTX 4090, Strix Halo or M5 Max boxes. Apache-2.0.
- Hy-Nomas Repo
Tencent Hunyuan's open interactive video world model: generates segment by segment and takes new text prompts or camera trajectories between turns, with a lifelong KV cache for consistency. 8-step distilled inference, 720×1280 at 24 fps, Cinematic and Robotics LoRAs, Nomas-PE prompt enhancer, single-RTX 4090 low-memory config (INT8/FP8, CPU offload), LoRA fine-tune recipe. CUDA 13. Weights on HF. Apache-2.0.
- MDMZ MiniMax H3 Character Swap workflow ComfyUI
MiniMax H3 + SAM 3 motion guide: swap a character into your own clip keeping lighting, motion and lip sync, or motion-transfer into the character's own scene. Up to 15 s at 24 fps, Gemini-node or free-chatbot auto-prompt, Turbo LoRA mode. 40+ GB of models. Workflow via the creator's Patreon post. Use only footage and likenesses you have rights to.
- LTX Desktop Repo
Lightricks' Apache-2.0 local LTX-2.5 app. v1.3.1 adds AlphaGen cutouts (no green screen), Motion Track paths, Pose/Depth/Canny guide video, 1440p/4K and a job-planning assistant. Weights: LTX Model License.
- FastH3 Omni ComfyUI
Open 8-step MiniMax H3 (Hao AI Lab FastVideo + Nuva Lab): prompt + up to 12 image/video/audio references → 5/10/15s clips with synced stereo audio. Native ComfyUI node; local recipes for RTX 5090/4090, PRO 6000, DGX Spark. Code Apache-2.0; weights MiniMax-H3 Community License.
- seriesmith ComfyUI
Series config + screenplay-style episode scripts in, cut episodes out: every shot compiled to its own ComfyUI job (MiniMax H3, LTX-2, Wan locally or Seedance/Kling/Veo via Comfy Router) with per-shot re-renders. #OpenSourceHiggsfield entry. MIT.
- ReAngle ComfyUI
Re-shoot a clip from a new camera angle with open weights: depth-lift the first frame, draw a camera path in 3D, then depth reprojection guides LTX-2.5 via the Layout-to-Render IC-LoRA. Demo on a 16 GB card. #OpenSourceHiggsfield. Apache-2.0.
- Comfield Repo
#OpenSourceHiggsfield entry: Higgsfield-style image + video studio on the Comfy Developer Platform. Cloud models via Comfy Router or open weights (LTX-2, MiniMax H3, FastH3) on your GPU; 27 camera-motion presets, A–D takes, live pricing. MIT.
- FXjutsu ComfyUI
Open-weights VFX presets on your footage (Vanish, Day→Night, Flood, Alpha Matte, painted-keyframe fire/smoke, restyles, composites) with LTX-2.5 + IC-LoRAs; no partner APIs. #OpenSourceHiggsfield entry.
- Sol-H3 in ComfyUI ComfyUI
NVlabs' accelerated MiniMax H3 studio over ComfyUI. README: 5s 768p clip with audio in 67.2s one-stage / 35.5s two-stage on an RTX 5090 (vs 989s stock). Needs Linux, 32 GB card, 90 GB RAM. Apache-2.0 repo.
- ComfyUI-Ref2VA-VSA ComfyUI
Sparse-attention MiniMax H3 reference-to-video node: README reports ~72s warm for a 5s 1344×768 clip on an RTX 4090, 4 steps, ~13.5 GB peak VRAM, with reference tokens kept out of pruning. Apache-2.0.
- ComfyUI-Drama3 ComfyUI
Fish Audio Drama 3 (35B TTS) natively in ComfyUI from 8 GB VRAM via experts-in-RAM: voice cloning, up to 16 speakers, inline cues, Qwen3-ASR, saved voices. Code GPL-3.0; weights under Fish Audio's licence.
- drama3-h3-talking ComfyUI
One ComfyUI workflow: text → Drama 3 voice → MiniMax H3 Ref2VA lip-synced talking video (take anchored at frame 0).
- ComfyUI Blender→Seedance 2.5 ComfyUI
Official ComfyUI workflow: Blender blockout controls the camera, Seedance 2.5 finishes the shot. Same pattern as LTX Layout-to-Render.
- ComfyUI Wan 3.0 templates ComfyUI
Official t2v / i2v / r2v graphs; @Image1/@Video1/@Audio1 reference tags (Partner Nodes, credits).
- Comfy-Org workflow_templates ComfyUI
Official template repo (Wan 3.0, vCube enhance, Meshy 3D...).
- ComfyUI-WanVideoWrapper ComfyUI
Community nodes for running Wan models locally.
- LTX-2.5 Layout-to-Render IC-LoRA ComfyUI
Clay/blocking playblast from Blender or Unreal + one styled first frame → finished shot that keeps your camera path. ComfyUI workflow included.
- LTX-Video Repo
Open video model family for self-hosting.
- Wan 3.0 (official) Repo
30 s native clips, up to 20 reference assets, instruction video editing.
- ACE-Step 1.5 Repo
Local song generation, covers, stems, LoRA from a few songs, BPM/key extraction.
- DiffRhythm Repo
Lyrics → full-length song, fast.
- MMAudio Repo
Video-to-audio synced foley (~6 GB VRAM). Check licence for commercial use.
- HunyuanVideo-Foley Repo
Foley generation from video + text.
- LatentSync Repo
Open audio-conditioned lip-sync (v1.6, 512px). Consenting subjects only.