Skip to content
BranchesTags

Latest commit

 

History

10 Commits

Folders and files

NameName
Last commit message
Last commit date
.github
.github
 
 
data
data
 
 
i18n
i18n
 
 
scripts
scripts
 
 
.gitignore
.gitignore
 
 
CONTRIBUTING.md
CONTRIBUTING.md
 
 
LICENSE
LICENSE
 
 
README.md
README.md
 
 
package-lock.json
package-lock.json
 
 
package.json
package.json
 
 

Repository files navigation

Awesome MiniMax H3

Awesome License: CC BY 4.0 Entries Verified

Everything built on top of MiniMax H3 — the open-weight omni-modal video model that generates picture and native audio together. Weights, quantizations, LoRAs, ComfyUI nodes, workflows and the tricks people found. Every entry checked against a live API and dated.

简体中文

Scope: H3 only. MiniMax ships plenty besides H3 — the M-series LLMs, Music, an agent stack with more stars than H3 itself. None of it is here. If an entry cannot be pointed at H3, it does not belong on this list.

Contents

How to read an entry. Every entry carries its popularity and the date it was last touched, both read from the live API on 2026-09-07. ⭐ marks a community project that MiniMax named itself, in its release roundup or on its community picks page — so it is not applied inside the Official section, where everything is official by definition.

Official

Start here. Weights, docs and the endpoints, straight from MiniMax.

Model and weights

Documentation, API and community

Run it

Getting H3 to produce a first clip, on whatever hardware you actually have.

In ComfyUI

Serving frameworks

Local and cross-platform

  • ⭐ antirez/h3.c 2.6k★ · 2026-08-11 — Redis author antirez rewrote H3 inference from scratch in C + Metal: safetensors read directly, Qwen text/vision encoders, DiT and both VAEs in one native Mac binary — no Python, no PyTorch. Covers t2av, first/last-frame and Ref2VA.
  • deepbeepmeep/Wan2GP 9.2k★ · 2026-09-07 — Runs H3 on low-VRAM consumer cards; pairs with the DeepBeepMeep packaged weights.
  • OpenCharAI/OpenChar 289★ · 2026-09-07 — Node-based AI filmmaking on your own GPU.
  • PipeNetwork/minimax-h3-mlx 73★ · 2026-08-10 — MLX port of the full 33B joint video+audio pipeline, validated against the reference implementation.
  • mrbizarro/Phosphene 211★ · 2026-09-07 — Video, image and character-LoRA training on a Mac, all in MLX: no PyTorch, no CUDA, no cloud, no API key. One-click Pinokio install, and it runs LTX-2.5 alongside H3.

Hardware-specific

Try it online

Fit it on your GPU

The full model is 123.6GB. These are the projects that got it down to something a desktop can hold.

Packaged builds

  • ⭐ Comfy-Org/MiniMax-H3 1.7k♥ · 19.6M↓ · 2026-09-06 — The ComfyUI packaged build, and by download count the way most people actually run H3: 40% of the modulation weights pruned, INT8 ConvRot and custom kernels take 123.6GB down to 42.5GB.
  • DeepBeepMeep/MiniMax-H3 56♥ · 464.2k↓ · 2026-09-06 — Low-VRAM packaging from the Wan2GP author.

Quantizations

Split-out components

Make it fast

Two separate levers: fewer sampling steps (Turbo LoRAs) and cheaper steps (caching, attention kernels). They compose.

Turbo LoRAs — 20 steps down to 4–8

Caching and attention

LoRAs and training

Teaching H3 a look it does not have — and the paths to training your own.

Style and capability LoRAs

Train your own

  • ⭐ fal H3 LoRA Trainer — Hosted trainer billed per training step, with t2v, flf2v, i2v and ref2va entry points. Reference material feeds both H3's visual conditioning sequence and the Qwen3-VL prompt side, so picture and sound can be trained together; frames, resolution, rank, steps and reference-conditioning ratio are all adjustable.
  • fal — how to train an H3 LoRA — Their working defaults: 10–200 clips at 24fps, rank 16 and up, lr 2e-4, A/B evaluation on a fixed seed.
  • ⭐ IAmIronMan42/MiniMax-H3-FineTuning 80★ · 2026-08-10 — Full fine-tuning on top of the official diffusers integration: separate noise schedules for picture and sound, real stereo audio in training. Run on 8×A800 over 2000+ clips at 30s / 448×768.
  • modelscope/DiffSynth-Studio 13.1k★ · 2026-09-07 — Complete H3 training path — LoRA scripts plus model docs.
  • DiffSynth — H3 LoRA training script — The FL2VA LoRA training entry point.
  • ostris/ai-toolkit 11.9k★ · 2026-09-06 — Added H3 t2v and i2v training.
  • unslothai/unsloth 75.7k★ · 2026-09-07 — Lists MiniMax-H3 among its supported training targets.

Unusual tricks

The most interesting corner of this ecosystem: things H3 turned out to do that it was never trained to do. MiniMax's own release roundup led with these rather than with benchmarks.

Continue a shot indefinitely

  • ⭐ NikoDemon80/ComfyUI-H3-Motion-Context 914★ · 2026-09-06 — Recursive generation: reads the previous segment's picture and audio latents directly, takes its last 22 frames as the next segment's opening condition, and re-aligns audio onto one timeline — avoiding the information loss of decoding and re-encoding between segments.
  • ethanfel/ComfyUI-MiniMaxH3-Context-Loop 383★ · 2026-09-07 — Reviewable long-form workflow: shot planning, per-segment preview, regeneration, checkpoints and automatic stitching.
  • tritant/ComfyUI_MiniMax_H3_Extender 192★ · 2026-09-06 — Folds Ref2VA conditioning, Motion Context, disk caching, multi-clip generation and image/audio references into one node, so long continuous sequences stop requiring a hand-built graph.

Use a video model to edit one image

  • ⭐ tori29umai0123/ComfyUI-MiniMaxH3-SingleFrame 83★ · 2026-08-19 — Turns H3 into an image editor with no extra training at all: pin the input image at frame 0, generate a single AV latent, decode it as an image. A second mode pins first and last frames and takes the middle. Ships Temporal RoPE Patch and Empty Single Frame Latent nodes.
  • Comfy-Org/ComfyUI#15416 — Tracking issue for single-frame VAE decode artifacts — read before you file your own.
  • thaakeno/ComfyUI-MiniMax-H3-Studio 95★ · 2026-08-20 — Turns H3 into an actual image workflow: text-to-image, editing, multi-reference generation, LightX acceleration, face refine, previews and benchmarking — without building the H3 graph yourself.

Other

  • matlowai/ComfyUI-MAINodes 140★ · 2026-09-05 — Contact-sheet diffusion (several views of one subject in a single generation) plus a motion lab for test-time de-roping.

Audio-only generation, no repository required

One of the three tricks MiniMax highlighted has no repo at all. Take the default image-to-video workflow, disconnect the input image, and drop the resolution to 32×32. Pull dialogue, ambience and effects out separately with Get Video Components. It runs close to real time, audio quality does not noticeably suffer, and the result can be fed back in as an audio reference. Credited to developer comfiestncoziest in the official one-week roundup; the original write-up appears to be a Reddit post we have not been able to locate — a pointer would be welcome.

Direct a film

H3 makes shots, not films. These projects add what sits above a shot: planning, continuity, batching, edit and export.

  • ⭐ chiphoton/MiniMax-H3-Codex-Drama 29★ · 2026-09-01 — An installable Codex plugin with nine built-in skills that drives local H3 through the ComfyUI MCP: character and scene design, storyboard keyframes, per-shot workflow selection, then FFmpeg edit, mix, subtitle, export and QC. Prompts, assets, candidate takes and selection results are all kept, so a run can be resumed and each new attempt keeps its own version.
  • ⭐ huangserva/ComfyUI_MiniMaxH3_Director 897★ · 2026-08-04 — Segment planning, conditioning, sampling, decoding and export folded into one node; PySceneDetect splits scenes automatically and the previous segment's closing motion and audio become the next one's context.
  • AIMixer/ComfyUI_MiniMaxH3_Director 1.4k★ · 2026-09-07 — Multi-segment directing, adapted to the official nodes.
  • seesee75-commits/ComfyUI-MiniMaxH3-Director 283★ · 2026-08-16 — Timeline storyboard.
  • open-video-ai/open-video 116★ · 2026-08-18 — "Ollama for MiniMax H3" — a local director layer over ComfyUI.
  • yg496/CS-H3-Multimodal-Director 57★ · 2026-08-05 — Multimodal directing setup.

Agent skills

Installable skills that let a coding agent — Claude Code, Codex, OpenCode and friends — do the prompting and drive the pipeline for you.

  • Rimagination/h3lite 407★ · 2026-08-23 — Hardware-aware Codex skill for local generation through ComfyUI: it profiles what you have and picks the route — low-VRAM fast, or quality — instead of making you choose. Bilingual docs.
  • benjiyaya/Minimax-H3-Prompt-AgentSkill 129★ · 2026-08-06 — Agent Skill that turns media plus a rough idea into a formatted H3 prompt.
  • SlavaSexton/ComfyUI-Agent-Kit 94★ · 2026-09-03 — One ComfyUI skill that every coding agent can drive — Claude Code, Codex, Gemini CLI, Qwen Code.
  • T8mars/minimax-h3-prompt-skill-T8 203★ · 2026-09-06 — Creative-DNA prompt cases packaged as installable skills.
  • unknowlei/minimax-h3-opencode-skills 108★ · 2026-08-09 — OpenCode skill suite: directing, routing, multishot planning and prompt generation.

Other ComfyUI nodes

Useful nodes that do not fit the categories above.

Prompting

H3 responds to a fairly specific prompt shape. These help you hit it.

Ready-made workflows

Drop-in graphs from Civitai and GitHub — including the low-VRAM configurations that actually run. Popular, and largely absent from other H3 lists.

On GitHub

On Civitai

Learn

Guides and news sources that keep up.

English

Chinese

Related lists

Other people covering the same ground, usefully.

  • wildminder/awesome-minimax-H3 471★ · 2026-09-06 — Deeply detailed on weights, quantizations and node inventory — the reference to reach for when you want file-level specifics.
  • AtlasCloudAI/awesome-minimax-h3-prompts 33★ · 2026-09-06 — Our companion list: H3 prompts from the official showcase, each with its generated preview, in 20 languages.

What is MiniMax H3?

H3 is MiniMax's open-weight omni-modal video model. Instead of treating text-to-video, image-to-video, reference conditioning and audio as separate models, it reads one multimodal context and generates picture and stereo audio jointly. That single design decision is why this ecosystem looks the way it does: the interesting projects below are mostly people discovering that one model already does things nobody trained it to do — editing a single image, generating audio on its own, continuing a shot indefinitely.

What gets in

  • SFW only. A meaningful slice of the H3 LoRA scene is adult-oriented. None of it is here, and neither are uncensored components such as stripped text encoders — including otherwise-good speed work that depends on them.
  • Popularity floor, with one exemption. 50+ stars on GitHub, 10+ likes on Hugging Face, 500+ downloads on Civitai. The floor keeps out the SEO shells that crowd this keyword. The single exemption is official recognition: anything MiniMax named itself goes in regardless — two of the best entries here have single-digit stars.
  • Every link resolves. The whole list was re-checked against the GitHub, Hugging Face and Civitai APIs on 2026-09-07. Entries that had gone private, 404'd, or that could not be verified programmatically were dropped rather than carried over on trust.
  • H3, not MiniMax. See the scope note above.

How this list stays current

Twice a week a GitHub Action re-checks every entry against the GitHub, Hugging Face and Civitai APIs: figures and dates are refreshed, renamed repositories are followed, and anything that stops resolving is withheld from the list — then deleted if it is still gone three weeks later. If more than a tenth of the run fails, it publishes nothing rather than stamp a verification date the entries did not earn.

A weekly sweep looks for H3 projects that are missing here and clear the bar, and files them as a review issue. Nothing is ever added automatically — whether a project does something no listed project already does is not a question a query can answer.

Entries live in data/ecosystem.json and both READMEs are generated from it by npm run generate; edit the JSON, not the Markdown. MiniMax keeps its own community picks page updated, and this list is reconciled against it.

Contributing

Found something missing? Open an issue or send a PR. Include the link, what it does that nothing else here does, and its current star or like count. Read CONTRIBUTING.md first — the bar above is applied literally.

License

CC BY 4.0. Linked projects keep their own licenses.

About

Everything built on top of MiniMax H3 — weights, quantizations, LoRAs, ComfyUI nodes, workflows and the tricks people found. H3 only, SFW only, every link verified.

Topics

ai-videoawesomeawesome-listcomfyuiimage-to-videoloraminimaxminimax-h3text-to-videovideo-generation

Resources

Readme
License

Contributing

Contributing
Activity
Custom properties

Stars

9 stars

Watchers

0 watching

Forks

1 fork
Report repository

Releases

Packages

Contributors

Languages