MiniMax H3 Integrations
A community-maintained index of models, tools and projects for MiniMax H3 and MiniMax-H3-Image.
English | 简体中文
Contents
MiniMax H3
MiniMax H3 is MiniMax's open-weight model for video with native stereo audio: it reads text, images, video and audio as context and generates clips of 4 to 15 seconds, up to 2K. It comes in two variants: FL2VA works from text and optional first and last frames, and Ref2VA takes up to 9 images, 3 videos and 3 audio clips as references.
Official
| MiniMax-H3 | Official repository: reference implementation and prompt guides. | |
| MiniMaxAI/MiniMax-H3 | Original weights for FL2VA (first and last frame) and Ref2VA (reference), with text encoder, video VAE and audio VAE. | |
| Comfy-Org/MiniMax-H3 | ComfyUI single-file weights, including AdaLN-pruned BF16, FP8 and INT8 builds. | |
| ComfyUI tutorial | Official ComfyUI guide for H3, from the day-0 announcement. | |
| ComfyUI workflow templates | Official T2V, I2V and R2V templates that ship with ComfyUI. | |
| Video prompt writing guides | Official guides for base (FL2VA) and reference (Ref2VA) prompts. | |
| MiniMax API | Official video generation API; for mainland China, use platform.minimaxi.com. | |
| Hailuo AI | Use H3 in the browser or the desktop app; for mainland China, use hailuoai.com and hub.minimaxi.com. | |
| ComfyUI API nodes | Built-in nodes for text, first and last frame and reference to video that run H3 and H3 Max through your Comfy account. | |
| MiniMax CLI | Official command-line tool for the MiniMax API, including H3 video generation. |
Quantizations
| unsloth/MiniMax-H3-GGUF | Pruned FL2VA and Ref2VA GGUF from Q2_K (6.26 GiB) to Q8_0, plus Qwen3-VL text-encoder GGUF. | |
| DmitryDB/MiniMax-H3-ComfyUI-Quants | INT8 ConvRot and NVFP4 builds of FL2VA and Ref2VA from 10.86 to 21.91 GiB, plus VAE files. | |
| DmitryDB/MiniMax-H3-DynTime-sQKV | INT8 DT-sQKV builds that need the repository's ComfyUI runtime patch. | |
| rockerBOO/minimax-h3-nvfp4-convrot | NVFP4 FL2VA and Ref2VA for Blackwell GPUs, recommended with INT8 ConvRot; experimental INT4 ConvRot FL2VA for other GPUs. | |
| Abiray/Minimax-H3-nvfp4-INT4-INT8-Convrot | Pruned NVFP4 (11.67 GiB), INT8 ConvRot and mixed INT4/INT8 ConvRot builds. | |
| tsolful/Minimax_H3_INT4MixedConvRot | Pruned mixed INT4/INT8 ConvRot builds of FL2VA and Ref2VA. | |
| Merserk/MiniMax-H3-INT4-ConvRot | INT4 ConvRot, 16.67 GiB full and 10.56 GiB pruned. | |
| AX1Y2JP/MiniMax-H3-W4A8-ConvRot | Pruned W4A8 ConvRot DiT (11.68 GiB) and a W4A8 text encoder. | |
| Kijai/MiniMax-H3-experimental | Experimental W4A8 mixed DiT builds and an INT8 ConvRot video VAE. | |
| Winnougan/MiniMax-H3-INT4_Convrot_ComfyUI | Pruned W4A8 ConvRot DiT with a matching quantized text encoder. | |
| rzgar/minimax_h3_fl2va_fp8_e4m3fn | FP8 E4M3FN, MXFP8, and FP8 with FP16 attention builds of FL2VA. | |
| abakanai/Minimax_h3_hybrid | Hybrid NVFP4 Ref2VA builds for Blackwell GPUs, 14.03 and 16.38 GiB. | |
| DiffSynth-Studio/MiniMax-H3-NF4 | NF4 DiT, text encoder and VAEs for DiffSynth-Studio, from 8 GB of VRAM. | |
| WaveCut/MiniMax-H3-OrbitQuant-W4A4 | W4A4 build with a quantized text encoder; needs the ComfyUI-OrbitQuant node. | |
| Abiray/MiniMax-H3-GGUF | Full FL2VA and Ref2VA GGUF from Q3_K_S to Q8_0, with text-encoder quants and a workflow. | |
| Abiray/MiniMax-H3-Pruned-GGUF | Pruned FL2VA GGUF, Q3_K_M and Q8_0. | |
| realrebelai/MiniMax-H3_GGUFs | Q3_K_M and Q4_K_M GGUF for FL2VA and Ref2VA, plus Q2_K and Q4_K_M text encoders; the Q2_K one is 7.91 GiB. | |
| vantagewithai/MiniMax-H3-comfyUI-GGUF | Full FL2VA and Ref2VA GGUF from Q3_K_M to Q8_0 for ComfyUI. | |
| MarxistLeninist/MiniMax-H3-FL2VA-Pruned-IQ1-GGUF | IQ1_S (3.78 GiB) and IQ1_M builds, the smallest published DiT. | |
| molbal/MiniMax-H3-GGUF | Pruned Ref2VA GGUF, including hand-tuned mixed layouts. | |
| leejet/MiniMax-H3-GGUF | Pruned FL2VA GGUF, Q2_K and Q4_K_M. | |
| joeygambino/MiniMax-H3-curve-GGUF | Pruned FL2VA and Ref2VA GGUF in Q4_0, Q5_1 and Q8_0; needs the H3-Multishot patch for ComfyUI-GGUF. | |
| jfar-z/MiniMax-H3-FL2VA-Ref2VA-Hybrid-NVFP4 | NVFP4 builds of the smhfacct FL2VA and Ref2VA hybrid in four block ranges, 11.67 GiB each, for Blackwell GPUs. | |
| lilcheaty/MiniMax-H3-NVFP4 | Pruned NVFP4 FL2VA and Ref2VA at 11.67 GiB each, plus full and mixed Ref2VA builds, for Blackwell GPUs. | |
| DeepBeepMeep/MiniMax-H3 | Weights WanGP downloads for the H3 family: FL2VA, Ref2VA, FastH3 Omni, HyperFlow Flash, Fun ControlNet and H3-Image, in BF16 and INT8. |
Components
| alibaba-pai/MiniMax-H3-Fun-Controlnet-Union | One control branch for Canny, depth, HED, MLSD and pose, plus video inpainting. | |
| lihaoyun6/MiniMax-H3-Ref-Patch | 148 MiB patch that lets FL2VA behave more like Ref2VA; needs its ComfyUI node. | |
| Kijai/MiniMax-H3-TAE | 9 MiB tiny autoencoder for previews in ComfyUI-KJNodes. | |
| Mamad8/MiniMax-H3-Image-VAE | Experimental single-frame VAE for still images from the video model; keep the original VAE for video. | |
| LBH-123-AI/Minimax_h3_latent_Upscaler | Learned latent upscaler weights, 1.0 to 4.0x, for its ComfyUI node. | |
| Open-H3-Regenerate-2K | NVIDIA's three-step refiner that takes a low-resolution H3 draft to 2560x1440, with code. | |
| ZawShiShawn/Gestura-MiniMax-H3-VAE-LiteRT | H3 VAE in LiteRT format for Gestura, an Android app for on-device video. | |
| speach1sdef178/MiniMax-H3-X2-Detail-VAE | Video VAE that decodes at twice the resolution; needs the fast VAE node from ComfyUI-MiniMaxH3_LatentUpscaler. | |
| Efficient-Large-Model/Open-H3-Regenerate-2K-IC | In-context LoRA version of NVIDIA's refiner that redraws a 768p draft at 1080p, 2K or 4K through Ref2VA. | |
| joeygambino/MiniMax-H3-encoder-GGUF | GGUF builds of the Qwen3-VL-32B text encoder for ComfyUI. |
Acceleration and distillation
| Minimax-H3-Turbo | LightX2V's 4- and 8-step distillation LoRAs, with weights on Hugging Face. | |
| larryvrh/MiniMax-H3-Turbo-Lora | Community 4-step Turbo re-bakes; v4_step600_ema is the usual default. | |
| drbaph/MiniMax-H3-Turbo-Lora-ComfyUI | Turbo LoRAs converted for the pruned ComfyUI checkpoints, including rank-21 resizes. | |
| Abiray/MiniMax-H3-Turbo-Lora-Pruned-ComfyUI | Turbo LoRAs for the pruned ComfyUI checkpoints. | |
| Kijai/MiniMax-H3_comfy | LightX2V Turbo LoRA in ComfyUI format, full and rank-21. | |
| InstantX/MiniMax-H3-Turbo-Lora-Diffusers | 4-step Turbo LoRA in diffusers format. | |
| joyfox/MiniMax-H3-Turbo | 4-step FL2VA LoRA. | |
| Tutu 20-to-8 NFE LoRA | Audio-video LoRA that cuts 20 evaluations to 8. | |
| t8star/minimax-h3-4step-turbo-loras-comfyui-exp | Experimental INT8 ConvRot Turbo LoRAs for ComfyUI. | |
| infosave/MiniMax-H3-Turbo-cmf | Turbo checkpoints in CMF format, Q4TP and Q2TP. | |
| alibaba-pai/MiniMax-H3-Acc-LoRAs | PDD 8-step acceleration LoRAs for FL2VA and Ref2VA, loaded natively by current ComfyUI. | |
| aptech0081/MiniMax-H3-PDD-ComfyUI | PDD LoRAs converted for ComfyUI, plus INT8 Ref2VA checkpoints with PDD baked in. | |
| t8star/Minimax-H3-Super-Acceleration-Comfy | PDD LoRAs repacked for the T8 audio nodes. | |
| 1ronman1993/MiniMax-H3-SVDQuant-fp4-pdd8 | SVDQuant FP4 build with the PDD 8-step LoRA folded in; an INT4 build is also available. | |
| FastH3 4-step Preview | FastVideo's data-free 4-step checkpoint with sparse attention; also as a LoRA. | |
| FastH3 8-Step V2 | 8-step text-to-audio-video checkpoint with 80% sparse attention. | |
| FastH3 Trim NVFP4 | Experimental 8-step model with 42 of 50 blocks, 11.1 GiB in NVFP4 for Blackwell and DGX Spark. | |
| FastH3-Omni 8-Step | 8-step reference-conditioned video and audio from image, video and audio references. | |
| FastVideo/FastVideo-FastH3-Comfy | FastH3 8-Step V2 repackaged for ComfyUI. | |
| realrebelai/FastH3_GGUFs | GGUF conversions of FastH3 4-step. | |
| Pika H3 Flex | Pika's six-step FL2VA and Ref2VA adapters for native 480p and 768p video, with CLI, diffusers and ComfyUI examples. | |
| PTD | EPFL's four- and six-step trajectory distillation, with weights and ComfyUI integration. | |
| DMAD | Four- and eight-step text-to-audio-video LoRAs with weights, workflows and training code. | |
| VDN | Video Delta Net: hybrid linear and softmax attention for H3, 2.6 to 2.9x faster per step, with weights and training code. | |
| drbaph/vdn-minimax-h3-int8-convrot-comfyui | INT8 ConvRot VDN in ComfyUI layout. | |
| t8star/Vdn-Minimax-H3-Comfy | Ready-to-use VDN model bundle for the T8 nodes. | |
| RAVEN Streaming LoRA | 4-step preview that turns H3 into a causal streaming generator, run through the RAVEN code. | |
| MATLOWAI/minimax-h3-fused-turbo-int8-convrot | One 21 GB INT8 ConvRot file with the Turbo and motion LoRAs folded in, for 4-step text-to-video, image-to-video and Ref2VA in ComfyUI. | |
| HyperFlow H3 Flash | Video Rebirth's 13.7B compressed H3 for video with stereo audio in 4 to 8 steps, INT8 and GGUF weights; also an 8-step LoRA for full H3. | |
| TaoMate-H3 | Three-step streaming generation for H3, with weights. | |
| TurboDiffusion | W8A8 kernels, sampler and multi-GPU runtime for H3 FL2VA, with TurboH3-FL2VA-I2VA weights. | |
| LongLive | NVIDIA's LongLive-Plug distillation, with few-step and CFG adapters for H3. | |
| FastGen | NVIDIA's distillation framework, with H3 among its video-with-audio models. | |
| cache-dit | Caching and parallel inference for diffusers, covering H3 T2VA, FL2VA and Ref2VA. | |
| Veda-Sparse/Minimax-H3-T2VA-Veda-8NFE-600Step-Preview | Learned sparse attention keeping about 10% of tiles in every mode, plus a tuned R2VA predictor; runs in ComfyUI or Miowtion. | |
| lightx2v/Minimax-h3-Turbo-SLA | 4-step FL2VA LoRA with 85% sparse-linear attention, about 2.5x faster on an RTX 5090 in LightX2V, with a ComfyUI build. | |
| taobao-content-ai/MiniMax-H3-PTD-LoRAs | Per-step PTD LoRAs that run Ref2VA in 4 or 8 steps, with diffusers code and ComfyUI nodes that need ComfyUI-PTD and the unpruned Ref2VA model. | |
| pdmd2026/pdmd_4NFE_lora | Rank-128 LoRA for 4-step generation, distilled with projected distribution matching, in diffusers key format. | |
| ChrisColeTech/minimax-h3-turbo-GGUF | Pruned FL2VA and Ref2VA GGUF with the LightX2V Turbo LoRA baked in, plus 4B and 8B text encoders for ComfyUI-ClipProj. |
LoRAs and fine-tunes
| minimax-h3-turnaround | Five rotated views of one subject from a single reference image. | |
| Realism - People | fal's LoRA for natural-looking people in everyday scenes. | |
| Looping Sketch Anime | Hand-drawn 2D anime look, made to loop. | |
| Krea2 Style | Experimental style LoRA toward Krea 2's look, for text-to-video only. | |
| smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models | FL2VA base with Ref2VA AdaLN blocks merged in, in four block ranges. | |
| t8star/minimax_h3_ref2va_patchin_hf102 | Experimental Ref2VA weight modification that boosts one high-frequency patch by 2%. | |
| Viggle Remake | Two Viggle LoRAs that recast people or change objects, backgrounds and weather; Viggle Animate recasts from one repainted frame. | |
| P-H3-Image | PrunaAI's fine-tune of the video model for text-to-image and image editing. | |
| Alissonerdx/Minimax-H3-ComfyUI | Experimental LoRAs for motion transfer, character and head swap, style transfer, VFX edits and live wallpapers. | |
| 360 Orbit | FL2VA LoRA that turns one photo into a frozen-time 360° orbit. | |
| Minimax-h3 Singularity | Fine-tuned merge of the FL2VA and Ref2VA checkpoints for text-, image-, reference- and video-to-video in ComfyUI; also a pruned build. | |
| prithivMLmods/MiniMax-H3-Facial-Realism-CloseUp | Experimental rank-16 LoRA for close-up faces with subtle expressions and micro-movements; prompts start with Facial Realism. | |
| LiconStudio/Minimax-H3-VBVR | Image-to-video LoRA trained on rebuilt VBVR data that reduces duplicated objects and characters passing through objects. | |
| alvdansen/h3-keyframe-animation | Three LoRAs that hold a hand-drawn animation cadence from one or two drawings; free for individuals and small businesses, gated. | |
| XGENlabs/XGEN-JING | Four-step first-person model on Ref2VA that makes video and audio from keyboard camera moves, reference images and text. | |
| SolarWM | World model whose H3 release streams camera-controlled video from a first frame and a prompt, with training code and the 4-step SolarRenderer. | |
| Wushu Action LoRA | Martial-arts action LoRA trained on 924 wushu clips; trigger word wushu_action. |
Runtimes
| ComfyUI | Native support since day 0, with INT8 in mainline. | |
| DiffSynth-Studio | MiniMaxH3Pipeline with NF4 inference from 8 GB of VRAM. | |
| LightX2V | Inference with parallelism, quantized DiT and feature caching; home of the Turbo LoRAs. | |
| Sana Sol-Engine | NVIDIA's per-GPU optimized runtimes, 3.55 to 4.52x faster than baseline, and the 4-step Sol-H3 runtime for up to 8 GPUs. | |
| FastVideo | Sparse-attention and step-distillation framework behind FastH3. | |
| VideoX-Fun | Training and inference code behind the Fun ControlNet-Union and the PDD LoRAs. | |
| h3.c | Native Metal engine for Apple silicon: text-to-video, first and last frame, and Ref2VA. | |
| Draw Things | Swift and Metal engine behind the Draw Things app: FL2VA, Ref2VA and FastH3 Omni on Apple silicon. | |
| mlx-serve | Apple silicon app and server with FastH3 Omni reference video in 8 steps. | |
| Raylight | Multi-GPU ComfyUI runtime that splits sequences across GPUs. | |
| Unsloth | Lists MiniMax-H3 among the models it can run and train. | |
| h3-flash | Inference scripts and a web demo for 8x B200 with sequence parallelism and fused kernels. | |
| TorchTPU + Diffusers | H3 on Cloud TPU v6e-4 with native PyTorch, torch.compile and tensor parallelism. | |
| stable-diffusion.cpp | C/C++ inference engine that runs H3 and its GGUF builds. | |
| LocalAI | Local OpenAI-compatible server; version 4.9.0 added H3 video with audio. | |
| fastllm | Native H3 text-to-audio-video with an HTTP demo. | |
| audio.cpp | C++ engine that runs H3 text-to-audio-video from Q4_K GGUF, optionally with an INT8 ConvRot DiT. | |
| 1Cat-vLLM | vLLM fork for V100 with native H3 audio-video generation. | |
| gufo | Inference engine for AMD Strix Halo that runs H3 video with sound. | |
| vllm.cpp | Python-free C++ engine that serves H3 from Q4_K_M GGUF through /v1/videos, measured at 34.6 s per step on one Jetson Thor. |
Apps and studios
| open-video | Runs H3 on your own GPU with install, pull and run, plus a skill that lets coding agents make video. | |
| FreeVideo | Local ComfyUI workspace built on VDN-H3, with an 8 GB VRAM path. | |
| Openjutsu | Character and object swaps in video with SAM 3.1 tracking. | |
| H3 Motion | Editor for keyframed subject motion and masked region regeneration. | |
| MiniMax-H3 Studio | Self-hosted web and API studio with a persistent queue, including FastH3 and Pika H3 Flex. | |
| TaleMondo Studio | Noncommercial Mac workspace for film production with local H3. | |
| Tiiny Studio | Desktop app for Tiiny Pocket hardware with H3-Turbo text- and image-to-video. | |
| LF Studio | Web studio for NVIDIA's Sol-H3 pipeline on one DGX Spark or one RTX 5090; formerly sol-h3-studio. | |
| Omnichar Studio | Character studio that keeps one face, body, outfit and voice across shots; generates with H3 and trains H3 LoRAs locally. Formerly Inline Studio. | |
| WanGP | Video generator for low-VRAM GPUs, with H3 among its models. | |
| SwarmUI | Web UI with H3 among its video models. | |
| Pallaidium | Generative movie studio inside Blender's video editor, with H3 and the Open-H3 Regenerate upscaler. | |
| Phosphene | Native Mac app that runs H3 and LTX-2.5 offline on Apple silicon; paid Pro version optional. | |
| freefield | Ready-made H3 apps on Modal: video from text or a photo, extending a clip, and new camera angles. | |
| nightmare-mobile | Android node-graph app for Snapdragon phones that makes H3 video with sound; non-commercial license. | |
| infinite-creation | Agent app that turns a novel or script into a video of any length with local ComfyUI H3; needs 12 GB or more of VRAM. | |
| Amuse | Desktop app built on .NET for local image, video and audio generation that lists H3 among its video pipelines, on CUDA or Vulkan. | |
| ltx-video-mac | Native macOS app that runs H3 on Apple silicon through h3.c, in BF16, INT8 or a 6-step Turbo build, with a local REST API. | |
| CodeVideoModel | Has a language model build a Three.js layout video, then turns it into realistic video with local H3 Ref2VA. | |
| GameFactory-3A | Game-asset framework whose CG-video stage renders cutscenes and trailers with local H3 at 720p. | |
| MiniMax H3 Video Studio | Self-hosted studio and CLI on ComfyUI for H3 text, image, reference and video-to-video, and segment-by-segment long videos. | |
| Shinsekai | AI RPG maker that can animate character sprites with H3 image-to-video on local ComfyUI. | |
| OpenShorts | Local-first short-video pipeline whose AI drama mode generates clips with H3 on your machine through stable-diffusion.cpp, or in the cloud. |
API integrations
| Hermes Agent | Self-improving agent whose fal video plugin runs H3, H3 Max and H3 Max Turbo, mapping usual resolutions to their tiers. | |
| n8n | Workflow automation tool whose MiniMax node can generate H3 video through the MiniMax API; fair-code license. | |
| MoneyPrinterTurbo | Turns a topic or keyword into a short video and can generate the footage with H3 through Metaso's API. | |
| OpenMontage | Agentic video production system with an H3 skill and tools that call the MiniMax API, fal or local ComfyUI. | |
| LiteLLM | Python SDK and AI gateway that can create H3 video tasks on the MiniMax API through its OpenAI-style video calls. | |
| New API | Self-hosted AI gateway whose Hailuo task plugin sends H3 multimodal jobs to MiniMax's v2 video API. | |
| Puter | Self-hostable web desktop whose MiniMax video provider uses H3 by default. | |
| OpenMAIC | Multi-agent classroom whose video adapter runs H3 and H3 Max on MiniMax's v2 API, directly or via TokenDance. | |
| Agno | Agent framework whose MiniMax tool generates H3 video from a text prompt through the MiniMax API. | |
| ArtCraft | Desktop app that stages scenes in 2D and 3D, then generates video with H3 through its own cloud or fal. | |
| QwenPaw | Personal assistant whose Creator plugin makes video with H3 or H3 Max on the MiniMax API, or with self-hosted H3 on SGLang. | |
| Magpie | Menu-bar app that sets up models for coding agents; with a MiniMax API key it also makes videos with H3. | |
| Sim | Agent workflow workspace whose video tool runs H3 and its Highspeed, Lite and Max versions on the MiniMax API. | |
| AI SDK | Vercel's TypeScript toolkit whose MiniMax provider makes H3 and H3 Max video and checks each mode's input limits. | |
| MaxKB | Enterprise agent platform whose MiniMax provider adds H3 and H3 Max text-to-video and image-to-video models. | |
| Toonflow | Short-drama studio from script and storyboard to video, calling H3 through the Metaso, CompShare or APIMart API. | |
| Huobao Drama | Generates a short drama from one sentence to finished video, with an H3 adapter for the MiniMax API; non-commercial license. | |
| Corsair | App integration platform whose MiniMax plugin lets agents generate H3 and H3 Max video through the MiniMax API. | |
| waoowaoo | Film and video production platform that calls H3 and H3 Max through OpenRouter; Elastic License 2.0. | |
| Bifrost | AI gateway that runs H3 and H3 Max as async tasks on MiniMax's v2 API, fitting size and seconds to each model's range. | |
| RikkaHub | Android chat app for many LLM providers that can also generate H3 video through the MiniMax API. | |
| xiaobei | Marketing agent for small businesses whose video skill makes clips with H3 through the MiniMax API; MIT with extra terms. | |
| JiuwenSwarm | AI agent whose video tools send H3 and H3 Max jobs to the MiniMax API. | |
| iPolloWork | Agent workbench with H3 video and H3 digital-human tools run as RunningHub workflows; personal non-commercial license. | |
| ArcReel | Agent workspace that turns novels and scripts into shots and video, with a built-in H3 endpoint for the MiniMax API. | |
| Monid | Marketplace of tool APIs for agents, with connectors for H3, H3 Max and H3 Max Turbo on MiniMax's v2 API. | |
| Raven | Multi-agent system whose MiniMax media tools create H3 and H3 Max videos through the v2 API. | |
| Pollinations | Open-source generative AI platform whose API serves H3, H3 Max and H3 Max Turbo video relayed from fal. | |
| MonkeyCode | Enterprise AI office workbench whose video service runs H3 text, frame and reference modes on the MiniMax API. | |
| YouTube Automation Agent | Runs a YouTube channel end to end and can render clips with H3 on the MiniMax API under a paid-seconds cap. | |
| RubyLLM | Ruby AI framework whose MiniMax video protocol lets its animate call create H3 videos. | |
| AgentField | Framework for running agents as services whose MiniMax media provider uses H3 as its default video model. | |
| Agency Orchestrator | Multi-agent workflow tool whose video steps can render H3 clips through the Metaso or APIMart API. | |
| LocalMiniDrama | Short-drama tool that keeps data on your machine and renders shots with H3 on the MiniMax API or through relays. | |
| BeefTV | Canvas video workspace with a live H3 Max Director view on fal and an H3 plugin for MiniMax's v2 API; source-available. | |
| OhMyGame | AI game studio app whose video tools offer H3 and H3 Max through the MiniMax API. | |
| open-ai-canvas | Workspace for AI film with H3 prompt cards and H3 video through MiniMax, Alibaba Bailian or Metaso. | |
| ai-fusion-video | Java agent platform for short dramas and comics, with H3 and H3 Max presets for the MiniMax API. | |
| Row-Bot | Local-first assistant whose video tool runs H3 and its Highspeed, Lite and Max versions on the MiniMax API. | |
| AICON | Canvas workflow agent for video creation whose default video model is H3, called through the project's own API gateway. | |
| infinite-canvas | AI creation canvas that makes H3 video over a MiniMax-protocol channel, with built-in H3 prompt-writing skills. | |
| ClipForge | Turns one product photo into a short sales video, with H3 and H3 Max available through Atlas Cloud. | |
| StoryGen-Atelier | Storyboard tool where Gemini writes shots and frames, with H3 on the MiniMax API as an alternative to Veo for clips. | |
| Reveric | Desktop studio for novels and short dramas that generates video with H3 through the MiniMax or Metaso API. | |
| TapCanvas | Agent-native board for AI film and comics whose bundled gateway speaks MiniMax's H3 v2 video protocol. | |
| wind-comic | Multi-agent pipeline that turns one line of text into a short drama, rendering shots with H3 on the MiniMax API. | |
| ReqLLM | Elixir LLM library whose generate_video calls H3 on the MiniMax API and can upload local reference files. | |
| Kivio | Screen-level AI assistant for macOS and Windows that can make H3 videos through MiniMax's v2 API. | |
| OpenTryOn | Virtual try-on and fashion toolkit with H3 and H3 Max through the MiniMax API, or H3 on a local GPU; non-commercial. | |
| Open Storyboard Canvas | AI shot-planning board with camera controls whose video nodes can use a preset MiniMax H3 provider. | |
| Loomic | Chat-driven design agent on an infinite canvas that makes H3 video from text or frames through Metaso's API. | |
| Kinema | Agent for AI comics, films and commentary videos, with a minimax-h3 provider on MiniMax's v2 API. | |
| BrewReel | Turns a product brief into a vertical promo video; its talking-head mode adds AI B-roll made with H3 on the MiniMax API. | |
| AI Video Production Editor | Electron app for AI filmmaking whose node pipelines can run H3 text- and image-to-video on fal. | |
| Mora | Local-first promo video workbench that can render clips with H3 through Atlas Cloud. | |
| Polox AI | Generation platform built on DeepSeek Harness, with H3 text, image and reference to video through WaveSpeed. | |
| Chibi | Self-hosted Telegram bot whose video generation can use H3 and H3 Max through the MiniMax API. | |
| WaveSpeed MCP Server | MCP server whose H3 family generates, edits and extends video on WaveSpeed's API. | |
| Open Media Gateway | Self-hosted service whose MiniMax provider covers H3 and its Highspeed, Lite and Max versions on the official API. | |
| KunFlix | Platform for film and TV ad creation whose video service calls H3 through MiniMax's v2 API; Apache-2.0 with extra terms. | |
| hf-studio | Self-hosted media studio that runs H3 video through Higgsfield and H3-Image through KIE. | |
| Magiviz | AI video creation engine whose video models include H3 through KIE's API. | |
| OpenFilm | Video agent and editor for coding agents that can replace the people in a clip with H3 Max Recast on fal. | |
| CozyClay | Browser previs tool: block a scene and camera, then render shots with fal's H3 Max Turbo or your own ComfyUI H3 workflow. | |
| OpenStory | Video sequence platform whose motion step offers H3 Max text, image and reference to video on fal. | |
| video-to-game | 31 games, each rebuilt from a 10 to 15 second H3 Max clip made on fal, with the asset pipeline. | |
| anymove | Motion capture for any creature from two-camera H3 Max clips on fal, with a skill that trains an H3 LoRA there. | |
| prompt2rig | Type a motion and a rigged character performs it, with H3 Max Turbo on fal as the fast video option. | |
| falcraft | Minecraft mod for fal-generated builds, including a wall TV whose live stream is rendered by H3 Max Director on fal. | |
| varg SDK | TypeScript video SDK with H3 Max text, image and reference to video models on fal; its JSX API is now deprecated. | |
| Banana Studio | Self-hosted media studio with H3 Max text, image and reference to video, extend, lip sync and recast on fal. | |
| Henji AI | Desktop app that brings many AI providers together, with H3 Max through fal or APIMart. | |
| pokemonlive | Local Pokémon battle prototype whose cinematics come from H3 Max Turbo and H3 Max reference-to-video on fal. | |
| live-classroom | Live TV channel that teaches any topic, with each lesson beat rendered just in time by H3 Max Turbo on fal. | |
| H3 Max Realtime Education | Science tutor you can interrupt by voice, with teaching video generated continuously by H3 Max on fal. | |
| SketchMaker | Self-hostable creative app whose animate, story and lip-sync features use H3 Max on fal by default. | |
| H3 Max Blender | Blender neural rendering demo: H3 Max on fal restyles a gray ship in four looks as its geometry changes. | |
| SPITE | Pre-production canvas for filmmakers whose video nodes can use H3 Max through fal with your own key. | |
| Interdimensional Game | Choose-your-own-adventure game with real-time video from H3 Max on fal. | |
| UNREEL | AI streaming app where a language model writes each episode and H3 Max Turbo on fal renders it shot by shot. | |
| StoryCast | Turns a topic into a story video, with shots and lip-synced lines from H3 Max on fal. | |
| OpenFlipbook | Clickable picture book where every page is AI-drawn, with optional page animation from H3 Max or H3 Max Turbo on fal. |
ComfyUI nodes
| comfyui-minimax-h3-audio-T8 | 62 nodes for audio, stills, conditioning, long video and speech. | |
| ComfyUI_MiniMaxH3_Director (huangserva) | Storyboard nodes with five importable templates: t2v, fl2v, r2v, v2v and rv2v. | |
| ComfyUI_MiniMaxH3_Director (AIMixer) | The original Director nodes. | |
| ComfyUI-MiniMaxH3-Director | Timeline editor: media on tracks, one prompt per shot, live preview and retakes. | |
| ComfyUI-MiniMaxDirector | Timeline of shots, camera moves and audio cues compiled into the structured prompt from the official guides. | |
| ComfyUI-MiniMaxH3-Easy | One compact workflow for T2V, I2V, first and last frame, and reference video. | |
| ComfyUI-FeiHou-Easy-H3 | Fork of MiniMaxH3-Easy with built-in slots for 9 images, 3 videos and 3 audio references. | |
| ComfyUI-H3-Motion-Context | Chains clips while keeping motion and sound continuous across them. | |
| ComfyUI-PainterNodes | MiniMaxRefToVideo2 node with the official reference and dialogue format. | |
| ComfyUI-H3-FunControl | Loads the Fun ControlNet-Union with its preprocessors and inpaint path. | |
| ComfyUI-MiniMaxH3_Ref-Patch | Applies the Ref Patch to FL2VA. | |
| VPack | Compresses Ref2VA reference-video conditioning to cut reference tokens. | |
| TARA-H3 | Experimental temporal alignment of Ref2VA video-reference attention, with a technical report. | |
| ComfyUI-Dials | Training-free steering vectors built from prompt pairs, for motion and style. | |
| LanPaint | Training-free video and audio inpainting. | |
| ComfyUI-MiniMaxH3_LatentUpscaler | Latent upscaler for H3's audio-video latents, for two-pass sampling. | |
| Comfyui_Minimax_h3_latent_Upscaler | Learned 1.0 to 4.0x latent upscaler that skips the VAE round-trip. | |
| ComfyUI-H3-ResCascade | Hires-fix nodes that refine about 1 MP renders to 1080p or 1440p, keeping the audio. | |
| ComfyUI Open-H3 Regenerate | Nodes and workflows for NVIDIA's Open-H3-Regenerate-2K refiner. | |
| Tessera | Tiled, chunked video refinement with streamed VAE encode and decode. | |
| ComfyUI-H3VAE_TRT | TensorRT video VAE, up to 1.7x faster decode. | |
| ComfyUI-Spectrum-MiniMax-H3 | Skips selected transformer steps by forecasting features; output is approximate. | |
| ComfyUI-SolAttn_triton | Training-free Sol-Attn sparse attention, tested with H3; deprecated now that ComfyUI has it built in. | |
| ComfyUI-VDN-H3 | Runs the VDN adapters in ComfyUI; a 24 GB fork offloads more. | |
| ComfyUI-Ref2VA-VSA | Video sparse attention for Ref2VA, about 72 s per 5 s clip on an RTX 4090. | |
| ComfyUI-MiniMax-H3-PDD-Acc | Loads the PDD LoRA with its 32 output heads on older ComfyUI. | |
| ComfyUI-Sol-H3 | ComfyUI front end for the Sol-H3 runtime (GPL-3.0). | |
| Spark-H3 | Adaptive block-sparse attention with diffusers and ComfyUI packages. | |
| ComfyUI-OrbitQuant | Loads the OrbitQuant W4A4 build. | |
| ComfyUI-FastH3-Lora-Converter | Converts the FastH3 LoRA to ComfyUI key order. | |
| ComfyUI-INT8-Fast | Largely superseded by native INT8; its convert_comfy_quant.py converts older INT8 files. | |
| ComfyUI-MiniMaxH3-TimelineDirector | Editable timeline of reference media for Ref2VA, with an all-in-one director workflow. | |
| ComfyUI-MiniMaxH3Mod | RefMod adapters that compress reference images and videos into small latents without training; still under construction. | |
| ComfyUI-H3-FaceRefine | Tracks small faces frame by frame, refines them with H3 and stitches them back. | |
| comfyui-AICG3D | Short-drama workbench that splits a script into segments for videos of any length. | |
| H3-Character-Sheet-Designer | Character reference sheets with up to 13 views from one character image. | |
| ComfyUI_MiniMax_H3_Extender | Chains clips into long video while keeping motion, picture and audio continuous. | |
| ComfyUI-Majoor-H3-GuideMaster | Places images and audio on the H3 timeline as guides. | |
| ComfyUI-Viggle-Animate-H3 | Nodes for Viggle Animate, which replaces a character in a driving video with a reference. | |
| ComfyUI-MiniMax-H3-Image-Studio | Still images and reference editing with the video model, plus Qwen detail refinement. | |
| ComfyUI-MiniMax-H3-Studio | Reference-aware still-image workflow built on the video model; alpha. | |
| TE-Speed-MiniMaxH3-OSS | Step cache that reuses late-block residuals when sigma barely changes, about 45% faster at 30 steps; needs a one-time ComfyUI patch. | |
| ComfyUI-MiniMax-H3-LongMedia | Long-form video and audio nodes with a Director timeline, single-clip rerenders and lip-sync redubbing. | |
| ComfyUI-H3-Continuum | Generates long video and audio in chunks with a prompt per chunk, lets you retry one chunk and resumes from saved takes. | |
| ComfyUI-H3-Motion-Context-MultiRef | Fork of H3-Motion-Context that adds music videos built around one song, V2V motion transfer, AV bridges between two clips and masked keyframes. | |
| comfyui-obvpm-timeline | Timeline node that generates, extends, prepends, bridges and loops clips like a small video editor; needs comfyui-obvpm. | |
| ComfyUI-YCNodes-MiniMax-H3 | Six nodes, including time-split prompts, dynamic CFG, low-noise refinement and tiled sampling. | |
| Goohai-MiniMax-H3_Integration | All-in-one conditioning node that picks T2VA, FL2VA, Ref2VA or hybrid mode from the uploaded media and writes the official tags. | |
| Comfyui-MMH3-UltimateUpscale | Upscales and re-samples a finished latent in time chunks and tiles, so long high-resolution clips fit limited VRAM. | |
| ComfyUI-MAINodes | Motion Lab renders a slowed copy and drops held frames to fix smeared fast motion; also turnaround and low-VRAM nodes. | |
| ComfyUI-MiniMaxH3-PrefixStream | Long-video continuation that writes the previous clip's clean latents into the start of the next, with a clip bin for picking which take to continue. | |
| ComfyUI-ClipProj | Swaps the 32B text encoder for Qwen3-VL 4B or 8B through a learned projection, down to 4.5 GB. | |
| ComfyUI_MinimaxH3HybridLoader | Loader that merges chosen Ref2VA tensor groups into FL2VA at load time by block range, without writing a new file. | |
| ComfyUI-Fizgig-H3-Still | Two nodes for true single-frame stills, with a one-frame latent and a clean decode at any size; no extra model. | |
| ComfyUI-H3-AudioRefine | Freezes the video of a sampled latent and runs extra steps on the audio only, for 4-step Turbo runs whose sound is too rough. | |
| ComfyUI-H3-Continuity-Studio | Multi-shot timeline that carries context between clips and across scene or camera changes, with an MCP adapter for coding agents. | |
| ComfyUI-sol-attn | Sol-Attn Triton kernel with H3 nodes for zero-copy attention, fused AdaLN modulation and chunked MLP, on RTX 30 series through Blackwell. | |
| MiniMaxH3-Studio | ComfyUI sidebar with 28 nodes for prompts, the official H3 workflows, long-video planning and a built-in editor. | |
| Veda-on-ComfyUI | One node that adds Veda sparse attention to native H3, 2.9x faster end to end on an RTX 5070, on NVIDIA, AMD ROCm and Apple silicon. | |
| WAS Node Suite | Node suite whose H3 nodes condition a multi-segment video on one node, extend it by window and steer segments with Fun ControlNet. | |
| ComfyUI-MiniMax-ContextIR | Nodes from prompt to sampling for local H3, with optional prompt rewriting through MiniMax's H3 Context-IR API. |
Workflows
| MiniMax-H3-Multishot-Workflow | Strings FL2VA and Ref2VA clips into one sequence with matched audio handoffs. | |
| comfy-MiniMax-H3-workflows | The low-VRAM acceleration stack in one importable workflow. | |
| MiniMax-H3-8GB-Workflows | Three-tier workflow pack for 8 GB laptops. | |
| FL2V GGUF workflow | Runs the GGUF-quantized FL2VA model in ComfyUI. | |
| OrbitQuant workflows | T2VA workflow for the W4A4 build, with T2VA and Ref2VA API forms. | |
| ComfyUI-MiniMax-H3-Workflow | First/last-frame, image-to-video and text-to-video in one workflow, with the Turbo LoRA and a control panel for resolution. |
Prompting and agent skills
| ComfyUI-MiniMax-H3-Promptor | Builds H3 prompts with <Picture X> tags placed inline in the action. | |
| ComfyUI-MiniMax-H3-Guide | Splits identity, keyframes, motion, voice and score into roles and compiles valid H3 prose. | |
| OpenH3-IR | Writes prompts in the Context-IR format MiniMax published, then checks and fixes them; CLI, HTTP or ComfyUI. | |
| comfyui-minimax-h3-prompt-enhancer-T8 | Server-side prompt enhancement in ComfyUI. | |
| MiniMax-H3-Prompt-Rewriter-LoRA | Qwen3.6-27B LoRA that rewrites short prompts into H3's three-part structure; an 8B vision version also reads keyframes. | |
| awesome-minimax-h3-prompts | Prompt collection with example videos, by category. | |
| MiniMax H3 1K prompts | 1,000 curated text-to-video prompts with attribution. | |
| minimax-h3-prompt-skill-T8 | Case library installable as an agent skill, with a desktop viewer. | |
| awesome-minimax-h3 (skills) | Agent skills that take a brief to a finished file through the hosted API. | |
| okaypic-video-skill | Agent skill that produces a captioned multi-shot episode through the okaypic API. | |
| ComfyUI-Agent-Kit | Drives a local ComfyUI from coding agents, with an H3 skill. | |
| minimax-h3-opencode-skills | OpenCode skills for directing, multi-shot planning, prompting and review. | |
| MiniMax-H3-Prompt-Rewriter-ComfyUI | ComfyUI nodes for LightX2V's Prompt Rewriter LoRA, plus reference captioning. | |
| ComfyUI-H3-VisionPromptor | Turns an idea and optional reference images into an H3 prompt with a local vision model. | |
| h3lite | Codex skill that picks a ComfyUI setup for your hardware and generates H3 video locally. | |
| Minimax-H3-Prompt-AgentSkill | Builds a formatted H3 prompt from your media and idea, then hands it back for you to run. | |
| ComfyUI-Fantastic-MiniMaxH3-PromptBuilder | Fillable prompt templates for every mode, checked against the official guide as you type. | |
| ComfyUI-MiniMaxH3-Prompt-Writer | Prompt-writing panel in ComfyUI that turns a short brief and reference media into the official H3 format, using Ollama, a local GGUF or an API model. | |
| minimax-h3-video-reverse-skill | Agent skill that reads a reference video and writes an H3 text-to-video prompt or per-shot image-to-video prompts. | |
| h3-storyboard-skill | Claude Code skill that breaks a script into H3 shots and directs facial performance, marking each rule as tested or inferred. | |
| oh-my-minimaxh3-director | Agent skill that splits a script into shots, batch-runs them through local ComfyUI workflows and assembles a JianYing draft. | |
| MiniMax-H3-Cloud | Codex plugin that rents a GPU on the CompShare cloud, sets up ComfyUI, runs text, keyframe or reference-to-video jobs and shuts the instance down. | |
| ComfyTV | Canvas app for ComfyUI whose built-in agents ship with an H3 cinematic-director skill. | |
| drama-skills | Short-drama skills from novel to storyboard to prompts, with an H3 mode that binds reference images and speakers. | |
| shuohao-skills | Short-drama skills for character, art and screenplay bibles and storyboards, with the storyboard step aligning cuts and keyframes to H3 prompts. | |
| h3-prompt-journal | Case studies of music-driven dance and multi-shot prompts, each with the full prompt and what made it work. | |
| awesome-MiniMax-H3-cases | Over 3,000 playable H3 videos with their sources, over 700 complete prompts and guides for running H3 locally. | |
| H3 Prompt Composer | Offline prompt builder in a single HTML file that runs in the browser. | |
| Scenario skills | Toolkit for agents on Scenario's API whose video skill covers H3 keyframe and reference modes and native audio prompts. | |
| ChatCut Agent Plugin | AI video editor for coding agents whose video skill submits H3 and H3 Max jobs to its hosted service. | |
| oh-story-dsh | DeepSeek Harness plugin for web novels and short dramas that renders video with H3 through the MiniMax API. | |
| God MiniMax H3 | Prompt library and agent skill for 30-second H3 films built from chained segments, with API parameters for each prompt. | |
| ancient-rap-battle | Claude Code skill for rap battles between historical rivals, with an optional H3 lip-sync route on the MiniMax API. | |
| short-drama-production | Codex skill that runs a short drama from outline to rough cut and submits H3 jobs to the MiniMax API or CompShare. | |
| h3-game-sprites | Agent skill that turns OpenRouter H3 clips into 2D sprite sheets, one clip per move, keyed and checked frame by frame. | |
| MiniMax H3 B-roll Generator | Codex skill that finds where a talk needs cutaways, writes H3 prompts and renders approved shots via MiniMax or Metaso. | |
| Goose Skills | Growth and marketing skills for coding agents, with a creator part that makes talking-creator clips with H3 Max on fal. | |
| Motion Graphics Music Video Skill | Claude plugin that turns a music track and a prompt into an animated video, with H3 Max image-to-video on fal. |
Training
| MiniMax-H3-FineTuning | Supervised LoRA training on the diffusers implementation, with audio in the loss. | |
| verl-omni | RL post-training of H3 LoRAs with FlowGRPO and DiffusionNFT. | |
| miles_diffusion | Flow-GRPO post-training with an H3 text-to-audio-video recipe. | |
| Fizgig | LoRA and LoKr training studio with an H3 preset. | |
| musubi-tuner | LoRA training scripts with experimental H3 support. | |
| diffusion-pipe | Pipeline-parallel trainer with H3 text-to-image and text-to-audio-video training. | |
| SimpleTuner | Fine-tuning kit for image, video and audio models, including H3. | |
| VeOmni | ByteDance's training framework with FL2VA and Ref2VA configs for H3. | |
| UniRL | Tencent Hunyuan's RL framework; H3 text-to-audio-video is supported on the training side only for now. | |
| Flow-Factory | Flow-matching fine-tuning framework with GRPO LoRA recipes for H3 text-, first/last-frame- and reference-to-audio-video. |
Serving and hosting
| SGLang | Native H3 serving with a cookbook of recipes verified from one RTX 4090 to B300 and AMD MI355X. | |
| vLLM-Omni | OpenAI-compatible /v1/videos serving with an official H3 recipe. | |
| UniServe | HTTP serving for FastH3 and FastH3-Omni on 1 to 8 GPUs. | |
| minimax-h3-gb300-ebpf | Reproducible GB300 benchmarking and tracing of SGLang H3 serving. | |
| H3 Max on fal | Hosted H3 Max, Director and Turbo models, post-trained by fal; closed weights, also reachable from ComfyUI Partner Nodes. | |
| Nuva Lab | Dedicated H3 and FastH3 hosts for one company, including customer fine-tuned weights. | |
| MiniMax H3 on fal | H3 itself on fal's API, with text, image and reference to video at up to 2K. | |
| MiniMax H3 on Replicate | H3 on Replicate's prediction API, with text, first and last frame and reference to video at 768P or 2K. | |
| MiniMax H3 on WaveSpeed | The open H3 weights served as an API, with text, image and reference to video. | |
| H3 on OpenRouter | H3 and H3 Max video through OpenRouter's API. | |
| MiniMax H3 on Cloudflare | H3 in Cloudflare's AI model catalog, with text, first and last frame and reference to video. |
Guides
| Run H3 locally | This repository's guide: VRAM recipes, every quantized file, components, compatibility and measured environments. | |
| GPU and VRAM requirements | Nuva Lab's notes on GPUs and VRAM for local H3, with a cost comparison against the API. | |
| minimax-h3-notes | Single-machine run notes. | |
| minimax-h3-turing | Field handbook for 22 GB Turing cards with W4A8 and Turbo. | |
| -h3-8gb-traps | Measured 8 GB laptop path, with a reproduction for each trap. |
MiniMax-H3-Image
MiniMax-H3-Image is the open-weight image model of the H3 family, with text-to-image and multi-image editing in one checkpoint. It generates images of about 16 megapixels directly in 6 steps, at aspect ratios from 1:3 to 3:1, and takes up to 16 reference images.
Official
| MiniMax-H3-Image | Official repository: usage guide, showcase, local deployment and a prompt writing skill. | |
| MiniMaxAI/MiniMax-H3-Image | Original weights for 16-megapixel image generation and editing with up to 16 reference images; also on ModelScope. | |
| MiniMaxAI/MiniMax-H3-Image-PE | Optional prompt expander that turns short prompts into the detailed English the model expects. | |
| Comfy-Org/MiniMax-H3-Image | ComfyUI repack for native support. | |
| ComfyUI tutorial | Official ComfyUI guide with text-to-image and editing templates; see also the launch post. | |
| MiniMax API | Official API for text-to-image, image editing and prompt expansion; for mainland China, use platform.minimax.cn. | |
| Hailuo AI | Use H3-Image in the browser or the desktop app; for mainland China, use hailuoai.com and agent.minimax.cn. | |
| ComfyUI API node | Built-in node that generates and edits images with H3-Image through your Comfy account, with an edit template. | |
| MiniMax CLI | Official command line that generates images with H3-Image and edits them from up to 16 references. | |
| MiniMax MCP | Official server that lets agents generate images with H3-Image by default and edit from up to 16 references. |
Quantizations and conversions
| unsloth/MiniMax-H3-Image-FP8 | INT8 ConvRot and FP8 DiT, 20.3 GB each, picked automatically by Unsloth Desktop. | |
| LSDJesus/MiniMax-H3-Image-fp8-nvfp4 | FP8 scaled (21 GB) and NVFP4 (12.5 GB) for ComfyUI; GGUF from Q6_K to Q2_K. | |
| Abiray/MiniMax-H3-Image-GGUF | Pruned-DiT GGUF from 6.88 GB with sensitive layers kept in FP32, for stock ComfyUI-GGUF. | |
| Hearthscale/MiniMax-H3-Image-GGUF | Q8_0, Q5_K_M, Q4_K_M and Q3_K_M for stable-diffusion.cpp. | |
| AtomicChat/MiniMax-H3-Image-GGUF | GGUF tuned with a layer sensitivity scan, 7.6 to 21.4 GB, with published measurements. | |
| lingjiu09/minimax-h3-image-gguf | Q3_K_M and Q4_0, tested on an 8 GB laptop. | |
| tsolful/Minimax_H3_Image_INT6_INT4GPTQ | GPTQ W6A8 (15.6 GiB) and W4A8 (12.8 GiB) for ComfyUI. | |
| speedyrulz/minimax-h3-image-int8-w4a8-hybrid | 14.8 GiB INT8 and W4A8 mix for 16 GB GPUs. | |
| PotatoForge/Minimax-H3-Image-W4A4 | Experimental W4A4 at 11.35 GB for older GPUs. | |
| WaveCut/MiniMax-H3-Image-OrbitQuant-W4A8 | W4A8 DiT and text encoder, 27.5 GB in total. | |
| WaveCut/MiniMax-H3-Image-Pruned | Unused weights removed, 135 GB down to 95 GB with bit-identical text-to-image output; build scripts. | |
| SD-Next SDNQ uint4 | SDNQ 4-bit for diffusers and SD.Next; also 4-bit and 8-bit builds by OzzyGT. | |
| nativ-community/MiniMax-H3-Image-4bit | MLX 4-bit (54.5 GB) for mlx-vlm; an 8-bit build is also available. | |
| VisualInference/MiniMax-H3-Image-MLX-bf16 | MLX BF16 for the minimax-h3-image-mlx runner, with 4-bit and 8-bit builds by cguldogan. | |
| Sawfwair/MiniMax-H3-Image-MLX-8bit | MLX 8-bit export (52 GB) for mere.run. |
Distillation and adapters
| minimax-h3-image-pdmd-2step | Two-step pDMD distillation LoRA, with a demo Space. | |
| Fizgig H3-Image Training Adapter | Rank-16 adapter kept on while training LoRAs. | |
| Felldude/MiniMaxImage_hdr_diffusers_vae | VAE with a fine-tuned decoder that trades pixel accuracy for more high-frequency detail, with metrics against the base VAE. |
Runtimes
| ComfyUI | Native support with official templates. | |
| SGLang | Serving from the main branch, with a cookbook of recipes from DGX Spark to B300 and offload for 12 to 48 GB cards. | |
| vLLM-Omni | Serving with an official recipe, including AMD Instinct. | |
| diffusers | Modular Diffusers blocks for text-to-image and reference editing, released in v0.43.0; see the docs. | |
| LightX2V | Text-to-image and editing with one or more references, on one GPU or several with TP/SP parallelism. | |
| DiffSynth-Studio | Inference, low-VRAM inference, and full or LoRA training. | |
| stable-diffusion.cpp | C/C++ inference with day-0 H3-Image support and GGUF models. | |
| Unsloth | Runs H3-Image in its desktop app and web UI with the INT8 or FP8 DiT. | |
| mmh3 | Rust engine with CUDA and Metal kernels. | |
| h3-image-mlx | MLX generation and editing on M1 to M4 Macs with W8A8 weights, about 7 GB peak memory at 4K. | |
| minimax-h3-image-mlx | BF16 MLX port that reads each block from disk while the previous one computes; 5 GB of memory at 2048×2048. | |
| TensorSharp | .NET engine that runs H3-Image text-to-image on ggml; editing is not implemented yet. | |
| minimax-h3-image-server | Server with a browser UI for generation and reference edits; about 21 GB of VRAM by default, down to 8 GB with W4A8 and weight streaming. | |
| MiniMax-H3-Image-Accel | SGLang tuned for L20 and RTX 5090, BF16 by default with optional FP8 for lower latency. | |
| mlx-vlm | MLX text-to-image and image editing for H3-Image on Apple silicon, with BF16, 8-bit and 4-bit checkpoints. |
Apps
| SwarmUI | Web UI that runs H3-Image with up to 16 prompt images for references or edits. | |
| Pallaidium | Blender video editor add-on with H3-Image text-to-image and edits from up to 10 references, in SDNQ 4-bit for 24 GB GPUs. | |
| diffusion-for-krita | Krita plugin on a ComfyUI backend that uses H3-Image to generate, inpaint and edit with up to 16 images. | |
| styk-paint | Godot add-on that has H3-Image repaint a 3D blockout view by view through ComfyUI, then projects the paint back onto the set. | |
| mmh3-studio | macOS and browser client for a resident mmh3 server, so generations skip the cold start. | |
| h3-image-editor | Windows workspace on ComfyUI for H3-Image generation, reference edits and masked repair, with Qwen, SAM and SeedVR2 tools. | |
| ImageGenerationServer | Multi-user web front end for a local ComfyUI, with H3-Image generation and edits from up to 16 pictures; non-commercial license. | |
| tongflow-modal-minimax-h3-image | TongFlow plugin that runs H3-Image generation, editing and fusion of up to 16 references on Modal. |
API integrations
| HyperFrames | HTML-to-video renderer for agents whose media command can make 2K or 4K images with H3-Image through HeyGen. | |
| AI SDK | Vercel's TypeScript toolkit whose MiniMax provider generates and edits images with H3-Image and its Highspeed version. | |
| Reveric | Desktop studio for novels and short dramas that creates and edits images with H3-Image through the MiniMax API. | |
| Dify MiniMax Image plugin | Tool for Dify apps that generates images with H3-Image and edits from 1 to 16 references through the MiniMax API. | |
| OpenDirector | Agent video workflow tool with an H3-Image adapter that edits from up to 16 references through the MiniMax API. | |
| Octafuse Gateway | Self-hosted API hub with presets for H3-Image generation and edits on MiniMax's OpenAI-compatible Images API. |
ComfyUI nodes and workflows
| ComfyUI-MiniMaxH3-Tools | Nodes that free the text encoder before the DiT loads, so H3-Image runs on a 24 GB GPU with 32 GB of RAM. | |
| ComfyUI-MiniMaxH3-Easy | Its Easy Image node does H3-Image text-to-image and reference edits; type @ in the prompt to insert a reference. | |
| ComfyUI-H3Image-MediaLoader | Loads up to 16 reference images with per-slot crop, rotate and reorder. | |
| H3-Video-LoRA-Loader-for-Image | Runs H3 video LoRAs on H3-Image. | |
| ComfyUI-H3-Continuity-Studio | Node pack for multi-clip H3 video, with an H3-Image node that makes the still frame guiding each shot. | |
| comfyui-minimax-h3-prompt-enhancer-T8 | Prompt enhancement nodes, with H3-Image support since v1.41.0. | |
| ComfyUI-OrbitQuant | Runs the OrbitQuant W4A8 build for generation and edits with up to 16 references. | |
| H3GIF | Has H3-Image draw a storyboard grid in one image, then splits it into frames for a looping GIF. | |
| Comfiest | Agent-driven ComfyUI workflow suite with H3-Image text-to-image and edit graphs for 1, 2 or 4 references. | |
| minimax-h3-colab-workflow | Colab notebooks that run H3 and H3-Image from open weights on an L4, through ComfyUI quants and stable-diffusion.cpp. | |
| comfyui-minimax-image | RunPod template that boots into ComfyUI with the H3-Image models. | |
| minimax-h3-v100-guide | Tested setup for H3 and H3-Image on one 16 GB V100 with 32 GB of RAM: configs, launchers and workflows. |
Training
| Fizgig | LoRA and fine-tune studio with H3-Image support since v9.6. | |
| ai-toolkit | GUI and CLI trainer; trains H3-Image editing when the dataset has control images. |
Prompting
| minimax-h3-image-prompts | Reusable skill for writing generation and editing prompts with up to 16 references. | |
| awesome-minimax-h3-image-prompts | 63 source-linked H3-Image examples with result images and editing references; only 5 include the prompt itself. |
Benchmarks
| h3-image-bench | 600 graded prompts in two rounds at about 16 MP on a 16 GB RTX 5080, with a live results page. | |
| M5 Pro Mac mini benchmark | ComfyUI INT8 timings on a 64 GB M5 Pro Mac mini: about 9 minutes per 4096×4096 image. | |
| 12 GB tests | INT8 on a 12 GB RTX 4070 SUPER with 96 GB of RAM: 11 to 12 s at 4 MP, about 41 s at 16 MP. |
Contributing
Add one row to the matching table: the owner's avatar, the name linked to the project, and one sentence on what it is. Add the same row to README_cn.md, then run python3 scripts/check_readme.py.
An entry needs its own README, docs or model card that says what it is and that it uses MiniMax H3 or MiniMax-H3-Image, locally or through an API. Measurements, VRAM recipes and per-file details belong in docs/run-h3-locally.md.
Acknowledgements
This index is only possible because other people made the models, tools, tests, and documentation it points to. Thank you to the MiniMax, ComfyUI, SGLang, vLLM, NVIDIA, and Unsloth teams, and to the independent maintainers who keep testing H3 on real hardware.
The structure and much of the resource discovery come from the community-maintained wildminder/awesome-minimax-H3. This page follows that work and adds MiniMax's own GitHub and Hugging Face scan.
More specific thanks go to:
- Comfy-Org for the official ComfyUI conversions, workflow templates, and day-one support.
- ModelTC / LightX2V for the Turbo distillation work and public training configuration.
Larryvrhfor the checkpoint comparisons behind the Turbo guidance.Kijaifor the NVIDIA Sol-Attn implementation and benchmark notes.- Alibaba PAI and
aigc-apps/VideoX-Funfor the PDD 8-step Acc LoRAs and the Fun ControlNet-Union branch. - OpenVDN for publishing Video Delta Net's training code alongside the weights.
- FastVideo for the FastH3 line (4-step preview, V2, Trim and Omni), its data-free distillation recipe, and UniServe.
- Nuva Lab for the production grounding behind FastH3, the FastH3-Omni co-development and reference dataset, and the production hardware behind the UniServe benchmarks.
- Pika Labs for the six-step Flex adapters and inference examples; NVIDIA for the released draft refiner and fast decoder, building on LBH, LightX2V and LMS components.
IAmIronMan42for the training work and the documented fixes.- Salvatore Sanfilippo (antirez) for the standalone H3 C/Metal inference engine.
- Every quantizer and workflow maintainer represented above. Their files, testing time, and write-ups make local H3 use much easier.
If you spot a wrong number, broken link, or missing compatibility note, please open an issue or send a correction.
Contact
For MiniMax H3 questions, contact model@minimax.io.
