larryvrh/MiniMax-H3-Turbo-Lora · Hugging Face

Source: https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora

https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora#minimax-h3-turbo-lora–4-step-audio-video-generation-early-previewMiniMax-H3 Turbo LoRA — 4-step audio-video generation (early preview)

A LoRA forMiniMax-H3that renders jointvideo + synchronized stereo audioin4 sampling stepsinstead of the usual ~20 — roughly a 5× speedup in sampling wall-clock.

⚠️**Early prototype — the weightsandthe tooling.**These weights are an early, under-trained checkpoint (not production quality), and the ComfyUI nodes below are prototype code:**functionality and compatibility are not guaranteed.**It is already a clear step up from the base model at 4 steps — sharper detail, cleaner and better-synced audio — but treat everything here as a work-in-progress taste, not a finished product.If something breaks, please open an issueon thenode repo.

https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora#use-it-in-comfyui-recommendedUse it in ComfyUI (recommended)

Custom nodes:Larryvrh/ComfyUI-MiniMax-H3-Turbo— or search**“MiniMax-H3 Turbo”**in ComfyUI-Manager.

1. Install the nodes (Manager, orgit cloneintoComfyUI/custom\_nodes).
2. Download a\.safetensorsfrom this repo intoComfyUI/models/loras/.
3. Start from the officialMiniMax-H3 workflow(text-to-video or image-to-video) and make two changes:- insertMiniMax-H3 Turbo LoRAbetween the model loader and the sampler; - replace the sampler feedingSamplerCustomAdvancedwithMiniMax-H3 Turbo Sampler (4-step), and set the scheduler to4 steps(simple).

Everything else stays as in the official workflow, so both t2v and i2v work. A ready-made t2v workflow is included here (minimax\_h3\_t2v\_turbo\.json) and in thenode repo— drag it into ComfyUI. The custom sampler is required: MiniMax-H3 runs video and audio on two different flow schedules, and a stock sampler over-steps the audio at 4 steps and it breaks.

• Steps: any count**≥ 4works andmore steps look better**— 4 is the fast default, 6 or 8 are cleaner. Keep the scheduler onsimple.
• Base model: use anon-prunedbase (thebf16or fullint8\_convrotDiT). Theprunedvariants (pruned\_int8,pruned\_fp8) use a different time-conditioning layer and arenot compatiblewith this LoRA.

https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora#weightsWeights

Three checkpoints, all bf16, ~744 MB, applied as a standard low-rank update (W\_eff = W \+ lora\_B @ lora\_A, alpha = rank so no extra scaling):

file~stepsnotesminimax\_h3\_turbo\_4step\_ema\_ckpt500\.safetensors~500newest, most-trained — recommended default(time-averaged)minimax\_h3\_turbo\_4step\.safetensors~200initial release, trained weights — a bit sharper on very fast motionminimax\_h3\_turbo\_4step\_ema\.safetensors~200initial release, time-averaged — softer; superseded by the ckpt500 file Training is ongoing; newer checkpoints will land here over time.

https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora#standalone-no-comfyui-graphStandalone (no ComfyUI graph)

generate\.pyis a single self-contained file — loads the base DiT + a LoRA, encodes the prompt, runs the 4-step dual-schedule sampler, decodes and muxes an mp4. It still needs a ComfyUI checkout for the H3 model / VAE / text-encoder definitions:

# ComfyUI (pinned to the commit these weights were validated against)
git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI && git checkout 14b05228cef127ce529bc0c08660770d4af3e9a8
pip install -r requirements.txt && cd ..
pip install -r requirements.txt   # this repo: torch, safetensors, imageio-ffmpeg

# base weights from Comfy-Org/MiniMax-H3 into a models/ tree, then:
python generate.py \
  --comfyui ./ComfyUI \
  --base   models/diffusion_models/minimax_h3_fl2va_bf16.safetensors \
  --lora   minimax_h3_turbo_4step_ema_ckpt500.safetensors \
  --te     models/text_encoders/qwen3vl_32b_minimax_h3_int8_convrot.safetensors \
  --video-vae models/vae/minimax_h3_video_vae_fp16.safetensors \
  --audio-vae models/vae/minimax_h3_audio_vae_fp32.safetensors \
  --prompt "A corgi in a chef hat flipping a pancake, sizzling sounds and a cheerful bark." \
  --width 1344 --height 768 --frames 124 --out corgi.mp4

https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora#notesNotes

• Resolution / duration: width/height are multiples of 32; the short edge is typically 768. Frame count is at 24 fps and snaps to the model’s 17·k+5 grid (124 ≈ 5 s). Validated range124–362 frames (5–15 s).
• VRAM: the base model is large (~33 B); an 80 GB GPU is comfortable (\-\-offload\-adalnin the standalone script trades ~13 GB of VRAM for CPU RAM).
• Steps: 4 is the design point, schedulersimple.
• Audio: 32 kHz stereo, aligned to the video; the two streams ride different flow schedules and are integrated each on its own clock.