MiniMax H3 Turbo LoRA
ComfyUI Guide
A source-checked guide to the community Turbo LoRA preview: what it changes, how to install it, which workflows are documented, and how to test it without turning preview behavior into a universal promise.
Treat this as a community preview, not an official release.
The creator's Hugging Face model card and ComfyUI node repository provide a real weight, an install path, and a sample workflow. They also label the current checkpoint as preview code with known artifacts.
Preview status
The current ckpt850 preview is documented as sharp at low steps, but the author reports plastic-looking skin and over-sharp grain or noise. Expect updates and retest after checkpoint changes.
Not an official MiniMax release
The LoRA and custom nodes are community work. Keep the base H3 model license and the LoRA repository license as separate questions.
Reproducible path
The repository gives a ComfyUI-Manager route, a git clone command, a LoRA folder, and an example T2V workflow.
The adapter targets low-step H3 generation.
Use the author-provided workflow as the reference point. The notes below describe the preview's documented behavior, not a guarantee for every prompt, GPU, base variant, or future checkpoint.
| Area | Documented behavior | Practical reading |
|---|---|---|
| Sampling | The current creator workflow uses a Turbo Sampler with a 4-step schedule. | Keep the matching sampler and scheduler together; do not copy only the LoRA into an unrelated 4-step graph. |
| Audio path | The creator documents separate flow schedules for H3 video and audio. | At low steps, use the Turbo Sampler from the node repository so the audio path is handled separately. |
| Quality trade-off | The current preview is described as sharp, with known plastic-looking skin and over-sharp grain or noise. | Record artifacts with the prompt and seed instead of claiming that the preview improves every scene. |
| Weight type | The model card lists standard low-rank update weights, with a current EMA ckpt850 file and older checkpoints. | Choose one exact file, record its name, and do not compare unnamed checkpoints. |
Match the adapter to the base and workflow.
The compatibility notes below are attributed to the creator's current documentation. ComfyUI nodes and checkpoints are actively evolving, so verify the source before each new setup.
| Item | Status in the creator docs | Do not assume |
|---|---|---|
| Full BF16 base | Listed as a compatible MiniMax H3 base. | That a BF16 result predicts the same artifact profile on a quantized base. |
| INT8 ConvRot base | Listed as a compatible full-base variant. | That an INT8 filename alone identifies the correct workflow. |
| Pruned INT8 or FP8 base | The node documentation says it can detect pruned bases. | That an older node version supports every current pruned variant. |
| T2V | Explicitly documented, with a ready-made example workflow. | That an unrelated sampler can replace the Turbo Sampler at 4 steps. |
| I2V | Explicitly documented as supported by the H3 workflow path. | That reference-image behavior is identical to every T2V prompt. |
| R2V | Not explicitly documented by the creator's current guide. | Do not publish R2V support as confirmed; test it separately if you experiment. |
For the base model, text encoder, and VAE file layout, use the Model Files and Downloads guide and the current official H3 workflow.
Install the node, LoRA, and matching workflow.
1 / Install the custom node
Search for MiniMax-H3 Turbo in ComfyUI-Manager, or install the creator repository manually.
2 / Place the LoRA
Download one safetensors file from the creator model card and put it in the ComfyUI LoRA folder.
3 / Keep base files ready
The Turbo LoRA does not replace the H3 diffusion model, text encoder, or video and audio VAEs.
cd ComfyUI/custom_nodes git clone https://github.com/larryvrh/ComfyUI-MiniMax-H3-Turbo
LoRA folder
ComfyUI/models/loras/
Current preview file
minimax_h3_turbo_4step_ema_ckpt850.safetensors
Approximate size
The model card lists the BF16 weights at approximately 744 MB. Check the live file entry before downloading.
After installing, restart ComfyUI and start from the creator's example workflow or the current official H3 T2V/I2V workflow. Do not assume a stale node graph is compatible with the latest checkpoint.
Change the sampler path, not the whole workflow.
The creator's documented setup keeps the H3 conditioning, VAE decode, and audio output path, then inserts the Turbo LoRA and Turbo Sampler at the model-to-sampler boundary.
Load Diffusion Model → MiniMax-H3 Turbo LoRA → MiniMax-H3 Turbo Sampler → SamplerCustomAdvanced
| Control | Starting point from the creator workflow | Test note |
|---|---|---|
| Sampler node | MiniMax-H3 Turbo Sampler | Use the node supplied by the creator repository. |
| Scheduler | simple | Keep it paired with the provided low-step setup. |
| Steps | 4 for the current ckpt850 preview | Re-test after changing the checkpoint or node version. |
| LoRA strength | Start at the creator default of 1.0 | Adjust only after a baseline; strength changes the sharpness and artifact trade-off. |
T2V
Use the creator's example workflow as the wiring reference.
I2V
Use the same documented Turbo path with the official H3 image-to-video inputs.
R2V
The creator guide does not confirm R2V. Treat any result as an experiment, not a supported recipe.
Compare like with like.
A low-step preview can be useful without being universally faster or better. Make the comparison reproducible before you draw a conclusion.
- Freeze the base model, text encoder, VAEs, prompt, seed, resolution, frame count, and output format.
- Run a baseline with the native H3 workflow and record wall-clock time, steps, GPU, and audio/video issues.
- Run the Turbo workflow with the exact checkpoint filename and the creator's sampler path.
- Record LoRA strength, scheduler, steps, node version, and whether low_vram was enabled.
- Inspect motion, faces, texture, audio sync, audio distortion, and repeated-frame artifacts separately.
- Keep the result marked as preview data; do not turn one prompt or one GPU into a universal benchmark.
Fix the graph before changing the prompt.
| Symptom | Likely cause | First check |
|---|---|---|
| Audio is distorted at 4 steps | A stock sampler is stepping audio and video on one schedule. | Use the creator's MiniMax-H3 Turbo Sampler and matching workflow. |
| Nodes are missing after import | The custom node is not installed, outdated, or not loaded after restart. | Update or reinstall the creator node, restart ComfyUI, and reopen the example workflow. |
| Output is blurry or smeared | LoRA strength, checkpoint, or sampler settings do not match the test. | Start with the creator default and exact current checkpoint, then change one variable. |
| Over-sharp grain or plastic-looking skin | Known preview artifacts or an overly strong LoRA. | Lower strength, record the checkpoint, and compare with the base workflow. |
| Out of memory | Base model, resolution, frame count, or runtime mode exceeds available memory. | Try low_vram or reduce resolution/frame count, then record the change. |
| R2V does not behave as expected | R2V compatibility is not documented by the creator guide. | Return to the documented T2V/I2V path before treating it as a bug in the LoRA. |
Frequently Asked Questions
Is MiniMax H3 Turbo LoRA an official MiniMax release?
No. The creator's model card presents it as a community preview. It is not an official MiniMax release, and the preview has known artifacts.
What does the MiniMax H3 Turbo LoRA installation need?
Install the creator's ComfyUI custom node, place a LoRA safetensors file in ComfyUI/models/loras/, keep the official H3 base files available, and start from the matching Turbo workflow.
Can I use the Turbo LoRA with R2V?
The creator documentation explicitly describes T2V and I2V. It does not document R2V compatibility, so treat R2V as unverified and test it separately.
Why is the audio distorted when I use four steps?
The creator documentation says H3 video and audio use different flow schedules. Use the Turbo Sampler in the provided workflow instead of a stock sampler at the low-step setting.
Use the creator's release and the official base-model sources together.
Last verified: August 7, 2026. The Turbo LoRA is a community preview; check the live model card and node repository before downloading or reporting compatibility.