MiniMax H3 Turbo — ComfyUI 4-Step T2V and I2V LoRA
This repository contains an inference acceleration LoRA for the ComfyUI release of MiniMax H3, covering Text-to-Video (T2V) and Image-to-Video (I2V). It targets H3's joint audio-video diffusion path and uses a fixed 4-step Euler inference contract.
This file is a LoRA, not a standalone diffusion checkpoint. It must be paired with the matching BF16 base model from
Comfy-Org/MiniMax-H3.
Important: ComfyUI v0.31.0 audio regression
ComfyUI v0.31.0 changed MiniMax H3 to the new ModelSamplingAV audio-carry sampling path. With the validated 4-step Euler + simple workflow, this change can produce a much higher noise floor, unstable stereo imaging, and excessive high-frequency artifacts. Generation still completes and the saved video contains a valid audio track, but the resulting sound can be clearly degraded compared with ComfyUI v0.30.0.
This is a ComfyUI sampling compatibility issue, not a problem with this LoRA, the MiniMax H3 Audio VAE, or the video-saving node.
Use the following compatibility node to restore the MiniMax H3 audio sampling behavior from ComfyUI v0.30.0 while keeping the rest of the current ComfyUI installation:
Place MiniMax H3 Legacy Audio Sampling after the LoRA and other model patches, connect its model output to both BasicGuider and BasicScheduler, and keep shift_video = 12.0 and shift_audio = 3.0. Do not chain it with the official ModelSamplingMiniMaxH3 node.
Supported generation modes
- Text-to-Video (T2V): generate a video directly from a text prompt with the FL2VA model and LoRA.
- Image-to-Video (I2V): animate an input image using the FL2VA first-frame path; an optional last frame can also guide the ending.
Comparison protocol
Within each comparison group, all results use the same MiniMax H3 BF16 base model, the same inference input, and the following standardized inference target. Measured generation times are shown directly in the corresponding comparison table.
- Sampling steps: 4
- LoRA strength: 1.0
- Video duration: 10 seconds
- Output resolution: approximately 0.9 megapixels per frame
First comparison group
No LoRA | | Ours LoRA |
|---|
| 190 s | 178 s | 180 s |
| | |
Second comparison group — Subway scene
No LoRA | | Ours LoRA |
|---|
| 177 s | 174 s | 175 s |
| | |
Third comparison group — Outdoor lawn scene
No LoRA | | Ours LoRA |
|---|
| 183 s | 174 s | 175 s |
| | |
Recommended inference settings
| Setting | Value |
|---|
| LoRA strength | 1.0 |
| Steps | 4 |
| Sampler | Euler |
| Scheduler | simple |
| CFG | Disabled |
| Sigma path | 1.0 → 0.97297297 → 0.92307692 → 0.8 → 0.0 |
The 4-step sigma path is part of the model contract. Different samplers, schedulers, step counts, or LoRA strengths may work experimentally, but are outside the validated configuration.
ComfyUI usage
- Place the
.safetensors file in ComfyUI/models/loras/.
- Load
minimax_h3_fl2va_pruned_bf16.safetensors as the diffusion model.
- Add the LoRA with strength
1.0.
- On affected ComfyUI v0.31.0 builds, install and connect ComfyUI-MiniMax-H3-LegacySampling.
- For T2V, use the FL2VA workflow without keyframe conditioning. For I2V, use
MiniMax H3 Image to Video with the first-frame input.
- Configure
Euler, simple, and 4 sampling steps.
How it works
Four-step flow update
MiniMax H3 predicts a conditional velocity field for a noisy joint audio-video latent. At each sampling boundary, Euler advances the latent with:
x_next = x + (sigma_next - sigma) * v(x, sigma, condition)
Running the original model with fewer scheduler evaluations is not enough by itself: each interval becomes much larger, so small velocity errors accumulate quickly. This Turbo LoRA modifies the model's velocity prediction for the fixed four-interval path. The acceleration therefore comes from a learned model correction, not merely from skipping scheduler steps.
The supplied adapter is tied to the published sigma boundaries. Euler + simple + 4 steps keeps the network evaluations and integration intervals aligned with that contract.
Joint audio-video denoising
H3 processes text and visual conditioning, audio latents, and video latents inside one denoising model. Shared transformer blocks model cross-modal timing and semantic relationships, while separate final projections predict audio and video velocity outputs. The LoRA covers both shared blocks and both output paths, so the four-step correction is applied coherently to the synchronized audio-video state.
FL2VA conditioning
First and last keyframes are encoded into latent-space keyframe conditions aligned to positions on the target video timeline. Their visual embeddings are also presented to the text/vision encoder before the prompt. This path supports text-only generation as well as control from boundary frames.
Technical profile
- ComfyUI-native LoRA key layout in
safetensors format.
- Covers attention, MLP, conditioning modulation, video output, and audio output paths.
- Designed for MiniMax H3's synchronized audio-video latent denoising.
- Structural loading was verified against ComfyUI successfully.
Notes and limitations
- The exact matching base model and the remaining MiniMax H3 components are required and are not included here.
- Four-step inference trades some fine detail and temporal stability for substantially lower sampling cost.
- Visual fidelity, motion, audio quality, and prompt adherence vary with input quality, aspect ratio, duration, and seed.
- This repository does not modify or redistribute the MiniMax H3 base checkpoints. Follow the upstream model's license and usage terms.
