Minimax-H3-Turbo provides MiniMax-H3 Turbo LoRA checkpoints, plus Diffusers batch inference and ComfyUI workflows.
| Model | Tasks | Training resolution |
Training shifts (video / audio) |
Distillation steps (NFE) |
Recommended inference steps (NFE) |
|---|---|---|---|---|---|
|
FL2VA Turbo 4-step v0.1 Diffusers · ComfyUI |
FL2VA / T2VA | 544p mixed aspect ratio |
12 / 3 | 4 | 4 |
|
FL2VA Turbo 8-step v1.0 Diffusers · ComfyUI |
FL2VA / T2VA | 544p mixed aspect ratio |
12 / 3 | 8 | 8 / 4 |
|
FL2VA Turbo 4-step v1.0 768p Diffusers · ComfyUI |
FL2VA / T2VA | 768p 1344x768 |
6 / 3 | 4 | 4 |
|
Ref2VA Turbo 4-step v0.1 Diffusers · ComfyUI |
Ref2VA | 544p mixed aspect ratio |
12 / 3 | 4 | 4 |
For NFE = N, define the N transformer evaluation points on the unshifted grid as
q_i = (N - i) / N, where i = 0, 1, ..., N - 1.
For example, with NFE = 4, video shift = 12, and audio shift = 3, the shared
grid is q = [1, 0.75, 0.5, 0.25], giving video sigma
[1, 0.9730, 0.9231, 0.8000] -> 0 and audio sigma
[1, 0.9000, 0.7500, 0.5000] -> 0; each list therefore uses exactly four NFEs.
The three reference-image resizing policies used by our workflows are based on
the ref_image_size implementation described in ComfyUI's MiniMax H3 R2V reference-image sizing guidance:
| Mode | Behavior | Scale factor before 32-pixel rounding |
|---|---|---|
match |
Matches the reference pixel area to the target canvas while preserving the reference aspect ratio. It never upscales a smaller reference. | min(1, sqrt(target_area / ref_area)) |
max |
Preserves the reference aspect ratio and only scales down references whose short edge exceeds 2048 pixels. | min(1, 2048 / ref_short_edge) |
diffusers |
Preserves the reference aspect ratio and forces the short edge to 2048 pixels, matching the original Diffusers behavior. | 2048 / ref_short_edge |
All three policies keep the reference aspect ratio, use the H3 resolution grid
(dimensions rounded to multiples of 32), and avoid cropping the reference
content. In our distillation training, we use match, so the reference-image
pixel budget follows the target training resolution.
The Ref2VA inference entry point in this repository exposes the same three
policies through --reference-resize-mode and defaults to match. Passing
--reference-resize-mode diffusers restores the original Diffusers behavior
(the fixed 2048-pixel short edge). For our distilled models, we recommend
selecting match so inference uses the same resizing policy as training.
See DIFFUSERS_SETUP_AND_INFERENCE.md for environment setup, checkpoint downloads, test JSON files, and single- or multi-GPU inference commands.
See COMFYUI_SETUP_AND_INFERENCE.md for ComfyUI requirements, model installation, inputs, prompts, and run instructions.
Ready-to-import graphs are in example_workflows. The T2VA and I2VA graphs default to FL2VA Turbo 8-step v1.0; the Ref2VA graph uses Ref2VA Turbo 4-step v0.1.
| Workflow | Task | Default resolution |
|---|---|---|
| video_minimax_h3_t2v_lightx2v_turbo.json | T2VA (text-to-video + audio) | 960×544 (16:9, 0.5 MP) |
| video_minimax_h3_i2v_lightx2v_turbo.json | I2VA / FL2VA (image-to-video + audio) | 864×480 (16:9, 0.4 MP) |
| video_minimax_h3_ref2v_lightx2v_turbo.json | Ref2VA (reference-to-video + audio) | 960×544 (16:9, 0.5 MP) |
All graphs wrap the same MiniMax-H3 subgraph. T2VA leaves first_frame /
last_frame unconnected; I2VA connects a LoadImage to first_frame, with
last_frame optional for first/last-frame interpolation. Ref2VA connects one
or more reference images through the reference-input branch.
For detailed workflow inputs and execution steps, see COMFYUI_SETUP_AND_INFERENCE.md.
- Improve the visual quality and consistency of Ref2VA and FL2VA Turbo.
Some Ref2VA test cases and reference assets are adapted from public showcases on the Hailuo website and from the MiniMax-H3 discussion on Hugging Face. We thank the community contributors for sharing their test assets and prompts.
Special thanks to the MiniMax-AI/MiniMax-H3 project and the MiniMax team for open-sourcing the MiniMax-H3 model.