MiniMax-H3 Turbo LoRAs for ComfyUI

This repository contains third-party MiniMax-H3 Turbo LoRA conversions for ComfyUI, including:

pruned/curve-form compatibility conversions of the Turbo LoRAs released by larryvrh;
dynamically resized 4-step and 8-step v1.0 LoRAs derived from the official ComfyUI weights in lightx2v/Minimax-h3-Turbo.

The Turbo LoRAs accelerate MiniMax-H3 joint video and synchronized-audio generation by reducing the required number of sampling steps.

Recommended pruned-model LoRA:
minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors

Recommended starting settings:
6–8 steps · Euler sampler · Beta scheduler · LoRA strength 1.0

Update — LightX2V v1.0 Dynamic-Rank LoRAs

Two resized BF16 LoRAs have been added from the official ComfyUI-compatible weights published in:

lightx2v/Minimax-h3-Turbo

New files

File	Source	Description
`minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_resized_avg_rank_21_bf16.safetensors`	`minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors`	Dynamically resized 4-step 768p LightX2V LoRA.
`minimax_h3_fl2v_turbo_8step_v1.0_comfyui_resized_avg_rank_21_bf16.safetensors`	`minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors`	Dynamically resized 8-step LightX2V LoRA.

These files were already supplied by LightX2V in the correct ComfyUI MiniMax-H3 layout. The update applies dynamic-rank compression without changing the model namespace or QKV layout expected by ComfyUI.

How the new LoRAs were resized

The LoRAs were resized using an exact compact SVD calculated directly from each LoRA A/B factor pair.

The process was:

Treat every attention, MLP and token-refiner projection independently.
Separate the fused QKV LoRA into its original Q, K and V projections.
Calculate the exact singular spectrum of each effective B @ A update without materializing the full dense weight matrix.
Select a dynamic rank independently for each projection.
Preserve one additional safety rank after the selected Frobenius-energy threshold.
Rebuild the factors using balanced singular values:
- A = sqrt(S) × Vh
- B = U × sqrt(S)
Re-fuse Q, K and V using a block-diagonal B matrix and concatenated A matrix.
Bake the original alpha / rank scale into lora_B.
Remove the now-unnecessary alpha tensors.
Save all resulting LoRA tensors as BF16.

The target was the same true average projection rank represented by an avg_rank_21 filename:

Rank sum: 6778
Projection count: 312
Average projection rank: 21.724359
Filename rank: floor(21.724359) = 21

Q, K and V are counted separately when calculating the average. The rank of a fused QKV tensor is the sum of its three independently selected projection ranks.

Resize results

LoRA	Size	Average rank	Average Frobenius norm retained	Baked scale
4-step v1.0 768p	284.36 MiB	21.724359	94.95%	1.0
8-step v1.0	311.89 MiB	21.724359	97.72%	0.0625

The thresholds were solved independently because the singular-value distributions differ between the 4-step and 8-step LoRAs. Using one fixed threshold would not produce the same average rank for both files.

Validation of the resized files

Each resized file contains:

Component	Count
BF16 tensors	416
LoRA A/B pairs	208
Independently ranked projections	312
QKV modules	52
Alpha tensors	0

Validation confirmed:

all tensors are finite BF16 values;
every LoRA A tensor has a matching LoRA B tensor;
all 208 adapters are recognized by ComfyUI;
QKV rank columns remain block-diagonal and correctly ordered as Q, K and V;
the effective original LoRA strength is preserved;
the exact average projection rank is 21.724359;
no incompatible or unresolved tensors remain.

Workflow Example

The following output was generated with the recommended pruned-model workflow using the v4 step-600 EMA LoRA.

Current workflow

Download the current ComfyUI workflow

Workflow: fl_minimax_h3_turbo_lora_example_workflow.json
LoRA: minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors
Steps: 6-8
Sampler: euler
Scheduler: beta
LoRA strength: 1.0
Video sigma shift: 12
Audio sigma shift: 4-6

The previous workflow is retained here:

Download the previous workflow

Its corresponding example output is:

View the previous workflow output

Included Files

LightX2V dynamic-rank update

File	Description
`minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_resized_avg_rank_21_bf16.safetensors`	Exact dynamic-SVD resize of the official LightX2V 4-step v1.0 768p ComfyUI LoRA.
`minimax_h3_fl2v_turbo_8step_v1.0_comfyui_resized_avg_rank_21_bf16.safetensors`	Exact dynamic-SVD resize of the official LightX2V 8-step v1.0 ComfyUI LoRA.

Pruned/curve-form compatibility conversions

File	Description
`minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors`	Recommended. Pruned-model conversion of the v4 step-600 EMA checkpoint.
`minimax_h3_turbo_v4_step600_pruned_comfyui.safetensors`	Non-EMA v4 step-600 variant.
`minimax_h3_turbo_4step_ema_ckpt850_pruned_comfyui.safetensors`	EMA checkpoint-850 conversion.
`minimax_h3_turbo_4step_ckpt850_pruned_comfyui.safetensors`	Non-EMA checkpoint-850 conversion.
`minimax_h3_turbo_4step_ema_ckpt500_pruned_comfyui.safetensors`	EMA checkpoint-500 conversion.
`minimax_h3_turbo_4step_ckpt500_pruned_comfyui.safetensors`	Non-EMA checkpoint-500 conversion.
`minimax_h3_turbo_4step_ema_pruned_comfyui.safetensors`	Initial EMA Turbo preview conversion.
`minimax_h3_turbo_4step_pruned_comfyui.safetensors`	Initial non-EMA Turbo preview conversion.

Workflows and assets

File	Description
`fl_minimax_h3_turbo_lora_example_workflow.json`	Current recommended ComfyUI workflow.
`fl_minimax_h3_turbo_lora_example_workflow_old.json`	Previous workflow retained for comparison.
`assets/MiniMax_H3_audio.mp4`	Current workflow example output.
`assets/MiniMax_H3_00060-audio.mp4`	Previous workflow example output.

EMA and non-EMA files contain different tensor values and are retained separately.

Which File Should I Use?

LightX2V 4-step v1.0

Use:

minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_resized_avg_rank_21_bf16.safetensors

Choose this for the newer LightX2V 4-step 768p FL2V workflow when maximum sampling speed is preferred.

LightX2V 8-step v1.0

Use:

minimax_h3_fl2v_turbo_8step_v1.0_comfyui_resized_avg_rank_21_bf16.safetensors

Choose this for the newer LightX2V 8-step workflow when prioritizing quality and motion consistency over the minimum possible step count.

Recommended pruned-model conversion

For the existing pruned/curve-form workflow in this repository, start with:

minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors

Recommended settings:

Steps: 6-8
Sampler: euler
Scheduler: beta
LoRA strength: 1.0
Video sigma shift: 12
Audio sigma shift: 4-6

The non-EMA, checkpoint-500, checkpoint-850 and initial preview files are provided primarily for testing, comparison and reproducibility.

Why the Pruned Compatibility Conversion Was Needed

The original full-model Turbo LoRAs use keys such as:

blocks.0.attn.qkv_proj.lora_A.weight

ComfyUI expects the corresponding keys under the model namespace:

diffusion_model.blocks.0.attn.qkv_proj.lora_A.weight

The following namespace conversion was applied:

blocks.* → diffusion_model.blocks.*
token_refiner.* → diffusion_model.token_refiner.*

The original full-model LoRAs also contain AdaLN adapters whose dimensions do not match the pruned/curve-form MiniMax-H3 checkpoint. Those incompatible adapters were removed.

Pruned conversion structure

Component	Count
Source tensors	518
Source LoRA A/B pairs	259
Retained tensors	416
Retained LoRA A/B pairs	208
Removed AdaLN tensors	102
Removed AdaLN A/B pairs	51

The retained adapters include:

all main-block attention adapters;
all main-block MLP adapters;
all token-refiner attention adapters;
all token-refiner MLP adapters.

The removed adapters consist of:

50 AdaLN projection pairs from blocks.0 through blocks.49;
one AdaLN pair from final_layer.adaln_proj.linear.

These pruned files are partial compatibility conversions and should not be assumed to behave identically to the original full-model LoRA used with its intended full MiniMax-H3 base model.

The newer LightX2V update does not use this pruning process because its source files were already released in the required ComfyUI MiniMax-H3 layout.

Usage

Place the LoRA files in:

ComfyUI/models/loras/

Load them with the compatible MiniMax-H3 model and workflow.

A reasonable starting strength is:

1.0

If adjustment is needed, test approximately:

0.8-1.2

When comparing LoRAs, keep the following identical:

prompt;
seed;
resolution;
frame count;
base model;
sampler;
scheduler;
LoRA strength;
sigma shifts;
workflow.

Sampling Guidance

4 steps

Maximum speed and useful for rapid previews.

6 steps

A good speed-to-quality compromise for the v4 step-600 EMA pruned conversion.

8 steps

Recommended when prioritizing quality, motion consistency and synchronized audio stability.

For the existing pruned-model workflow:

8 steps
Euler sampler
Beta scheduler

Follow the LightX2V project recommendations when using the new v1.0 FL2V LoRAs:

lightx2v/Minimax-h3-Turbo

Audio and Scheduler Notes

MiniMax-H3 generates video and synchronized audio as separate streams with different scheduling requirements.

Incorrect audio scheduling may cause:

severe distortion;
noise-like output;
unstable audio energy;
blown-out audio;
completely broken audio.

If audio is unstable, check:

sampler and scheduler configuration;
audio sigma shift;
video sigma shift;
ComfyUI version;
MiniMax-H3 sampling implementation;
custom sampler nodes;
the workflow intended for the selected LoRA.

Broken audio should not automatically be attributed to the LoRA.

Optional Acceleration

These compatibility LoRAs have been used with optional ComfyUI acceleration methods including:

SageAttention
Sol Attention
Gradient
Spectrum

These methods are optional and are not included in this repository. Compatibility depends on the corresponding ComfyUI implementation and workflow.

Original Projects and Attribution

LightX2V v1.0 LoRAs

The new dynamically resized 4-step and 8-step LoRAs are derived from the official ComfyUI-compatible files released by:

lightx2v/Minimax-h3-Turbo

Credit for the source LoRA weights, training and release belongs to the LightX2V project and its contributors.

This repository only provides the dynamic-rank BF16 resize described above.

Original Turbo LoRAs

The older Turbo LoRA weights, training work and distillation method were released by:

larryvrh/MiniMax-H3-Turbo-Lora

Original ComfyUI Turbo custom node:

ComfyUI-MiniMax-H3-Turbo

The custom node includes a dedicated MiniMax-H3 Turbo LoRA loader and sampling implementation for MiniMax-H3 video/audio scheduling.

This repository does not claim authorship of:

the original LoRA weights;
LoRA training or distillation;
the few-step sampling methods;
MiniMax-H3 video/audio scheduling;
the MiniMax-H3 base model;
the pruned/curve-form checkpoint;
the LightX2V source weights;
the original project documentation or generation code.

The modifications provided here are limited to:

pruned/curve-form ComfyUI compatibility conversions;
namespace adjustments;
removal of incompatible AdaLN adapters where required;
exact dynamic-rank SVD resizing of the LightX2V ComfyUI LoRAs;
BF16 safetensors packaging and validation.

Please retain attribution and comply with the licenses and terms of the original LoRAs and MiniMax-H3 base model when using or redistributing these files.

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for drbaph/MiniMax-H3-Turbo-Lora-ComfyUI

Base model

MiniMaxAI/MiniMax-H3

Finetuned

Comfy-Org/MiniMax-H3

Adapter

(14)

this model