This repository contains third-party ComfyUI compatibility conversions of the original MiniMax-H3 Turbo LoRA created by larryvrh.
The Turbo LoRA enables few-step MiniMax-H3 audio-video generation, producing joint video and synchronized audio using substantially fewer sampling steps than the standard MiniMax-H3 workflow.
This repository includes compatibility conversions of the initial Turbo weights as well as later checkpoint-500, checkpoint-850, and v4 step-600 releases.
These LoRAs are intended for the pruned/curve-form MiniMax-H3 checkpoint used by ComfyUI.
The original LoRA weights, training work, distillation method, audio/video sampling implementation, documentation and source files were created and released by larryvrh.
I did not train, distill or create the original Turbo LoRA weights.
This repository provides modified compatibility versions intended to allow the compatible LoRA adapters to load with ComfyUI's built-in MiniMax-H3 LoRA loader when using the pruned/curve-form model.
Full credit for the original MiniMax-H3 Turbo LoRA, training and distillation method belongs to larryvrh.
The MiniMax-H3 Turbo LoRA is designed to generate joint video and synchronized stereo audio using dramatically fewer sampling steps than the standard MiniMax-H3 generation process.
The original project demonstrated generation using substantially fewer sampling steps than the normal MiniMax-H3 workflow, providing a large reduction in sampling time.
The newer Turbo checkpoints continue to improve upon the initial preview release.
Refer to the original repository for the author's current documentation, training updates, standalone generation implementation and sampling recommendations:
Without that namespace, ComfyUI may report warnings such as:
[WARNING] lora key not loaded: blocks.0.attn.out_proj.lora_A.weight
The following namespace conversion was applied to the retained tensors:
blocks.* → diffusion_model.blocks.*
token_refiner.* → diffusion_model.token_refiner.*
The retained tensor values, BF16 dtype, tensor shapes, LoRA rank and A/B orientation were preserved.
Conversion Details
Each original full-model Turbo LoRA contained:
Component
Count
Source tensors
518
Source LoRA A/B pairs
259
Each pruned-model compatibility conversion contains:
Component
Count
Retained tensors
416
Retained LoRA A/B pairs
208
Pruned AdaLN tensors
102
Pruned AdaLN A/B pairs
51
The pruned adapters were:
50 AdaLN projection pairs from blocks.0 through blocks.49;
1 AdaLN projection pair from final_layer.adaln_proj.linear.
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 retained attention, MLP and token-refiner tensors have dimensions compatible with the pruned/curve-form MiniMax-H3 checkpoint.
Validation
The converted LoRA structure was checked for:
complete LoRA A/B adapter pairs;
no orphan A or B tensors;
retained keys using the diffusion_model.* namespace;
no duplicated diffusion_model.diffusion_model.* prefixes;
no incompatible adaln_proj tensors;
preserved BF16 dtype;
preserved tensor shapes;
preserved LoRA ranks;
separate EMA and non-EMA weights;
no unresolved retained keys during ComfyUI loader-level validation.
The original source files were not overwritten.
Optional Acceleration
The LoRAs have been tested working in ComfyUI with optional acceleration methods including:
SageAttention
Sol Attention
Gradient
Spectrum
These acceleration methods are optional and are not included in this repository.
They may also be combined with the Turbo LoRA depending on your ComfyUI workflow and available custom nodes.
Important Limitation
These files are partial ComfyUI compatibility conversions intended specifically for the pruned/curve-form MiniMax-H3 architecture.
The original full-model Turbo LoRAs contain AdaLN adapters that do not directly match the corresponding dimensions of the pruned/curve-form MiniMax-H3 checkpoint.
Those incompatible AdaLN tensors are therefore excluded from these compatibility versions.
The retained attention, MLP and token-refiner LoRA tensors remain compatible with the pruned architecture.
Because these are partial compatibility conversions, they should not be assumed to be numerically identical in behaviour to using the original full Turbo LoRA with its intended full MiniMax-H3 base.
Usage
Place the LoRA files inside:
ComfyUI/models/loras/
Use them with the pruned/curve-form MiniMax-H3 checkpoint for ComfyUI.
The node includes a MiniMax-H3 Turbo LoRA loader and dedicated sampling functionality designed for MiniMax-H3's video/audio scheduling behaviour.
A workflow is also included in the custom-node repository.
The custom node evolves alongside the upstream Turbo weights, so users choosing that implementation should follow the settings documented in its current README.
The settings documented in this repository refer to the tested pruned-model compatibility workflow, where the recommended starting point is:
These LoRAs are intended for the pruned/curve-form MiniMax-H3 checkpoint used by ComfyUI.
The compatibility conversions are particularly useful when the original full-model LoRA cannot be loaded directly because of incompatible AdaLN dimensions.
Optional acceleration methods confirmed working with these conversions include:
SageAttention
Sol Attention
Gradient
Spectrum
Accelerators can generally be stacked where the corresponding ComfyUI implementations support it.
Naming Notes
The upstream project contains multiple generations of Turbo training.
Older naming
Files such as:
minimax_h3_turbo_4step_ema_ckpt500.safetensors
and:
minimax_h3_turbo_4step_ema_ckpt850.safetensors
belong to the older training line.
In these filenames, 4step refers to the few-step Turbo sampling target.
Newer v4 naming
Files such as:
minimax_h3_turbo_v4_step600_ema.safetensors
use the newer naming convention.
Here:
v4 refers to the newer training recipe;
step600 refers to training step 600;
ema refers to the exponential-moving-average weights.
The corresponding pruned ComfyUI compatibility conversion in this repository is:
This repository contains third-party, partial ComfyUI compatibility versions for the pruned/curve-form MiniMax-H3 checkpoint, including:
initial Turbo preview weights;
checkpoint-500 variants;
checkpoint-850 variants;
v4 step-600 variants.
I do not claim authorship of:
the original Turbo LoRA weights;
the original LoRA training;
the original distillation method;
the original few-step sampling method;
the original dual video/audio scheduling implementation;
the original generation scripts;
the original documentation;
the MiniMax-H3 base model;
the pruned/curve-form MiniMax-H3 checkpoint.
The modifications in this repository are compatibility conversions intended to make the compatible portions of the upstream Turbo LoRAs usable with the pruned/curve-form MiniMax-H3 architecture in ComfyUI.
Please retain attribution to larryvrh and comply with the licences and terms of the original LoRA and MiniMax-H3 base model when using or redistributing these files.
