Trying out the MiniMax H3 Turbo LoRA in ComfyUI

It has been three years (three days) since the release of MiniMax H3, hasn't it?

Since quite a bit of time has passed, and the acceleration methods have evolved since I published them two years (two days) ago, I will give them a try.

It feels like a LoRA for generating videos in 4 steps has been released.
I will try using this.

Environment

OS: Windows 11
GPU: GeForce RTX 4090
CPU: i9-13900KF
memory: 128G

Setup Procedure

CUDA 13 is required, so please refer to procedures like this one to update your CUDA version.

Please install the following custom nodes and update ComfyUI to the latest version.

cd ComfyUI\custom_nodes
git clone https://github.com/kijai/ComfyUI-SolAttn_triton.git

Models

minimax_h3_fl2va_pruned_int8_convrot.safetensors download and place in 'ComfyUI\models\diffusion_models'

qwen3vl_32b_minimax_h3_int8_convrot.safetensors download and place in 'ComfyUI\models\text_encoders'
(For 50X0 series, qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors is also acceptable)

minimax_h3_video_vae_fp16.safetensors download and place in 'ComfyUI\models\vae'

minimax_h3_video_vae_int8_convrot.safetensors download and place in 'ComfyUI\models\vae'

minimax_h3_turbo_4step_ckpt500_pruned_comfyui.safetensors download and place in 'ComfyUI\models\loras'

Workflow

The workflow is as follows.

I have provided a brief explanation of the workflow below, so please take a look if you are interested.

Generation Results

The generation results are as follows.
Generation time: 210 seconds.
CPU memory: 75GB
GPU memory: 67GB (21GB VRAM + 46GB shared memory)

Compared to 20 steps without LoRA, it is as follows.
For 20 steps, the memory usage is almost the same, but the generation time is about 10 minutes.

Impressions

I do feel some degradation, to say the least. The audio is catastrophic.
The generation speed is good, though.
I think there is potential for both audio and video to turn out well if optimized, but in its current state, it is only good enough for creating prototypes.

Leftovers

In the end, I only ended up introducing the LoRA and quietly changing the VAE, but I have been tinkering with various things while playing around with it.

The only thing that seems to have a practical use is the following.

Install the custom node below and swap out the VAE.

cd ComfyUI\custom_nodes
git clone https://github.com/Saganaki22/ComfyUI-sol-attn.git

qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors Download and place it in 'ComfyUI\models\text_encoders' (even for a 4090).

The generation time is 230 seconds, which is slightly slower, but the CPU memory is 60GB and the GPU memory is 55GB (19GB VRAM + 36GB shared memory).
It seems more suitable for generating higher resolution or longer duration videos.

Next are the tools that I probably won't use this time.

First is this custom node.
EasyCache wins in terms of generation speed.
The audio results are supposedly better, but since the audio breaks when using this LoRA anyway, it has no redeeming qualities.

There are also 4-step LoRAs below, but they either have the same contents or are not intended for pruned models, so I settled on this one.

Update (20260807)

This is the audio-improved version.

Install the following custom nodes and swap out the LoRA.

cd ComfyUI\custom_nodes
git clone https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo.git

minimax_h3_turbo_4step_ckpt500.safetensors and place it in 'ComfyUI\models\loras'

The generation time is 210 seconds, with 62GB of CPU memory and 55GB of GPU memory (19GB VRAM + 36GB shared memory) used.

The audio is fine, and both the memory usage and generation time are optimal, so I feel like this is the way to go.