I found MiniMax-H3-Turbo-Lora ~ 4-step

#5
by sunnyboxs - opened

I found MiniMax-H3-Turbo-Lora, but it's currently in training. The authors released one weight, but it doesn't yet support ComfyUI.

MiniMax-H3 Turbo LoRA — 4-step audio-video generation (early preview)

https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora

minimax_h3_turbo_4step.safetensors

It currently cannot be loaded in ComfyUI; could you convert and test it to see if it actually works?
4-step sampling would significantly boost generation speed.

Waiting for 8 steps to accelerate LoRA

It worked well for me after I used this script.

https://huggingface.co/Evados/DiffSynth-Studio-Lora-Wan2.1-ComfyUI/blob/main/minimax_h3_turbo_4step_testv1_comfy.safetensors

import sys
from safetensors.torch import load_file, save_file

input_lora = sys.argv[1]
output_lora = sys.argv[2]

src = input_lora
dst = output_lora

lora = load_file(src)

out = {}

for k, v in lora.items():

    new = k

    if (
        k.startswith("blocks.")
        or k.startswith("token_refiner.")
        or k.startswith("final_layer.")
    ):
        new = "diffusion_model." + k

    out[new] = v

    print(new)

save_file(out, dst)

print("done")

sampler: euler
sheduler: beta
8 steps and lora strength 1.8 res: 1280x720 + RTX Super resolution.
8/8 [06:16<00:00, 47.03s/it] On a RTX 4070ti super 16 gb vram and 32gb ram.

Don't work with the pruned model.

Sign up or log in to comment