Sign In

WackScratch - MiniMax H3 Style LoRA

0

Updated: Oct 8, 2026

style

Not loaded

Download

1 variant available

bf16 SafeTensor

wackscratch_style_h3_ep50.safetensors

BF16, good balance • 68.31 MB

Verified:

Type
LoRA
Stats

53

0

Generation

Not loaded

Reviews
Published

Oct 8, 2026

Base Model

MiniMax H3

Training
Epochs: 50
Usage Tips
Strength: 1
Hash
AutoV2
4188280DFD
Trigger Words
wackscratchstyle
default creator card background decoration
Followers - 82

82

Likes - 366

366

Generation, training and LoRA distribution on Civitai are covered by Civitai’s own license agreement with MiniMax. If you download these weights and run them yourself, your use is instead governed by the MiniMax H3 Community License Agreement, whose grant excludes the European Union, the United Kingdom, the Republic of Korea and the United States of America.

MiniMax H3

WackScratch look. A STYLE LoRA for MiniMax H3 video - it changes the look, not the subject.

EXPERIMENTAL: lightly tested - the sample clips on this page are the only test so far. Strength: start at 1.0, but this is a drawn / animated style and H3's base pulls toward photorealism, so you may need to turn it UP a little - 1.1 to 1.2 holds the look best; past that it starts to break. Tell me what you find.

Trigger word: wackscratchstyle (put it first in the prompt). Strength 1.0 to 1.2. One file: epoch 50.

Base: MiniMax H3 (fl2va). Loads in ComfyUI with LoraLoaderModelOnly (the file must sit in models/loras/).

How it was made: 150-image Krea 2 dataset generated with the source style LoRA, every subject unique, 50% people, cinematic shot vocabulary in every caption; trained with Fizgig "MiniMax H3 Style (LoRA 8)" - rank 8, LR 2e-4, 50 epochs, blocks 0-3,6-47, adamw, shift 0.6667.

Best results: give H3 a pencil-sketch storyboard plus character references and let the LoRA fill in the look. With a photoreal start frame it mostly reinforces what is already there.

Samples: two text-to-video clips, rendered locally in WanGP with MiniMax H3 FL2VA (pruned 20B) at 1920x1080 "Two Phases" (half-res pass -> latent upscale -> refine), 4 steps, euler, seed 77, 5 s; a 4-step distill LoRA (larryvrh v4) on phase 1 and this LoRA at 1.1 on both phases. Plain text prompts with no reference images, so the look you see is the LoRA. Audio stripped.

Credit: derived from "WackScratch" by Daalis (Krea 2) - https://civitai.com/models/1259116/wackscratch?modelVersionId=3287012 - please support the original creator.