import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("Comfy-Org/MiniMax-H3-Image", dtype=torch.bfloat16, device_map="cuda")
pipe.load_lora_weights("syntaxdiffusion/tanaami-style-h3-image-lora")
prompt = "T4N4AM1"
image = pipe(prompt).images[0]Tanaami-style LoRA for MiniMax H3 Image
A style LoRA for MiniMax H3 Image: dense psychedelic pop-art illustration with saturated flat colour, heavy contour lines and rainbow ribbons.
- File:
tanaami_h3_image_lora_2000.safetensors(rank 16, 2,000 steps) - Trigger: start the prompt with
T4N4AM1 style illustration. - Strength: 0.75 to 1.0. At 0.5 the base model's look still dominates.
Use
ComfyUI: put the file in models/loras and add a LoraLoaderModelOnly node between the H3 Image
model loader and the sampler. Sample at 6 steps, CFG 1, euler / simple, around 4 MP.
Write prompts as full sentences. The style is strongest when the prompt names its motifs (eyes, teeth, rainbow ribbons, flames, stripes, collage with no empty space); ordinary subjects come out as a bold, colourful illustration with a lighter touch of the style.
Samples
Left to right: base model, LoRA at 0.5, 0.75 and 1.0. Same prompt and seed.
Training
Trained with AI Toolkit (minimax_h3_image architecture) on 17 images at about 4 MP, batch 1,
learning rate 1e-4, adamw8bit, 2,000 steps, with the ostris/minimax_h3_training_adapter assistant LoRA.
Licence
This LoRA is a derivative of MiniMax H3 Image and is subject to the MiniMax H3 Image Community License.
- Downloads last month
- 104
Model tree for syntaxdiffusion/tanaami-style-h3-image-lora
Base model
MiniMaxAI/MiniMax-H3