FLUX.3 Tarot LoRA

A Rider-Waite tarot-style DreamBooth LoRA for the FLUX.3 image (klein) model, trained with the diffusers DreamBooth FLUX.3 trainer. Trigger the style with trtcrd, e.g. "a trtcrd of ..., tarot style".

Inference

import torch
from diffusers.modular_pipelines import Flux3AutoBlocks

pipe = Flux3AutoBlocks().init_pipeline("black-forest-labs/flux-3-image-klein-base")
pipe.load_components(dtype=torch.bfloat16)
pipe.to("cuda")
pipe.load_lora_weights("diffusers-modular/flux3-tarot-lora", weight_name="pytorch_lora_weights.safetensors")

image = pipe(
    prompt="a trtcrd of a dog eating a big bowl of spaghetti, tarot style",
    height=1344, width=800, num_inference_steps=28, negative_prompt="",
    max_sequence_length=512, text_encoder_out_layers=(4, 8, 12, 16, 20, 24, 28, 32),
    output="images",
)[0]
image.save("tarot.png")

Trigger

Use trtcrd in the prompt: "a trtcrd of {your subject}, tarot style".

Training

• Base model: black-forest-labs/flux-3-image-klein-base
• Dataset: multimodalart/1920-raider-waite-tarot-public-domain
• 1500 steps · seed 0 · resolution 1024 · aspect-ratio buckets
• Default weights (pytorch_lora_weights.safetensors): AdamW (8-bit), rank 16, lr 1e-4

Sweep variants

This repo also includes the configs from a small sweep (one setting changed vs the default). Load one with weight_name="<folder>/pytorch_lora_weights.safetensors":