terminusresearch/flux-lokr-garfield-masked

simpletuner-lora

This is a LyCORIS adapter derived from black-forest-labs/FLUX.1-dev.

The main validation prompt used during training was:

A photo-realistic image of a cat

Validation settings

• CFG: 3.0
• CFG Rescale: 0.0
• Steps: 20
• Sampler: None
• Seed: 42
• Resolution: 1776x512

Note: The validation settings are not necessarily the same as the training settings.

You can find some example images in the following gallery:

Prompt

unconditional (blank prompt)

Negative Prompt

blurry, cropped, ugly

Prompt

a comic strip of garfield, by jim davis. the first panel has garfield saying Help!. the second panel has garfield saying My clungus is leaking! and the third panel has Odie saying uh oh!

Negative Prompt

blurry, cropped, ugly

Prompt

a comic strip by jim davis, showcasing odie in his full demonic form while garfield cowers in the background

Negative Prompt

blurry, cropped, ugly

Prompt

a picture of garfield in walmart, shopping amongst the real people

Negative Prompt

blurry, cropped, ugly

Prompt

A photo-realistic image of a cat

Negative Prompt

blurry, cropped, ugly

The text encoder was not trained. You may reuse the base model text encoder for inference.

Training settings

• Training epochs: 2
• Training steps: 2000
• Learning rate: 0.0001
• Effective batch size: 2
  • Micro-batch size: 2
  • Gradient accumulation steps: 1
  • Number of GPUs: 1
• Prediction type: flow-matching
• Rescaled betas zero SNR: False
• Optimizer: optimi-lion
• Precision: bf16
• Quantised: Yes: fp8-quanto
• Xformers: Not used
• LyCORIS Config:

{
    "algo": "lokr",
    "multiplier": 1.0,
    "linear_dim": 10000,
    "linear_alpha": 1,
    "factor": 16,
    "apply_preset": {
        "target_module": [
            "Attention",
            "FeedForward"
        ],
        "module_algo_map": {
            "Attention": {
                "factor": 16
            },
            "FeedForward": {
                "factor": 8
            }
        }
    }
}

Datasets

garfield

• Repeats: 0
• Total number of images: 2206
• Total number of aspect buckets: 4
• Resolution: 512 px
• Cropped: False
• Crop style: None
• Crop aspect: None

Inference

import argparse
import torch
from helpers.models.flux.pipeline import FluxPipeline as DiffusionPipeline
from lycoris import create_lycoris_from_weights
from huggingface_hub import hf_hub_download

def generate_image(pipeline, prompt, output_file, num_inference_steps, width, height, guidance_scale, seed, device):
    # Set device
    pipeline.to(device)

    # Generate image
    generator = torch.Generator(device=device).manual_seed(seed)
    image = pipeline(
        prompt=prompt,
        num_inference_steps=num_inference_steps,
        generator=generator,
        width=width,
        height=height,
        guidance_scale=guidance_scale,
    ).images[0]

    # Save image
    output_file = "output.png"
    image.save(output_file, format="PNG")
    print(f"Image saved as {output_file}")

def main():
    parser = argparse.ArgumentParser(description="Generate images using a custom diffusion pipeline with LoRA weights.")
    parser.add_argument("--model_id", type=str, default='black-forest-labs/FLUX.1-dev', help="Model ID from Hugging Face Hub.")
    parser.add_argument("--adapter_id", type=str, default='pytorch_lora_weights.safetensors', help="LoRA weights file.")
    parser.add_argument("--lora_scale", type=float, default=1.0, help="Scale for LoRA weights.")
    parser.add_argument("--output_file", type=str, default="output.png", help="Output file name for the generated image.")
    parser.add_argument("--num_inference_steps", type=int, default=30, help="Number of inference steps.")
    parser.add_argument("--guidance_scale", type=float, default=3.5, help="Guidance scale for the generation.")
    parser.add_argument("--seed", type=int, default=1641421826, help="Random seed for reproducibility.")
    parser.add_argument("--device", type=str, default='cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu', help="Device to run the model on.")
    
    args = parser.parse_args()

    # Load model and weights
    hf_hub_download(repo_id="terminusresearch/flux-lokr-garfield-nomask", filename=args.adapter_id, local_dir="./")
    pipeline = DiffusionPipeline.from_pretrained(args.model_id, torch_dtype=torch.bfloat16)

    # Apply LoRA weights
    wrapper, _ = create_lycoris_from_weights(args.lora_scale, args.adapter_id, pipeline.transformer)
    wrapper.merge_to()

    print("Model loaded successfully. Ready to generate images.")
    
    while True:
        user_input = input("Enter a prompt or 'quit' to exit: ")
        if user_input.lower() == 'quit':
            break

        # Check for resolution command
        if user_input.startswith("resolution:"):
            resolution = user_input.split(":")[1]
            width, height = map(int, resolution.split("x"))
            print(f"Resolution set to {width}x{height}")
            continue

        prompt = user_input
        output_file = args.output_file.replace(".png", f"_{prompt.replace(' ', '_')}.png")

        # Use default or previously set resolution
        width = locals().get('width', 1024)
        height = locals().get('height', 1024)

        generate_image(
            pipeline=pipeline,
            prompt=prompt,
            output_file=output_file,
            num_inference_steps=args.num_inference_steps,
            width=width,
            height=height,
            guidance_scale=args.guidance_scale,
            seed=args.seed,
            device=args.device
        )

if __name__ == "__main__":
    main()