hugovntr/flux-schnell-realism

LoRA checkpoint aiming to improve FLUX.1 [schnell] photorealism.

The weights are provided as .safetensors file, so they are suited for every use-cases, including but not limited to:

• SD-WebUI
• ComfyUI
• InvokeAI
• Forge
• diffusers
• ...

Changelog

• Sep. 21st 2024 - Uploaded v2.3, trained on 1000+ synthetic images manually captionned
• Sep. 7th 2024 - Uploaded v1, trained on 100+ synthetic images manually captionned

Training details

This checkpoint has been trained with a small dataset of 1000+ synthetic images generated and upscaled using Runtime44 All images have been manually captionned without any specific trigger word.

Training was done using kohya-ss scripts on a local machine with 16Gb of VRAM and took ~1h

Usage

With 🧨Diffusers library

from diffusers import AutoPipelineForText2Image
import torch

pipeline = AutoPipelineForText2Image.from_pretrained('black-forest-labs/FLUX.1-schnell', torch_dtype=torch.bfloat16).to('cuda')
pipeline.load_lora_weights('hugovntr/flux-schnell-realism', weight_name='schnell-realism_v1')
image = pipeline('Moody kitchen at dusk, warm golden ...').images[0]
image.save("output.png")