sdxl-flash-lora

SDXL Flash with LoRA in collaboration with Project Fluently

Introducing the new fast model SDXL Flash, we learned that all fast XL models work fast, but the quality decreases, and we also made a fast model, but it is not as fast as LCM, Turbo, Lightning and Hyper, but the quality is higher. Below you will see the study with steps and cfg.

--> Work with LoRA <--

• Trigger word:

  <lora:sdxl-flash-lora:0.55>
  

• Optimal LoRA multiplier: 0.45-0.6 (the best - 0.55)
• Optimal base model: fluently/Fluently-XL-v4

Steps and CFG (Guidance)

Optimal settings

• Steps: 6-9
• CFG Scale: 2.5-3.5
• Sampler: DPM++ SDE

Diffusers usage

pip install torch diffusers

import torch
from diffusers import StableDiffusionXLPipeline, DPMSolverSinglestepScheduler
# Load model.
pipe = StableDiffusionXLPipeline.from_pretrained("sd-community/sdxl-flash", torch_dtype=torch.float16).to("cuda")
# Ensure sampler uses "trailing" timesteps.
pipe.scheduler = DPMSolverSinglestepScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")
# Image generation.
pipe("a happy dog, sunny day, realism", num_inference_steps=7, guidance_scale=3).images[0].save("output.png")