sdxl-flash-lora / README.md
dreamdrop-art's picture
Update README.md
5b56967 verified
metadata
license: creativeml-openrail-m
library_name: diffusers
pipeline_tag: text-to-image
base_model: fluently/Fluently-XL-v4
tags:
  - safetensors
  - stable-diffusion
  - lora
  - template:sd-lora
  - sdxl
  - flash
  - sdxl-flash
  - lightning
  - turbo
  - lcm
  - hyper
  - fast
  - fast-sdxl
  - sd-community
instance_prompt: <lora:sdxl-flash-lora:0.55>
inference:
  parameters:
    num_inference_steps: 7
    guidance_scale: 3
    negative_prompt: >-
      (deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, wrong
      anatomy, extra limb, missing limb, floating limbs, (mutated hands and
      fingers:1.4), disconnected limbs, mutation, mutated, ugly, disgusting,
      blurry, amputation

SDXL Flash with LoRA in collaboration with Project Fluently

preview

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)

steps_and_cfg_grid_test

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")