support for SDXL

[wzq728]

Thanks for your nice work! I want to know if tome support SDXL? And if it is, how to use it.

[dbolya]

I haven't looked into it. How does SDXL differ from normal SD? If it's similar, there's probably a way to get it to work.

[theAdamColton]

I haven't done any detailed tests, but wrapping a huggingface pipeline.unet seems to work without crashing for training and inference, and produces images that are ok

this is with r=0.5 at 672x672

[dbolya]

Does it speed it up? I think the default behavior of the diffusers implementation is to do nothing when wrapping the wrong thing, so it might not actually be doing anything.

[theAdamColton]

import tomesd
from diffusers import StableDiffusionXLPipeline, StableDiffusionPipeline
import torch
import time

pipeline = StableDiffusionXLPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16, variant="fp16").to("cuda")

batch_size = 4
resolution = 896
trials = 2

tt = 0
for _ in range(trials):
    st = time.time()
    pipeline(prompt="Laundromat Stories: Inside a laundromat on a rainy day. People load clothes into washing machines and read magazines while waiting. Charcoal drawing, chiaroscuro, dramatic
 lighting from overhead fluorescents.", num_inference_steps=20, num_images_per_prompt=batch_size, width = resolution, height=resolution)
    tt += time.time() - st
print("SDXL no tomesd: avg time", tt/trials)

pipeline = tomesd.apply_patch(pipeline, ratio=0.75, max_downsample = 4)

tt = 0
for _ in range(trials):
    st = time.time()
    pipeline(prompt="Laundromat Stories: Inside a laundromat on a rainy day. People load clothes into washing machines and read magazines while waiting. Charcoal drawing, chiaroscuro, dramatic
 lighting from overhead fluorescents.", num_inference_steps=20, num_images_per_prompt=batch_size, width = resolution, height=resolution)
    tt += time.time() - st
print("SDXL w/ tomesd: avg time", tt/trials)

I get around a 12% speedup on a 3090: 18.9267s vs 16.891s

[haoweiz23]

Hi @theAdamColton,

I found SDXL does not achieve any acceleration when 'max_downsample = 1' , do you have any advise about such results? Thanks!