Reproducible generation with xFormers for SD2.x or variant models

readme.md

Reproducible generation with xFormers for SD2.x or variant models

More details and background: Generating (almost) reproducible pictures using Diffusers with xFormers

Limitation

Unfortunately, Flash Attention won't accept SD1.x attention shapes, but will accept SD2.x or variants.

Test Result

Automatic1111 WebUI (API) result:

$  python launch.py --xformers --listen --api

$  python launch.py --xformers --xformers-flash-attention --listen --api

Performance

without --xformers-flash-attention option:

100%|█████████████████████████████████████████████████████████████████| 20/20 [00:05<00:00,  3.89it/s]
100%|█████████████████████████████████████████████████████████████████| 20/20 [00:04<00:00,  4.30it/s]
100%|█████████████████████████████████████████████████████████████████| 20/20 [00:04<00:00,  4.23it/s]
100%|█████████████████████████████████████████████████████████████████| 20/20 [00:04<00:00,  4.30it/s]
Total progress: 80it [00:19,  4.18it/s]

with --xformers-flash-attention option:

100%|█████████████████████████████████████████████████████████████████| 20/20 [00:05<00:00,  3.91it/s]
100%|█████████████████████████████████████████████████████████████████| 20/20 [00:04<00:00,  4.35it/s]
100%|█████████████████████████████████████████████████████████████████| 20/20 [00:04<00:00,  4.26it/s]
100%|█████████████████████████████████████████████████████████████████| 20/20 [00:04<00:00,  4.33it/s]
Total progress: 80it [00:19,  4.22it/s]

test_t2i_repeat_same_setting.py

import json
import requests
import io
import base64
from PIL import Image
import numpy as np
import matplotlib.pyplot as plt
import sys
from pprint import pprint

plt.rcParams["figure.figsize"] = (8, 6)
plt.rcParams['figure.facecolor'] = 'white'

url = "http://127.0.0.1:7860"

# prompt by @p1atdev_art
# https://twitter.com/p1atdev_art/status/1616557167087353856
prompt = """masterpiece, best quality, high quality, 1girl, sun hat, 
frilled white dress, looking at viewer, summer, sky, beach, semi-realistic"""
neg_prompt = """nsfw, worst quality, low quality, medium quality, deleted, 
lowres, bad anatomy, bad hands, text, error, missing fingers, extra digits, 
fewer digits, cropped, jpeg artifacts, signature, watermark, username, blurry,"""
t2i_payload = {
    "prompt": prompt,
    "negative_prompt": neg_prompt,
    "seed": 0,
    "batch_size": 1,
    "n_iter": 1,
    "steps": 20,
    "cfg_scale": 11,
    "width": 512,
    "height": 512,
    "sampler_index": "DPM++ 2S a Karras",
    "model": "plat-v1-3-1-fp16", # https://huggingface.co/p1atdev/pd-archive/tree/main
}

def t2i_repeate(fn_prefix='run', seed=0, runs=4):
    t2i_payload['seed'] = seed
    infotext = ''
    for run in range(runs):
        response = requests.post(url=f'{url}/sdapi/v1/txt2img', json=t2i_payload)
        r = response.json()
        infotext = json.loads(r['info'])['infotexts'][0]
        i = r['images'][0]
        image = Image.open(io.BytesIO(base64.b64decode(i.split(",",1)[0])))
        fn = f"{fn_prefix}_{run}.png"
        image.save(fn)
        print(f'{fn} generated.')
    return infotext

def calc_difference_image(image0, image1, normalize=False):
    image0 = np.array(image0, dtype=np.int32)
    image1 = np.array(image1, dtype=np.int32)
    abs_diff = np.abs(image0 - image1)
    if normalize:
        abs_diff = abs_diff / abs_diff.max() * 255
    return Image.fromarray(abs_diff.astype(np.uint8))

def render_figure(title, footer, fn_prefix, fig_fn, runs=4):
    def plot_normal_row(axs, fn_prefix):
        for i, ax in enumerate(axs):
            ax.set_title(f'Run#{i}')
            ax.imshow(Image.open(f'{fn_prefix}_{i}.png'))
    def plot_diff_row(axs, fn_prefix):
        for i, ax in enumerate(axs):
            ax.set_title(f'Run#{i} - Run#0')
            ref = Image.open(f'{fn_prefix}_{0}.png')
            res = Image.open(f'{fn_prefix}_{i}.png')
            ax.imshow(calc_difference_image(ref, res))

    fig, axs = plt.subplots(2, runs)
    for ax in axs.flatten():
        ax.set_aspect('equal', 'box')
        ax.axis('off')
    plot_normal_row(axs[0], fn_prefix)
    plot_diff_row(axs[1], fn_prefix)
    fig.text(.02, .05, footer, wrap=True, horizontalalignment='left', fontsize=8)
    fig.suptitle(title, fontsize=16)
    fig.tight_layout(rect=[0,0.11,1,1])
    fig.savefig(fig_fn, dpi=72*2)

if __name__ == '__main__':
    title = sys.argv[1]
    fn_prefix='run'
    runs = 4
    footer = t2i_repeate(fn_prefix=fn_prefix, seed=7, runs=runs)
    fig_fn = f"fig_{title.replace(' ', '_').replace('-', '_')}.png"
    render_figure(title=title, footer=footer, fn_prefix=fn_prefix, fig_fn=fig_fn, runs=runs)