imgsample-hacked.py

#!/bin/env python

# This is a hacked front end, to use "opendiffusionai/sd-flow-alpha"
# to generate images,
# without having to use the git version of diffusers.


def argproc():
    import argparse
    p = argparse.ArgumentParser()
    p.add_argument("--model", type=str, default="/workspace/sd-flow-alpha")
    p.add_argument("--seed",  type=int, default=10)
    p.add_argument("--steps",  type=int, default=30)
    p.add_argument("--prompt", nargs="+", type=str,
                   default="a blonde woman sitting in a cafe", 
                   help="one or more prompt strings")
    return p.parse_args()

args=argproc()

from diffusers import DiffusionPipeline
import torch.nn as nn, torch, types
import os,sys

from PIL import Image, PngImagePlugin


MODEL = args.model

print("HAND HACKING FLOWMATCH MODULE")
#from diffusers import FlowMatchEulerDiscreteScheduler
#def scale_model_input(self, sample, timestep):
#    return sample
#FlowMatchEulerDiscreteScheduler.scale_model_input = scale_model_input

print(f"Loading from {MODEL}")
if MODEL.endswith(".safetensors") or MODEL.endswith(".st"):
    raise ValueError("Cannot acccept single-file models. "
    "Need diffusers directory tree or hf reference")
else:
    pipe = DiffusionPipeline.from_pretrained(
        MODEL, use_safetensors=True,
        safety_checker=None, requires_safety_checker=False,
        torch_dtype=torch.bfloat16,
    )

pipe.safety_checker=None

print("model initialized. ")
#pipe.enable_sequential_cpu_offload()
# The above obviates the need for to("cuda") I guess..
pipe.to("cuda")

prompt=args.prompt
seed=args.seed

generator = torch.Generator(device="cuda").manual_seed(seed)

print(f"Trying render of '{prompt}' using seed {seed}...")
images = pipe(prompt, num_inference_steps=args.steps, generator=generator).images

OUTDIR=MODEL if os.path.isdir(MODEL) else "./"
for i,image in enumerate(images):
    meta = PngImagePlugin.PngInfo()
    meta.add_text("Comment", f"prompt={prompt}")
    fname=f"{OUTDIR}/sample{i}_s{seed}.png"
    print(f"saving to {fname}")
    image.save(fname, pnginfo=meta)