File size: 5,023 Bytes
2213223 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 | #!/usr/bin/env python3
"""End-to-end Z-Image-Turbo txt2img comparison across diffusion model variants.
Runs the standard txt2img flow (CLIP encode -> KSampler -> VAE decode) once per
given diffusion checkpoint, using the same prompt/seed/size, and reports load,
sample and decode timings per variant. Outputs PNGs + a timing log.
Usage:
python test_generate.py --te qwen_3_4b.safetensors --vae qwen_image_vae.safetensors \
--diffusion bf16.safetensors --diffusion w4a8.safetensors --diffusion int8_convrot.safetensors \
--prompt "..." --outdir out
"""
import argparse
import json
import os
import sys
import time
import torch
REPO = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "ComfyUI")
if REPO not in sys.path:
sys.path.insert(0, REPO)
import comfy.sample
import comfy.sd
import comfy.utils
def load_clip_cond(te_path, prompt, negative):
clip = comfy.sd.load_clip([te_path], clip_type=comfy.sd.CLIPType.QWEN_IMAGE)
positive = clip.encode_from_tokens_scheduled(clip.tokenize(prompt))
negative = clip.encode_from_tokens_scheduled(clip.tokenize(negative))
return clip, positive, negative
def save_png(tensor, path):
img = torch.clamp(tensor, 0.0, 1.0).cpu().numpy()
img = (img * 255.0).astype("uint8")
try:
import torchvision.transforms.functional as F
F.to_pil_image(torch.from_numpy(img).permute(2, 0, 1)).save(path)
except ImportError:
from PIL import Image
Image.fromarray(img).save(path)
def generate(args, label, dm_path, positive, negative, vae):
t0 = time.time()
patcher = comfy.sd.load_diffusion_model(dm_path)
t_load = time.time() - t0
latent_format = patcher.get_model_object("latent_format")
batch = 1
latent = torch.zeros([batch, latent_format.latent_channels,
args.height // 8, args.width // 8], dtype=torch.float32)
noise = comfy.sample.prepare_noise(latent, args.seed)
t0 = time.time()
samples = comfy.sample.sample(
patcher, noise, args.steps, args.cfg, args.sampler, args.scheduler,
positive, negative, latent, denoise=1.0,
disable_pbar=not args.pbar, seed=args.seed)
t_sample = time.time() - t0
t0 = time.time()
images = vae.decode(samples)
t_decode = time.time() - t0
out_path = os.path.join(args.outdir, f"{label}.png")
save_png(images[0], out_path)
peak = 0
if torch.cuda.is_available():
peak = torch.cuda.max_memory_allocated() / 2**30
return {
"label": label,
"output": out_path,
"load_s": round(t_load, 2),
"sample_s": round(t_sample, 2),
"decode_s": round(t_decode, 2),
"total_s": round(t_load + t_sample + t_decode, 2),
"peak_vram_gb": round(peak, 2),
}
def main():
ap = argparse.ArgumentParser(description=__doc__)
ap.add_argument("--te", required=True, help="text encoder safetensors (BF16)")
ap.add_argument("--vae", required=True, help="VAE safetensors")
ap.add_argument("--diffusion", action="append", required=True, help="diffusion model path (repeatable)")
ap.add_argument("--prompt", default="A cute corgi sitting on a mossy rock in a forest, soft sunlight, detailed fur, photographic")
ap.add_argument("--negative", default="")
ap.add_argument("--steps", type=int, default=8)
ap.add_argument("--cfg", type=float, default=1.0)
ap.add_argument("--sampler", default="euler")
ap.add_argument("--scheduler", default="beta")
ap.add_argument("--width", type=int, default=1024)
ap.add_argument("--height", type=int, default=1024)
ap.add_argument("--seed", type=int, default=42)
ap.add_argument("--outdir", default="zimage_out")
ap.add_argument("--pbar", action="store_true")
args = ap.parse_args()
os.makedirs(args.outdir, exist_ok=True)
clip, positive, negative = load_clip_cond(args.te, args.prompt, args.negative)
sd, _ = comfy.utils.load_torch_file(args.vae, return_metadata=True)
vae = comfy.sd.VAE(sd=sd)
vae.throw_exception_if_invalid()
results = []
for i, dm in enumerate(args.diffusion):
label = os.path.splitext(os.path.basename(dm))[0]
print(f"\n=== [{i + 1}/{len(args.diffusion)}] {label} ===", flush=True)
try:
r = generate(args, label, dm, positive, negative, vae)
results.append(r)
print(json.dumps(r, ensure_ascii=False, indent=2), flush=True)
except Exception as e:
import traceback
traceback.print_exc()
results.append({"label": label, "error": str(e)})
if torch.cuda.is_available():
torch.cuda.empty_cache()
log_path = os.path.join(args.outdir, "timings.json")
with open(log_path, "w") as f:
json.dump({"settings": vars(args), "results": results}, f, ensure_ascii=False, indent=2)
print(f"\nWrote {log_path}")
for r in results:
print(json.dumps(r, ensure_ascii=False))
if __name__ == "__main__":
main()
|