ComfyUI comes up with very different images than Diffusers

[komninoschatzipapas]

I'm using this Python script:

# Use Pytorch 2!
import torch
from diffusers import (
    StableDiffusionPipeline,
    DiffusionPipeline,
    AutoencoderKL,
    UNet2DConditionModel,
    DDPMScheduler,
)
from transformers import CLIPTextModel

# Any model currently on Huggingface Hub.
model_id = "stabilityai/stable-diffusion-xl-base-1.0"
pipeline = DiffusionPipeline.from_pretrained(model_id)

# Optimize!
pipeline.unet = torch.compile(pipeline.unet)
scheduler = DDPMScheduler.from_pretrained(model_id, subfolder="scheduler")

# Remove this if you get an error.
torch.set_float32_matmul_precision("high")

pipeline.to("cuda")
prompts = {
    "car": "car",
}
for shortname, prompt in prompts.items():
    # old prompt: ''
    image = pipeline(
        prompt=prompt,
        negative_prompt="",
        num_inference_steps=35,
        generator=torch.Generator(device="cuda").manual_seed(42),
        width=832,
        height=1216,
        guidance_scale=5,
    ).images[0]
    image.save(f"./{shortname}_nobetas.png", format="PNG")

Which yields this image:

Trying to replicate this in ComfyUI (719fb2c):

I'm not sure what I'm missing here but although both setups seem to be using the same model and parameters, the results are quite different.

[RandomGitUser321]

Two potential things it could be:

# Remove this if you get an error.
torch.set_float32_matmul_precision("high")

and

generator=torch.Generator(device="cuda").manual_seed(42),

ComfyUI is likely generating the seed on CPU, vs that code performing it on the GPU (in the comfy samplers dropdown, any sampler that says GPU in it is generating the seed on the GPU instead of the CPU). A seed of 42 on a GPU isn't the same as it is for a CPU. Also, when performed on GPU, they can be driver or architecture specific. That's why most people recommend using CPU based seeding for reproducibility.

[komninoschatzipapas]

@RandomGitUser321, thank you! Generating the seed on the CPU matched the image I was getting from ComfyUI.