app.py

import os
import sys
import gradio as gr
import torch
import random
import numpy as np
from PIL import Image

# Setup and model loading
os.chdir('/content')
!git clone -b totoro2 https://github.com/camenduru/ComfyUI /content/TotoroUI
os.chdir('/content/TotoroUI')

# Create requirements.txt if it doesn't exist
requirements_content = """torch
torchsde
einops
diffusers
accelerate
xformers==0.0.26.post1
gradio"""

with open("requirements.txt", "w") as f:
    f.write(requirements_content)

# Install dependencies from requirements.txt
!pip install -r requirements.txt

# Install aria2
!apt -y install -qq aria2

# Download model weights
!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/adamo1139/stable-diffusion-3-medium-ungated/resolve/main/sd3_medium_incl_clips_t5xxlfp8.safetensors -d /content/TotoroUI/model -o sd3_medium_incl_clips_t5xxlfp8.safetensors

# Add TotoroUI to sys.path
sys.path.append('/content/TotoroUI')

# Import after adding to sys.path
import node_helpers
from totoro.sd import load_checkpoint_guess_config
import nodes

# Check for GPU availability and CUDA
use_cuda = torch.cuda.is_available()

model_patcher, clip, vae, clipvision = load_checkpoint_guess_config(
    "/content/TotoroUI/model/sd3_medium_incl_clips_t5xxlfp8.safetensors", 
    output_vae=True, output_clip=True, embedding_directory=None
)

def zero_out(conditioning):
    c = []
    for t in conditioning:
        d = t[1].copy()
        if "pooled_output" in d:
            d["pooled_output"] = torch.zeros_like(d["pooled_output"])
        n = [torch.zeros_like(t[0]), d]
        c.append(n)
    return (c, )

def generate_image(prompt, negative_prompt, steps):
    with torch.inference_mode():
        latent = {"samples": torch.ones([1, 16, 1024 // 8, 1024 // 8]) * 0.0609}
        
        cond, pooled = clip.encode_from_tokens(clip.tokenize(prompt), return_pooled=True)
        cond = [[cond, {"pooled_output": pooled}]]
        
        n_cond, n_pooled = clip.encode_from_tokens(clip.tokenize(negative_prompt), return_pooled=True)
        n_cond = [[n_cond, {"pooled_output": n_pooled}]]

        n_cond1 = node_helpers.conditioning_set_values(n_cond, {"start_percent": 0, "end_percent": 0.1})
        n_cond2 = zero_out(n_cond)
        n_cond2 = node_helpers.conditioning_set_values(n_cond2[0], {"start_percent": 0.1, "end_percent": 1.0})
        n_cond = n_cond1 + n_cond2

        seed = random.randint(0, 18446744073709551615)
        
        sample = nodes.common_ksampler(
            model=model_patcher, 
            seed=seed, 
            steps=steps, 
            cfg=4.5, 
            sampler_name="dpmpp_2m", 
            scheduler="sgm_uniform", 
            positive=cond, 
            negative=n_cond,
            latent=latent, 
            denoise=1
        )
        
        sample = sample[0]["samples"].to(torch.float16)
        
        if use_cuda:
            vae.first_stage_model.cuda()
        decoded = vae.decode_tiled(sample).detach()

    return Image.fromarray(np.array(decoded*255, dtype=np.uint8)[0])

# Gradio interface
interface = gr.Interface(
    fn=generate_image,
    inputs=[
        gr.Textbox(label="Prompt"),
        gr.Textbox(label="Negative Prompt"),
        gr.Slider(label="Steps", minimum=1, maximum=200, step=1, default=28)
    ],
    outputs=gr.Image(label="Generated Image")
)

interface.launch()