Spaces:

nftnik
/

Redux

Running on Zero

File size: 9,959 Bytes

import os
import sys
import random
import torch
from pathlib import Path
from PIL import Image
import gradio as gr
from huggingface_hub import hf_hub_download
import spaces
from typing import Union, Sequence, Mapping, Any
import folder_paths
from nodes import NODE_CLASS_MAPPINGS, init_extra_nodes
from comfy import model_management

# Configuração de diretórios
BASE_DIR = os.path.dirname(os.path.realpath(__file__))
output_dir = os.path.join(BASE_DIR, "output")
models_dir = os.path.join(BASE_DIR, "models")
os.makedirs(output_dir, exist_ok=True)
os.makedirs(models_dir, exist_ok=True)

# Configurar caminhos dos modelos
for model_folder in ["style_models", "text_encoders", "vae", "unet", "clip_vision"]:
    folder_path = os.path.join(models_dir, model_folder)
    os.makedirs(folder_path, exist_ok=True)
    folder_paths.add_model_folder_path(model_folder, folder_path)

# Download dos modelos
print("Baixando modelos necessários...")
hf_hub_download(repo_id="black-forest-labs/FLUX.1-Redux-dev", 
                filename="flux1-redux-dev.safetensors", 
                local_dir=os.path.join(models_dir, "style_models"))
hf_hub_download(repo_id="comfyanonymous/flux_text_encoders", 
                filename="t5xxl_fp16.safetensors", 
                local_dir=os.path.join(models_dir, "text_encoders"))
hf_hub_download(repo_id="zer0int/CLIP-GmP-ViT-L-14", 
                filename="ViT-L-14-TEXT-detail-improved-hiT-GmP-TE-only-HF.safetensors", 
                local_dir=os.path.join(models_dir, "text_encoders"))
hf_hub_download(repo_id="black-forest-labs/FLUX.1-dev", 
                filename="ae.safetensors", 
                local_dir=os.path.join(models_dir, "vae"))
hf_hub_download(repo_id="black-forest-labs/FLUX.1-dev", 
                filename="flux1-dev.safetensors", 
                local_dir=os.path.join(models_dir, "unet"))
hf_hub_download(repo_id="google/siglip-so400m-patch14-384", 
                filename="model.safetensors", 
                local_dir=os.path.join(models_dir, "clip_vision"))

# Diagnóstico CUDA
print("Python version:", sys.version)
print("Torch version:", torch.__version__)
print("CUDA disponível:", torch.cuda.is_available())
print("Quantidade de GPUs:", torch.cuda.device_count())
if torch.cuda.is_available():
    print("GPU atual:", torch.cuda.get_device_name(0))

# Inicializar nós extras
print("Inicializando ComfyUI...")
init_extra_nodes()

# Helper function
def get_value_at_index(obj: Union[Sequence, Mapping], index: int) -> Any:
    try:
        return obj[index]
    except KeyError:
        return obj["result"][index]

# Inicializar modelos
print("Inicializando modelos...")
with torch.inference_mode():
    # CLIP
    dualcliploader = NODE_CLASS_MAPPINGS["DualCLIPLoader"]()
    CLIP_MODEL = dualcliploader.load_clip(
        clip_name1="t5xxl_fp16.safetensors",
        clip_name2="ViT-L-14-TEXT-detail-improved-hiT-GmP-TE-only-HF.safetensors",
        type="flux"
    )

    # Style Model
    stylemodelloader = NODE_CLASS_MAPPINGS["StyleModelLoader"]()
    STYLE_MODEL = stylemodelloader.load_style_model(
        style_model_name="flux1-redux-dev.safetensors"
    )

    # VAE
    vaeloader = NODE_CLASS_MAPPINGS["VAELoader"]()
    VAE_MODEL = vaeloader.load_vae(
        vae_name="ae.safetensors"
    )

    # UNET
    unetloader = NODE_CLASS_MAPPINGS["UNETLoader"]()
    UNET_MODEL = unetloader.load_unet(
        unet_name="flux1-dev.safetensors",
        weight_dtype="fp8_e4m3fn"
    )

    # CLIP Vision
    clipvisionloader = NODE_CLASS_MAPPINGS["CLIPVisionLoader"]()
    CLIP_VISION = clipvisionloader.load_clip(
        clip_name="sigclip_vision_patch14_384.safetensors"
    )

model_loaders = [CLIP_MODEL, VAE_MODEL, UNET_MODEL, CLIP_VISION]
model_management.load_models_gpu([
    loader[0].patcher if hasattr(loader[0], 'patcher') else loader[0] 
    for loader in model_loaders
])

@spaces.GPU
def generate_image(prompt, input_image, lora_weight, guidance, downsampling_factor, weight, seed, width, height, batch_size, steps, progress=gr.Progress(track_tqdm=True)):
    try:
        with torch.inference_mode():
            # Text Encoding
            cliptextencode = NODE_CLASS_MAPPINGS["CLIPTextEncode"]()
            encoded_text = cliptextencode.encode(
                text=prompt,
                clip=CLIP_MODEL[0]
            )

            # Load Input Image
            loadimage = NODE_CLASS_MAPPINGS["LoadImage"]()
            loaded_image = loadimage.load_image(image=input_image)

            # Load LoRA
            loraloadermodelonly = NODE_CLASS_MAPPINGS["LoraLoaderModelOnly"]()
            lora_model = loraloadermodelonly.load_lora_model_only(
                lora_name="NFTNIK_FLUX.1[dev]_LoRA.safetensors",
                strength_model=lora_weight,
                model=UNET_MODEL[0]
            )

            # Flux Guidance
            fluxguidance = NODE_CLASS_MAPPINGS["FluxGuidance"]()
            flux_guidance = fluxguidance.append(
                guidance=guidance,
                conditioning=encoded_text[0]
            )

            # Redux Advanced
            reduxadvanced = NODE_CLASS_MAPPINGS["ReduxAdvanced"]()
            redux_result = reduxadvanced.apply_stylemodel(
                downsampling_factor=downsampling_factor,
                downsampling_function="area",
                mode="keep aspect ratio",
                weight=weight,
                autocrop_margin=0.1,
                conditioning=flux_guidance[0],
                style_model=STYLE_MODEL[0],
                clip_vision=CLIP_VISION[0],
                image=loaded_image[0]
            )

            # Empty Latent Image
            emptylatentimage = NODE_CLASS_MAPPINGS["EmptyLatentImage"]()
            empty_latent = emptylatentimage.generate(
                width=width,
                height=height,
                batch_size=batch_size
            )

            # KSampler
            ksampler = NODE_CLASS_MAPPINGS["KSampler"]()
            sampled = ksampler.sample(
                seed=seed,
                steps=steps,
                cfg=1,
                sampler_name="euler",
                scheduler="simple",
                denoise=1,
                model=lora_model[0],
                positive=redux_result[0],
                negative=flux_guidance[0],
                latent_image=empty_latent[0]
            )

            # VAE Decode
            vaedecode = NODE_CLASS_MAPPINGS["VAEDecode"]()
            decoded = vaedecode.decode(
                samples=sampled[0],
                vae=VAE_MODEL[0]
            )

            # Salvar imagem
            temp_filename = f"Flux_{random.randint(0, 99999)}.png"
            temp_path = os.path.join(output_dir, temp_filename)
            Image.fromarray((decoded[0] * 255).astype("uint8")).save(temp_path)

            return temp_path

    except Exception as e:
        print(f"Erro ao gerar imagem: {str(e)}")
        return None

# Interface Gradio
with gr.Blocks() as app:
    gr.Markdown("# FLUX Redux Image Generator")
    
    with gr.Row():
        with gr.Column():
            prompt_input = gr.Textbox(
                label="Prompt",
                placeholder="Enter your prompt here...",
                lines=5
            )
            input_image = gr.Image(
                label="Input Image",
                type="filepath"
            )
            
            with gr.Row():
                with gr.Column():
                    lora_weight = gr.Slider(
                        minimum=0,
                        maximum=2,
                        step=0.1,
                        value=0.6,
                        label="LoRA Weight"
                    )
                    guidance = gr.Slider(
                        minimum=0,
                        maximum=20,
                        step=0.1,
                        value=3.5,
                        label="Guidance"
                    )
                    downsampling_factor = gr.Slider(
                        minimum=1,
                        maximum=8,
                        step=1,
                        value=3,
                        label="Downsampling Factor"
                    )
                    weight = gr.Slider(
                        minimum=0,
                        maximum=2,
                        step=0.1,
                        value=1.0,
                        label="Model Weight"
                    )
                with gr.Column():
                    seed = gr.Number(
                        value=random.randint(1, 2**64),
                        label="Seed",
                        precision=0
                    )
                    width = gr.Number(
                        value=1024,
                        label="Width",
                        precision=0
                    )
                    height = gr.Number(
                        value=1024,
                        label="Height",
                        precision=0
                    )
                    batch_size = gr.Number(
                        value=1,
                        label="Batch Size",
                        precision=0
                    )
                    steps = gr.Number(
                        value=20,
                        label="Steps",
                        precision=0
                    )
            
            generate_btn = gr.Button("Generate Image")
        
        with gr.Column():
            output_image = gr.Image(label="Generated Image", type="filepath")
    
    generate_btn.click(
        fn=generate_image,
        inputs=[
            prompt_input,
            input_image,
            lora_weight,
            guidance,
            downsampling_factor,
            weight,
            seed,
            width,
            height,
            batch_size,
            steps
        ],
        outputs=[output_image]
    )

if __name__ == "__main__":
    app.launch()