Update app.py

app.py

@@ -1,48 +1,131 @@
-import torch
-from diffusers.utils import load_image
-from diffusers import FluxControlNetPipeline, FluxControlNetModel
 import os
+import torch
+import gradio as gr
+import numpy as np
+from PIL import Image
+from einops import rearrange
+import requests
+import spaces
+from huggingface_hub import login
+from gradio_imageslider import ImageSlider  # Import ImageSlider
+
+from image_datasets.canny_dataset import canny_processor, c_crop
+from src.flux.sampling import denoise_controlnet, get_noise, get_schedule, prepare, unpack
+from src.flux.util import load_ae, load_clip, load_t5, load_flow_model, load_controlnet, load_safetensors
+
+# Download and load the ControlNet model
+model_url = "https://huggingface.co/XLabs-AI/flux-controlnet-canny-v3/resolve/main/flux-canny-controlnet-v3.safetensors?download=true"
+model_path = "./flux-canny-controlnet-v3.safetensors"
+if not os.path.exists(model_path):
+    response = requests.get(model_url)
+    with open(model_path, 'wb') as f:
+        f.write(response.content)
+
+# Source: https://github.com/XLabs-AI/x-flux.git
+name = "flux-dev"
+device = torch.device("cuda")
+offload = False
+is_schnell = name == "flux-schnell"
+
+model, ae, t5, clip, controlnet = None, None, None, None, None
+
+def load_models():
+    global model, ae, t5, clip, controlnet
+    t5 = load_t5(device, max_length=256 if is_schnell else 512)
+    clip = load_clip(device)
+    model = load_flow_model(name, device=device)
+    ae = load_ae(name, device=device)
+    controlnet = load_controlnet(name, device).to(device).to(torch.bfloat16)
+
+    checkpoint = load_safetensors(model_path)
+    controlnet.load_state_dict(checkpoint, strict=False)
+
+load_models()
+
+def preprocess_image(image, target_width, target_height, crop=True):
+    if crop:
+        image = c_crop(image)  # Crop the image to square
+        original_width, original_height = image.size
 
-# Clear unnecessary memory
-torch.cuda.empty_cache()
+        # Resize to match the target size without stretching
+        scale = max(target_width / original_width, target_height / original_height)
+        resized_width = int(scale * original_width)
+        resized_height = int(scale * original_height)
 
-# Set the environment variable to handle memory fragmentation
-os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+        image = image.resize((resized_width, resized_height), Image.LANCZOS)
+        
+        # Center crop to match the target dimensions
+        left = (resized_width - target_width) // 2
+        top = (resized_height - target_height) // 2
+        image = image.crop((left, top, left + target_width, top + target_height))
+    else:
+        image = image.resize((target_width, target_height), Image.LANCZOS)
+    
+    return image
 
-base_model = 'black-forest-labs/FLUX.1-dev'
-controlnet_model = 'InstantX/FLUX.1-dev-Controlnet-Union'
+def preprocess_canny_image(image, target_width, target_height, crop=True):
+    image = preprocess_image(image, target_width, target_height, crop=crop)
+    image = canny_processor(image)
+    return image
 
-# Use mixed precision with float16
-controlnet = FluxControlNetModel.from_pretrained(controlnet_model, torch_dtype=torch.float16)
-pipe = FluxControlNetPipeline.from_pretrained(base_model, controlnet=controlnet, torch_dtype=torch.float16)
+@spaces.GPU(duration=120)
+def generate_image(prompt, control_image, num_steps=50, guidance=4, width=512, height=512, seed=42, random_seed=False):
+    if random_seed:
+        seed = np.random.randint(0, 10000)
+    
+    if not os.path.isdir("./controlnet_results/"):
+        os.makedirs("./controlnet_results/")
 
-# Offload some parts to CPU
-pipe.enable_model_cpu_offload()
+    torch_device = torch.device("cuda")
+    
+    model.to(torch_device)
+    t5.to(torch_device)
+    clip.to(torch_device)
+    ae.to(torch_device)
+    controlnet.to(torch_device)
 
-pipe.to("cuda")
+    width = 16 * width // 16
+    height = 16 * height // 16
+    timesteps = get_schedule(num_steps, (width // 8) * (height // 8) // (16 * 16), shift=(not is_schnell))
+    
+    processed_input = preprocess_image(control_image, width, height)
+    canny_processed = preprocess_canny_image(control_image, width, height)
+    controlnet_cond = torch.from_numpy((np.array(canny_processed) / 127.5) - 1)
+    controlnet_cond = controlnet_cond.permute(2, 0, 1).unsqueeze(0).to(torch.bfloat16).to(torch_device)
 
-control_image_canny = load_image("https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Union-alpha/resolve/main/images/canny.jpg")
-controlnet_conditioning_scale = 0.5
-control_mode = 0
+    torch.manual_seed(seed)
+    with torch.no_grad():
+        x = get_noise(1, height, width, device=torch_device, dtype=torch.bfloat16, seed=seed)
+        inp_cond = prepare(t5=t5, clip=clip, img=x, prompt=prompt)
 
-# Reduce the image resolution if needed to fit into memory
-width, height = control_image_canny.size
-width = width // 2
-height = height // 2
+        x = denoise_controlnet(model, **inp_cond, controlnet=controlnet, timesteps=timesteps, guidance=guidance, controlnet_cond=controlnet_cond)
+        
+        x = unpack(x.float(), height, width)
+        x = ae.decode(x)
 
-prompt = 'A bohemian-style female travel blogger with sun-kissed skin and messy beach waves.'
+    x1 = x.clamp(-1, 1)
+    x1 = rearrange(x1[-1], "c h w -> h w c")
+    output_img = Image.fromarray((127.5 * (x1 + 1.0)).cpu().byte().numpy())
+    
+    return [processed_input, output_img]  # Return both images for slider
 
-image = pipe(
-    prompt, 
-    control_image=control_image_canny,
-    control_mode=control_mode,
-    width=width,
-    height=height,
-    controlnet_conditioning_scale=controlnet_conditioning_scale,
-    num_inference_steps=24, 
-    guidance_scale=3.5,
-).images[0]
-image.save("image.jpg")
+interface = gr.Interface(
+    fn=generate_image,
+    inputs=[
+        gr.Textbox(label="Prompt"),
+        gr.Image(type="pil", label="Control Image"),
+        gr.Slider(step=1, minimum=1, maximum=64, value=28, label="Num Steps"),
+        gr.Slider(minimum=0.1, maximum=10, value=4, label="Guidance"),
+        gr.Slider(minimum=128, maximum=2048, step=128, value=1024, label="Width"),
+        gr.Slider(minimum=128, maximum=2048, step=128, value=1024, label="Height"),
+        gr.Number(value=42, label="Seed"),
+        gr.Checkbox(label="Random Seed")
+    ],
+    outputs=ImageSlider(label="Before / After"),  # Use ImageSlider as the output
+    title="FLUX.1 Controlnet Canny",
+    description="Generate images using ControlNet and a text prompt.\n[[non-commercial license, Flux.1 Dev](https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md)]"
+)
 
-# Empty cache after the operation to free up memory
-torch.cuda.empty_cache()
+if __name__ == "__main__":
+    interface.launch()
+