Update app.py

app.py

@@ -5,19 +5,45 @@ import time
 from os import path
 from safetensors.torch import load_file
 from huggingface_hub import hf_hub_download
+import imageio
+import numpy as np
+import torch
+import rembg
+from PIL import Image
+from torchvision.transforms import v2
+from pytorch_lightning import seed_everything
+from omegaconf import OmegaConf
+from einops import rearrange, repeat
+from tqdm import tqdm
+from diffusers import DiffusionPipeline, EulerAncestralDiscreteScheduler
+import gradio as gr
+import shutil
+import tempfile
+from functools import partial
+from optimum.quanto import quantize, qfloat8, freeze
+from diffusers import FluxPipeline
 
+from src.utils.train_util import instantiate_from_config
+from src.utils.camera_util import (
+    FOV_to_intrinsics, 
+    get_zero123plus_input_cameras,
+    get_circular_camera_poses,
+)
+from src.utils.mesh_util import save_obj, save_glb
+from src.utils.infer_util import remove_background, resize_foreground, images_to_video
+
+# Set up cache path
 cache_path = path.join(path.dirname(path.abspath(__file__)), "models")
 os.environ["TRANSFORMERS_CACHE"] = cache_path
 os.environ["HF_HUB_CACHE"] = cache_path
 os.environ["HF_HOME"] = cache_path
 
-import gradio as gr
-import torch
-from diffusers import FluxPipeline
+huggingface_token = os.getenv("HUGGINGFACE_TOKEN")
 
-torch.backends.cuda.matmul.allow_tf32 = True
+if not path.exists(cache_path):
+    os.makedirs(cache_path, exist_ok=True)
 
-huggingface_token = os.getenv("HUGGINGFACE_TOKEN")
+torch.backends.cuda.matmul.allow_tf32 = True
 
 class timer:
     def __init__(self, method_name="timed process"):
@@ -29,84 +55,213 @@ class timer:
         end = time.time()
         print(f"{self.method} took {str(round(end - self.start, 2))}s")
 
-if not path.exists(cache_path):
-    os.makedirs(cache_path, exist_ok=True)
+def find_cuda():
+    cuda_home = os.environ.get('CUDA_HOME') or os.environ.get('CUDA_PATH')
+    if cuda_home and os.path.exists(cuda_home):
+        return cuda_home
+    nvcc_path = shutil.which('nvcc')
+    if nvcc_path:
+        cuda_path = os.path.dirname(os.path.dirname(nvcc_path))
+        return cuda_path
+    return None
+
+cuda_path = find_cuda()
+if cuda_path:
+    print(f"CUDA installation found at: {cuda_path}")
+else:
+    print("CUDA installation not found")
+
+
+device = torch.device('cuda')
+
+base_model = "black-forest-labs/FLUX.1-dev"
+pipe = FluxPipeline.from_pretrained(base_model, torch_dtype=torch.bfloat16, token=huggingface_token).to(device)
+
+# Load and fuse LoRA BEFORE quantizing
+print('Loading and fusing lora, please wait...')
+lora_path = hf_hub_download("gokaygokay/Flux-Game-Assets-LoRA-v2", "game_asst.safetensors")
+pipe.load_lora_weights(lora_path)
+pipe.fuse_lora(lora_scale=1.0)
+pipe.unload_lora_weights()
+
+
+# Load 3D generation models
+config_path = 'configs/instant-mesh-large.yaml'
+config = OmegaConf.load(config_path)
+config_name = os.path.basename(config_path).replace('.yaml', '')
+model_config = config.model_config
+infer_config = config.infer_config
+
+IS_FLEXICUBES = True if config_name.startswith('instant-mesh') else False
+
+# Load diffusion model for 3D generation
+print('Loading diffusion model ...')
+pipeline = DiffusionPipeline.from_pretrained(
+    "sudo-ai/zero123plus-v1.2", 
+    custom_pipeline="zero123plus",
+    torch_dtype=torch.float16,
+)
+pipeline.scheduler = EulerAncestralDiscreteScheduler.from_config(
+    pipeline.scheduler.config, timestep_spacing='trailing'
+)
 
-pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16, token=huggingface_token)
-pipe.load_lora_weights(hf_hub_download("gokaygokay/Flux-Game-Assets-LoRA-v2", "game_asst.safetensors"))
-pipe.fuse_lora(lora_scale=1)
-pipe.to(device="cuda", dtype=torch.bfloat16)
-pipe.enable_model_cpu_offload()
+# Load custom white-background UNet
+unet_ckpt_path = hf_hub_download(repo_id="TencentARC/InstantMesh", filename="diffusion_pytorch_model.bin", repo_type="model")
+state_dict = torch.load(unet_ckpt_path, map_location='cpu')
+pipeline.unet.load_state_dict(state_dict, strict=True)
+
+pipeline = pipeline.to(device)
+
+# Load reconstruction model
+print('Loading reconstruction model ...')
+model_ckpt_path = hf_hub_download(repo_id="TencentARC/InstantMesh", filename="instant_mesh_large.ckpt", repo_type="model")
+model = instantiate_from_config(model_config)
+state_dict = torch.load(model_ckpt_path, map_location='cpu')['state_dict']
+state_dict = {k[14:]: v for k, v in state_dict.items() if k.startswith('lrm_generator.') and 'source_camera' not in k}
+model.load_state_dict(state_dict, strict=True)
+
+model = model.to(device)
+
+print('Loading Finished!')
+
+def get_render_cameras(batch_size=1, M=120, radius=2.5, elevation=10.0, is_flexicubes=False):
+    c2ws = get_circular_camera_poses(M=M, radius=radius, elevation=elevation)
+    if is_flexicubes:
+        cameras = torch.linalg.inv(c2ws)
+        cameras = cameras.unsqueeze(0).repeat(batch_size, 1, 1, 1)
+    else:
+        extrinsics = c2ws.flatten(-2)
+        intrinsics = FOV_to_intrinsics(50.0).unsqueeze(0).repeat(M, 1, 1).float().flatten(-2)
+        cameras = torch.cat([extrinsics, intrinsics], dim=-1)
+        cameras = cameras.unsqueeze(0).repeat(batch_size, 1, 1)
+    return cameras
+
+def preprocess(input_image, do_remove_background):
+    rembg_session = rembg.new_session() if do_remove_background else None
+    if do_remove_background:
+        input_image = remove_background(input_image, rembg_session)
+        input_image = resize_foreground(input_image, 0.85)
+    return input_image
+
+ts_cutoff = 2
+
+@spaces.GPU
+def generate_flux_image(prompt, height, width, steps, scales, seed):
+    return pipe(
+        prompt=prompt, 
+        width=int(height),
+        height=int(width),
+        num_inference_steps=int(steps), 
+        generator=torch.Generator().manual_seed(int(seed)),
+        guidance_scale=float(scales),
+        timestep_to_start_cfg=ts_cutoff,
+    ).images[0]
+
+
+@spaces.GPU
+def generate_mvs(input_image, sample_steps, sample_seed):
+    seed_everything(sample_seed)
+    z123_image = pipeline(
+        input_image, 
+        num_inference_steps=sample_steps
+    ).images[0]
+    show_image = np.asarray(z123_image, dtype=np.uint8)
+    show_image = torch.from_numpy(show_image)
+    show_image = rearrange(show_image, '(n h) (m w) c -> (n m) h w c', n=3, m=2)
+    show_image = rearrange(show_image, '(n m) h w c -> (n h) (m w) c', n=2, m=3)
+    show_image = Image.fromarray(show_image.numpy())
+    return z123_image, show_image
+
+@spaces.GPU
+def make3d(images):
+    global model
+    if IS_FLEXICUBES:
+        model.init_flexicubes_geometry(device, use_renderer=False)
+    model = model.eval()
+
+    images = np.asarray(images, dtype=np.float32) / 255.0
+    images = torch.from_numpy(images).permute(2, 0, 1).contiguous().float()
+    images = rearrange(images, 'c (n h) (m w) -> (n m) c h w', n=3, m=2)
+
+    input_cameras = get_zero123plus_input_cameras(batch_size=1, radius=4.0).to(device)
+    render_cameras = get_render_cameras(batch_size=1, radius=2.5, is_flexicubes=IS_FLEXICUBES).to(device)
+
+    images = images.unsqueeze(0).to(device)
+    images = v2.functional.resize(images, (320, 320), interpolation=3, antialias=True).clamp(0, 1)
+
+    mesh_fpath = tempfile.NamedTemporaryFile(suffix=f".obj", delete=False).name
+    mesh_basename = os.path.basename(mesh_fpath).split('.')[0]
+    mesh_dirname = os.path.dirname(mesh_fpath)
+    mesh_glb_fpath = os.path.join(mesh_dirname, f"{mesh_basename}.glb")
+
+    with torch.no_grad():
+        planes = model.forward_planes(images, input_cameras)
+        mesh_out = model.extract_mesh(
+            planes,
+            use_texture_map=False,
+            **infer_config,
+        )
+        vertices, faces, vertex_colors = mesh_out
+        vertices = vertices[:, [1, 2, 0]]
+        save_glb(vertices, faces, vertex_colors, mesh_glb_fpath)
+        save_obj(vertices, faces, vertex_colors, mesh_fpath)
+    
+    return mesh_fpath, mesh_glb_fpath
 
 with gr.Blocks(theme=gr.themes.Soft()) as demo:
     gr.Markdown(
         """
         <div style="text-align: center; max-width: 650px; margin: 0 auto;">
-            <h1 style="font-size: 2.5rem; font-weight: 700; margin-bottom: 1rem; display: contents;">Hyper-FLUX-8steps-LoRA</h1>
-            <p style="font-size: 1rem; margin-bottom: 1.5rem;">AutoML team from ByteDance</p>
+            <h1 style="font-size: 2.5rem; font-weight: 700; margin-bottom: 1rem;">Flux Image to 3D Model Generator</h1>
         </div>
         """
     )
 
     with gr.Row():
         with gr.Column(scale=3):
-            with gr.Group():
-                prompt = gr.Textbox(
-                    label="Your Image Description",
-                    placeholder="E.g., A serene landscape with mountains and a lake at sunset",
-                    lines=3
-                )
-                
-                with gr.Accordion("Advanced Settings", open=False):
-                    with gr.Group():
-                        with gr.Row():
-                            height = gr.Slider(label="Height", minimum=256, maximum=1152, step=64, value=1024)
-                            width = gr.Slider(label="Width", minimum=256, maximum=1152, step=64, value=1024)
-                        
-                        with gr.Row():
-                            steps = gr.Slider(label="Inference Steps", minimum=10, maximum=50, step=1, value=28)
-                            scales = gr.Slider(label="Guidance Scale", minimum=0.0, maximum=5.0, step=0.1, value=3.5)
-                        
-                        seed = gr.Number(label="Seed (for reproducibility)", value=3413, precision=0)
-                
-                generate_btn = gr.Button("Generate Image", variant="primary", scale=1)
+            prompt = gr.Textbox(
+                label="Your Image Description",
+                placeholder="E.g., A serene landscape with mountains and a lake at sunset",
+                lines=3
+            )
+            
+            with gr.Accordion("Advanced Settings", open=False):
+                with gr.Group():
+                    with gr.Row():
+                        height = gr.Slider(label="Height", minimum=256, maximum=1152, step=64, value=1024)
+                        width = gr.Slider(label="Width", minimum=256, maximum=1152, step=64, value=1024)
+                    
+                    with gr.Row():
+                        steps = gr.Slider(label="Inference Steps", minimum=10, maximum=50, step=1, value=28)
+                        scales = gr.Slider(label="Guidance Scale", minimum=0.0, maximum=5.0, step=0.1, value=3.5)
+                    
+                    seed = gr.Number(label="Seed (for reproducibility)", value=3413, precision=0)
+            
+            generate_btn = gr.Button("Generate 3D Model", variant="primary")
 
         with gr.Column(scale=4):
-            output = gr.Image(label="Your Generated Image")
-    
-    gr.Markdown(
-        """
-        <div style="max-width: 650px; margin: 2rem auto; padding: 1rem; border-radius: 10px; background-color: #f0f0f0;">
-            <h2 style="font-size: 1.5rem; margin-bottom: 1rem;">How to Use</h2>
-            <ol style="padding-left: 1.5rem;">
-                <li>Enter a detailed description of the image you want to create.</li>
-                <li>Adjust advanced settings if desired (tap to expand).</li>
-                <li>Tap "Generate Image" and wait for your creation!</li>
-            </ol>
-            <p style="margin-top: 1rem; font-style: italic;">Tip: Be specific in your description for best results!</p>
-        </div>
-        """
-    )
+            flux_output = gr.Image(label="Generated Flux Image")
+            mv_show_images = gr.Image(label="Generated Multi-views")
+            with gr.Row():
+                with gr.Tab("OBJ"):
+                    output_model_obj = gr.Model3D(label="Output Model (OBJ Format)")
+                with gr.Tab("GLB"):
+                    output_model_glb = gr.Model3D(label="Output Model (GLB Format)")
+
+    mv_images = gr.State()
 
-    @spaces.GPU
-    def process_image(height, width, steps, scales, prompt, seed):
-        global pipe
-        with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16), timer("inference"):
-            return pipe(
-                prompt=[prompt],
-                generator=torch.Generator().manual_seed(int(seed)),
-                num_inference_steps=int(steps),
-                guidance_scale=float(scales),
-                height=int(height),
-                width=int(width),
-                max_sequence_length=256
-            ).images[0]
+    def process_pipeline(prompt, height, width, steps, scales, seed):
+        flux_image = generate_flux_image(prompt, height, width, steps, scales, seed)
+        processed_image = preprocess(flux_image, do_remove_background=True)
+        mv_images, show_image = generate_mvs(processed_image, steps, seed)
+        obj_path, glb_path = make3d(mv_images)
+        return flux_image, show_image, obj_path, glb_path
 
     generate_btn.click(
-        process_image,
-        inputs=[height, width, steps, scales, prompt, seed],
-        outputs=output
+        fn=process_pipeline,
+        inputs=[prompt, height, width, steps, scales, seed],
+        outputs=[flux_output, mv_show_images, output_model_obj, output_model_glb]
     )
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch()