Update app.py

app.py

@@ -1,168 +1,268 @@
 import spaces
-import gradio as gr
+import argparse
+import os
+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 transformers import AutoProcessor, AutoModelForCausalLM, pipeline
-from diffusers import DiffusionPipeline
-import random
-import numpy as np
-import os
-import subprocess
-subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
+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
 
-# Initialize models
-device = "cuda" if torch.cuda.is_available() else "cpu"
-dtype = torch.bfloat16
+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
 
 huggingface_token = os.getenv("HUGGINGFACE_TOKEN")
 
-from diffusers import FluxPipeline, FluxTransformer2DModel
+if not path.exists(cache_path):
+    os.makedirs(cache_path, exist_ok=True)
+
+torch.backends.cuda.matmul.allow_tf32 = True
+
+class timer:
+    def __init__(self, method_name="timed process"):
+        self.method = method_name
+    def __enter__(self):
+        self.start = time.time()
+        print(f"{self.method} starts")
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        end = time.time()
+        print(f"{self.method} took {str(round(end - self.start, 2))}s")
+
+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')
 
-transformer = FluxTransformer2DModel.from_single_file(
-    "https://huggingface.co/Kijai/flux-fp8/blob/main/flux1-dev-fp8.safetensors",
+base_model = "black-forest-labs/FLUX.1-dev"
+pipe = FluxPipeline.from_pretrained(base_model, torch_dtype=torch.bfloat16, token=huggingface_token)
+
+# 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()
+pipe.enable_model_cpu_offload()
+
+
+# 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,
 )
-
-pipe = FluxPipeline.from_pretrained(
-    "black-forest-labs/FLUX.1-dev",
-    transformer=None,
-    torch_dtype=torch.bfloat16,
-    token=huggingface_token
+pipeline.scheduler = EulerAncestralDiscreteScheduler.from_config(
+    pipeline.scheduler.config, timestep_spacing='trailing'
 )
 
-pipe.transformer = transformer
+# 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)
 
-# Initialize Florence model
-florence_model = AutoModelForCausalLM.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True).to(device).eval()
-florence_processor = AutoProcessor.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True)
+pipeline = pipeline.to(device)
 
-# Prompt Enhancer
-enhancer_long = pipeline("summarization", model="gokaygokay/Lamini-Prompt-Enchance-Long", device=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)
 
-MAX_SEED = np.iinfo(np.int32).max
-MAX_IMAGE_SIZE = 2048
+model = model.to(device)
 
-# Florence caption function
-@spaces.GPU
-def florence_caption(image):
-    # Convert image to PIL if it's not already
-    if not isinstance(image, Image.Image):
-        image = Image.fromarray(image)
-    
-    inputs = florence_processor(text="<MORE_DETAILED_CAPTION>", images=image, return_tensors="pt").to(device)
-    generated_ids = florence_model.generate(
-        input_ids=inputs["input_ids"],
-        pixel_values=inputs["pixel_values"],
-        max_new_tokens=1024,
-        early_stopping=False,
-        do_sample=False,
-        num_beams=3,
-    )
-    generated_text = florence_processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
-    parsed_answer = florence_processor.post_process_generation(
-        generated_text,
-        task="<MORE_DETAILED_CAPTION>",
-        image_size=(image.width, image.height)
-    )
-    return parsed_answer["<MORE_DETAILED_CAPTION>"]
-
-# Prompt Enhancer function
-def enhance_prompt(input_prompt):
-    result = enhancer_long("Enhance the description: " + input_prompt)
-    enhanced_text = result[0]['summary_text']
-    return enhanced_text
-
-@spaces.GPU(duration=190)
-def process_workflow(image, text_prompt, use_enhancer, seed, randomize_seed, width, height, guidance_scale, num_inference_steps, progress=gr.Progress(track_tqdm=True)):
-    if image is not None:
-        # Convert image to PIL if it's not already
-        if not isinstance(image, Image.Image):
-            image = Image.fromarray(image)
-        
-        prompt = florence_caption(image)
-        print(prompt)
+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:
-        prompt = text_prompt
-    
-    if use_enhancer:
-        prompt = enhance_prompt(prompt)
-    
-    if randomize_seed:
-        seed = random.randint(0, MAX_SEED)
-    
-    generator = torch.Generator(device=device).manual_seed(seed)
-    
-    image = pipe(
-        prompt=prompt,
-        generator=generator,
-        num_inference_steps=num_inference_steps,
-        width=width,
-        height=height,
-        guidance_scale=guidance_scale
+        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 image, prompt, seed
-
-custom_css = """
-.input-group, .output-group {
-    border: 1px solid #e0e0e0;
-    border-radius: 10px;
-    padding: 20px;
-    margin-bottom: 20px;
-    background-color: #f9f9f9;
-}
-.submit-btn {
-    background-color: #2980b9 !important;
-    color: white !important;
-}
-.submit-btn:hover {
-    background-color: #3498db !important;
-}
-"""
-
-title = """<h1 align="center">FLUX.1-dev with Florence-2 Captioner and Prompt Enhancer</h1>
-<p><center>
-<a href="https://huggingface.co/black-forest-labs/FLUX.1-dev" target="_blank">[FLUX.1-dev Model]</a>
-<a href="https://huggingface.co/microsoft/Florence-2-base" target="_blank">[Florence-2 Model]</a>
-<a href="https://huggingface.co/gokaygokay/Lamini-Prompt-Enchance-Long" target="_blank">[Prompt Enhancer Long]</a>
-<p align="center">Create long prompts from images or enhance your short prompts with prompt enhancer</p>
-</center></p>
-"""
-
-with gr.Blocks(css=custom_css, theme=gr.themes.Soft(primary_hue="blue", secondary_hue="gray")) as demo:
-    gr.HTML(title)
-    
+    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;">Flux Image to 3D Model Generator</h1>
+        </div>
+        """
+    )
+
     with gr.Row():
-        with gr.Column(scale=1):
-            with gr.Group(elem_classes="input-group"):
-                input_image = gr.Image(label="Input Image (Florence-2 Captioner)")
+        with gr.Column(scale=3):
+            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):
-                text_prompt = gr.Textbox(label="Text Prompt (optional, used if no image is uploaded)")
-                use_enhancer = gr.Checkbox(label="Use Prompt Enhancer", value=False)
-                seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0)
-                randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
-                width = gr.Slider(label="Width", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=1024)
-                height = gr.Slider(label="Height", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=1024)
-                guidance_scale = gr.Slider(label="Guidance Scale", minimum=1, maximum=15, step=0.1, value=3.5)
-                num_inference_steps = gr.Slider(label="Inference Steps", minimum=1, maximum=50, step=1, value=28)
+                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", elem_classes="submit-btn")
-        
-        with gr.Column(scale=1):
-            with gr.Group(elem_classes="output-group"):
-                output_image = gr.Image(label="Result", elem_id="gallery", show_label=False)
-                final_prompt = gr.Textbox(label="Final Prompt Used")
-                used_seed = gr.Number(label="Seed Used")
-    
+            generate_btn = gr.Button("Generate 3D Model", variant="primary")
+
+        with gr.Column(scale=4):
+            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()
+
+    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(
-        fn=process_workflow,
-        inputs=[
-            input_image, text_prompt, use_enhancer, seed, randomize_seed,
-            width, height, guidance_scale, num_inference_steps
-        ],
-        outputs=[output_image, final_prompt, used_seed]
+        fn=process_pipeline,
+        inputs=[prompt, height, width, steps, scales, seed],
+        outputs=[flux_output, mv_show_images, output_model_obj, output_model_glb]
     )
 
-demo.launch(debug=True)
+if __name__ == "__main__":
+    demo.launch()