app.py

import gradio as gr
import spaces

import os
import shutil
import json
os.environ['TOKENIZERS_PARALLELISM'] = 'true'
os.environ['SPCONV_ALGO'] = 'native'
from typing import *
import torch
import numpy as np
import imageio
from easydict import EasyDict as edict
from trellis.pipelines import TrellisTextTo3DPipeline
from trellis.representations import Gaussian, MeshExtractResult
from trellis.utils import render_utils, postprocessing_utils

import traceback
import sys
import time


# Add JSON encoder for NumPy arrays
class NumpyEncoder(json.JSONEncoder):
    def default(self, obj):
        if isinstance(obj, np.ndarray):
            return obj.tolist()
        return json.JSONEncoder.default(self, obj)


MAX_SEED = np.iinfo(np.int32).max
TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'tmp')
os.makedirs(TMP_DIR, exist_ok=True)


def start_session(req: gr.Request):
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    os.makedirs(user_dir, exist_ok=True)
    
    
def end_session(req: gr.Request):
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    # Use shutil.rmtree with ignore_errors=True for robustness
    shutil.rmtree(user_dir, ignore_errors=True)


def pack_state(gs: Gaussian, mesh: MeshExtractResult) -> dict:
    return {
        'gaussian': {
            **gs.init_params,
            '_xyz': gs._xyz.cpu().numpy(),
            '_features_dc': gs._features_dc.cpu().numpy(),
            '_scaling': gs._scaling.cpu().numpy(),
            '_rotation': gs._rotation.cpu().numpy(),
            '_opacity': gs._opacity.cpu().numpy(),
        },
        'mesh': {
            'vertices': mesh.vertices.cpu().numpy(),
            'faces': mesh.faces.cpu().numpy(),
        },
    }
    
    
def unpack_state(state: dict) -> Tuple[Gaussian, edict, str]:
    gs = Gaussian(
        aabb=state['gaussian']['aabb'],
        sh_degree=state['gaussian']['sh_degree'],
        mininum_kernel_size=state['gaussian']['mininum_kernel_size'],
        scaling_bias=state['gaussian']['scaling_bias'],
        opacity_bias=state['gaussian']['opacity_bias'],
        scaling_activation=state['gaussian']['scaling_activation'],
    )
    # Ensure tensors are created on the correct device ('cuda')
    gs._xyz = torch.tensor(state['gaussian']['_xyz'], device='cuda', dtype=torch.float32)
    gs._features_dc = torch.tensor(state['gaussian']['_features_dc'], device='cuda', dtype=torch.float32)
    gs._scaling = torch.tensor(state['gaussian']['_scaling'], device='cuda', dtype=torch.float32)
    gs._rotation = torch.tensor(state['gaussian']['_rotation'], device='cuda', dtype=torch.float32)
    gs._opacity = torch.tensor(state['gaussian']['_opacity'], device='cuda', dtype=torch.float32)
    
    mesh = edict(
        vertices=torch.tensor(state['mesh']['vertices'], device='cuda', dtype=torch.float32),
        faces=torch.tensor(state['mesh']['faces'], device='cuda', dtype=torch.int64), # Faces are usually integers
    )
    
    return gs, mesh


def get_seed(randomize_seed: bool, seed: int) -> int:
    """
    Get the random seed.
    """
    return np.random.randint(0, MAX_SEED) if randomize_seed else seed


@spaces.GPU
def text_to_3d(
    prompt: str,
    seed: int,
    ss_guidance_strength: float,
    ss_sampling_steps: int,
    slat_guidance_strength: float,
    slat_sampling_steps: int,
    req: gr.Request,
) -> dict: # MODIFIED: Now returns only the state dict
    """
    Convert a text prompt to a 3D model state object.
    Args:
        prompt (str): The text prompt.
        seed (int): The random seed.
        ss_guidance_strength (float): The guidance strength for sparse structure generation.
        ss_sampling_steps (int): The number of sampling steps for sparse structure generation.
        slat_guidance_strength (float): The guidance strength for structured latent generation.
        slat_sampling_steps (int): The number of sampling steps for structured latent generation.
    Returns:
        dict: The JSON-serializable state object containing the generated 3D model info.
    """
    # Ensure user directory exists (redundant if start_session is always called, but safe)
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    os.makedirs(user_dir, exist_ok=True) 
    
    print(f"[{req.session_hash}] Running text_to_3d for prompt: {prompt}") # Add logging
    
    outputs = pipeline.run(
        prompt,
        seed=seed,
        formats=["gaussian", "mesh"],
        sparse_structure_sampler_params={
            "steps": ss_sampling_steps,
            "cfg_strength": ss_guidance_strength,
        },
        slat_sampler_params={
            "steps": slat_sampling_steps,
            "cfg_strength": slat_guidance_strength,
        },
    )

    # REMOVED: Video rendering logic moved to render_preview_video
    
    # Create the state object and ensure it's JSON serializable for API calls
    state = pack_state(outputs['gaussian'][0], outputs['mesh'][0])
    # Convert to serializable format
    serializable_state = json.loads(json.dumps(state, cls=NumpyEncoder))
    
    print(f"[{req.session_hash}] text_to_3d completed. Returning state.") # Modified log message
    
    torch.cuda.empty_cache()
        
    # --- REVERTED DEBUGGING --- 
    # Remove the temporary simple dictionary return
    # print("[DEBUG] Returning simple dict for API test.")
    # return {"status": "test_success", "received_prompt": prompt}
    # --- END REVERTED DEBUGGING ---
    
    # Original return line (restored):
    return serializable_state # MODIFIED: Return only state

# --- NEW FUNCTION ---
@spaces.GPU
def render_preview_video(state: dict, req: gr.Request) -> str:
    """
    Renders a preview video from the provided state object.
    Args:
        state (dict): The state object containing Gaussian and mesh data.
        req (gr.Request): Gradio request object for session hash.
    Returns:
        str: The path to the rendered video file.
    """
    if not state:
        print(f"[{req.session_hash}] render_preview_video called with empty state. Returning None.")
        return None 

    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    os.makedirs(user_dir, exist_ok=True) # Ensure directory exists
    
    print(f"[{req.session_hash}] Unpacking state for video rendering.") 
    # Only unpack gs, as mesh causes type errors with render_utils after unpacking
    gs, _ = unpack_state(state) # We still need the mesh for GLB, but not for this video preview
    
    print(f"[{req.session_hash}] Rendering video (Gaussian only)...") 
    # Render ONLY the Gaussian splats, as rendering the unpacked mesh fails
    video = render_utils.render_video(gs, num_frames=120)['color']
    # REMOVED: video_geo = render_utils.render_video(mesh, num_frames=120)['normal']
    # REMOVED: video = [np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))]
    
    video_path = os.path.join(user_dir, 'preview_sample.mp4') 
    print(f"[{req.session_hash}] Saving video to {video_path}") 
    # Save only the Gaussian render
    imageio.mimsave(video_path, video, fps=15)
    
    torch.cuda.empty_cache()
    return video_path
# --- END NEW FUNCTION ---


@spaces.GPU(duration=90)
def extract_glb(
    state: dict,
    mesh_simplify: float,
    texture_size: int,
    req: gr.Request,
) -> Tuple[str, str]:
    """
    Extract a GLB file from the 3D model state.
    Args:
        state (dict): The state of the generated 3D model.
        mesh_simplify (float): The mesh simplification factor.
        texture_size (int): The texture resolution.
    Returns:
        str: The path to the extracted GLB file (for Model3D component).
        str: The path to the extracted GLB file (for DownloadButton).
    """
    if not state:
       print(f"[{req.session_hash}] extract_glb called with empty state. Returning None.")
       return None, None # Return Nones if state is missing
       
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    os.makedirs(user_dir, exist_ok=True)
    
    print(f"[{req.session_hash}] Unpacking state for GLB extraction.") # Add logging
    gs, mesh = unpack_state(state)
    
    print(f"[{req.session_hash}] Extracting GLB (simplify={mesh_simplify}, texture={texture_size})...") # Add logging
    glb = postprocessing_utils.to_glb(gs, mesh, simplify=mesh_simplify, texture_size=texture_size, verbose=False)
    glb_path = os.path.join(user_dir, 'sample.glb')
    print(f"[{req.session_hash}] Saving GLB to {glb_path}") # Add logging
    glb.export(glb_path)
    
    torch.cuda.empty_cache()
    # Return the same path for both Model3D and DownloadButton components
    return glb_path, glb_path


@spaces.GPU
def extract_gaussian(state: dict, req: gr.Request) -> Tuple[str, str]:
    """
    Extract a Gaussian PLY file from the 3D model state.
    Args:
        state (dict): The state of the generated 3D model.
    Returns:
        str: The path to the extracted Gaussian file (for Model3D component).
        str: The path to the extracted Gaussian file (for DownloadButton).
    """
    if not state:
       print(f"[{req.session_hash}] extract_gaussian called with empty state. Returning None.")
       return None, None # Return Nones if state is missing
       
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    os.makedirs(user_dir, exist_ok=True)
    
    print(f"[{req.session_hash}] Unpacking state for Gaussian extraction.") # Add logging
    gs, _ = unpack_state(state)
    
    gaussian_path = os.path.join(user_dir, 'sample.ply')
    print(f"[{req.session_hash}] Saving Gaussian PLY to {gaussian_path}") # Add logging
    gs.save_ply(gaussian_path)
    
    torch.cuda.empty_cache()
    # Return the same path for both Model3D and DownloadButton components
    return gaussian_path, gaussian_path


# --- NEW COMBINED API FUNCTION ---
@spaces.GPU(duration=120) # Allow more time for combined generation + extraction
def generate_and_extract_glb(
    # Inputs mirror text_to_3d and extract_glb settings
    prompt: str,
    seed: int,
    ss_guidance_strength: float,
    ss_sampling_steps: int,
    slat_guidance_strength: float,
    slat_sampling_steps: int,
    mesh_simplify: float, # Added from extract_glb
    texture_size: int,    # Added from extract_glb
    req: gr.Request, # Keep req for potential session info if needed
) -> Optional[str]: # MODIFIED: Explicitly show it can return None
    """
    Combines 3D model generation and GLB extraction into a single step 
    for API usage, avoiding the need to transfer the state object.
    
    Args:
        prompt (str): Text prompt for generation.
        seed (int): Random seed.
        ss_guidance_strength (float): Sparse structure guidance.
        ss_sampling_steps (int): Sparse structure steps.
        slat_guidance_strength (float): Structured latent guidance.
        slat_sampling_steps (int): Structured latent steps.
        mesh_simplify (float): Mesh simplification factor for GLB.
        texture_size (int): Texture resolution for GLB.
        req (gr.Request): Gradio request object.
        
    Returns:
        Optional[str]: The absolute path to the generated GLB file within the Space's filesystem,
                       or None if any step fails.
    """
    session_hash = "API_CALL" # Use a generic identifier for API calls if req is None or lacks session
    if req and hasattr(req, 'session_hash') and req.session_hash:
        session_hash = req.session_hash
        
    user_dir = os.path.join(TMP_DIR, str(session_hash))
    try:
        os.makedirs(user_dir, exist_ok=True)
    except Exception as e:
        print(f"[{session_hash}] API: ERROR creating directory {user_dir}: {e}")
        return None # Cannot proceed without directory
    
    print(f"[{session_hash}] API: ===== generate_and_extract_glb START ====")
    print(f"[{session_hash}] API: Prompt: '{prompt}', Seed: {seed}, Simplify: {mesh_simplify}, Texture: {texture_size}")
    
    gs_output = None
    mesh_output = None
    
    # --- Step 1: Generate 3D Model (adapted from text_to_3d) ---
    try:
        print(f"[{session_hash}] API: Step 1 - Running generation pipeline...")
        t_start_gen = time.time() # Add timing
        outputs = pipeline.run(
            prompt,
            seed=seed,
            formats=["gaussian", "mesh"], # Need both for GLB extraction
            sparse_structure_sampler_params={
                "steps": ss_sampling_steps,
                "cfg_strength": ss_guidance_strength,
            },
            slat_sampler_params={
                "steps": slat_sampling_steps,
                "cfg_strength": slat_guidance_strength,
            },
        )
        t_end_gen = time.time()
        print(f"[{session_hash}] API: Step 1 - Generation pipeline completed in {t_end_gen - t_start_gen:.2f}s.")
        
        # Validate outputs immediately
        if not outputs or 'gaussian' not in outputs or not outputs['gaussian'] or 'mesh' not in outputs or not outputs['mesh']:
            print(f"[{session_hash}] API: ERROR - Pipeline output is missing expected keys or values.")
            return None
            
        gs_output = outputs['gaussian'][0]
        mesh_output = outputs['mesh'][0]
        
        if gs_output is None or mesh_output is None:
             print(f"[{session_hash}] API: ERROR - Pipeline returned None for gs_output or mesh_output.")
             return None
             
        print(f"[{session_hash}] API: Step 1 - Outputs obtained (gs type: {type(gs_output)}, mesh type: {type(mesh_output)}).")
        
    except Exception as e:
        print(f"[{session_hash}] API: ERROR during generation pipeline step: {e}")
        # Print detailed traceback
        traceback.print_exc()
        # Clean up CUDA memory before returning
        try: 
            torch.cuda.empty_cache()
            print(f"[{session_hash}] API: CUDA cache cleared after generation error.")
        except Exception as cache_e:
            print(f"[{session_hash}] API: Error clearing CUDA cache after generation error: {cache_e}")
        return None # Return None on failure
        
    # --- Step 2: Extract GLB (adapted from extract_glb) ---
    glb_path = None # Initialize glb_path
    try:
        print(f"[{session_hash}] API: Step 2 - Extracting GLB (simplify={mesh_simplify}, texture={texture_size})...")
        # Check if inputs from previous step are valid
        if gs_output is None or mesh_output is None:
             print(f"[{session_hash}] API: ERROR - Cannot proceed with GLB extraction, gs_output or mesh_output is None.")
             return None
             
        t_start_glb = time.time()
        # Directly use the outputs from the pipeline
        glb = postprocessing_utils.to_glb(gs_output, mesh_output, simplify=mesh_simplify, texture_size=texture_size, verbose=False)
        t_end_glb = time.time()
        print(f"[{session_hash}] API: Step 2 - GLB object created in {t_end_glb - t_start_glb:.2f}s.")
        
        if glb is None:
            print(f"[{session_hash}] API: ERROR - postprocessing_utils.to_glb returned None.")
            return None
            
        glb_path = os.path.join(user_dir, f'api_generated_{session_hash}.glb') # Use unique name
        print(f"[{session_hash}] API: Step 2 - Saving GLB to {glb_path}...")
        t_start_save = time.time()
        glb.export(glb_path)
        t_end_save = time.time()
        print(f"[{session_hash}] API: Step 2 - GLB saved in {t_end_save - t_start_save:.2f}s.")
        print(f"[{session_hash}] API: Step 2 - GLB extraction completed successfully.")
        
    except Exception as e:
        print(f"[{session_hash}] API: ERROR during GLB extraction step: {e}")
        # Print detailed traceback
        traceback.print_exc()
        # Clean up CUDA memory before returning
        try: 
            torch.cuda.empty_cache()
            print(f"[{session_hash}] API: CUDA cache cleared after extraction error.")
        except Exception as cache_e:
            print(f"[{session_hash}] API: Error clearing CUDA cache after extraction error: {cache_e}")
        return None # Return None on failure

    # --- Final Cleanup and Return ---
    try:
        torch.cuda.empty_cache()
        print(f"[{session_hash}] API: Final CUDA cache cleared.")
    except Exception as cache_e:
        print(f"[{session_hash}] API: Error clearing final CUDA cache: {cache_e}")
        
    if glb_path and os.path.exists(glb_path):
        print(f"[{session_hash}] API: ===== generate_and_extract_glb END (Success) ===== Returning GLB path: {glb_path}")
        return glb_path # Return only the path to the generated GLB
    else:
        print(f"[{session_hash}] API: ===== generate_and_extract_glb END (Failure) ===== GLB path not generated or does not exist.")
        return None # Ensure None is returned if glb_path wasn't set or file doesn't exist
# --- END NEW COMBINED API FUNCTION ---


# State object to hold the generated model info between steps
output_buf = gr.State() 
# Video component placeholder (will be populated by render_preview_video)
# video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300) # Defined later inside the Blocks

with gr.Blocks(delete_cache=(600, 600)) as demo:
    gr.Markdown("""
    ## Text to 3D Asset with [TRELLIS](https://trellis3d.github.io/)
    * Type a text prompt and click "Generate" to create a 3D asset.
    * The preview video will appear after generation.
    * If you find the generated 3D asset satisfactory, click "Extract GLB" or "Extract Gaussian" to extract the file and download it.
    """)
    
    with gr.Row():
        with gr.Column():
            text_prompt = gr.Textbox(label="Text Prompt", lines=5)
            
            with gr.Accordion(label="Generation Settings", open=False):
                seed = gr.Slider(0, MAX_SEED, label="Seed", value=0, step=1)
                randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
                gr.Markdown("Stage 1: Sparse Structure Generation")
                with gr.Row():
                    ss_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=7.5, step=0.1)
                    ss_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=25, step=1)
                gr.Markdown("Stage 2: Structured Latent Generation")
                with gr.Row():
                    slat_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=7.5, step=0.1)
                    slat_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=25, step=1)

            generate_btn = gr.Button("Generate")
            
            with gr.Accordion(label="GLB Extraction Settings", open=False):
                mesh_simplify = gr.Slider(0.9, 0.98, label="Simplify", value=0.95, step=0.01)
                texture_size = gr.Slider(512, 2048, label="Texture Size", value=1024, step=512)
            
            with gr.Row():
                # Buttons start non-interactive, enabled after generation
                extract_glb_btn = gr.Button("Extract GLB", interactive=False)
                extract_gs_btn = gr.Button("Extract Gaussian", interactive=False)
            gr.Markdown("""
                        *NOTE: Gaussian file can be very large (~50MB), it will take a while to display and download.*
                        """)

        with gr.Column():
            # Define UI components here
            video_output = gr.Video(label="Generated 3D Asset Preview", autoplay=True, loop=True, height=300)
            model_output = gr.Model3D(label="Extracted GLB/Gaussian", height=300)
            
            with gr.Row():
                 # Buttons start non-interactive, enabled after extraction
                download_glb = gr.DownloadButton(label="Download GLB", interactive=False)
                download_gs = gr.DownloadButton(label="Download Gaussian", interactive=False)  
    
    # Define the state buffer here, outside the component definitions but inside the Blocks scope
    output_buf = gr.State()

    # --- Handlers ---
    demo.load(start_session)
    demo.unload(end_session)

    # --- MODIFIED UI CHAIN ---
    # 1. Get Seed
    # 2. Run text_to_3d -> outputs state to output_buf
    # 3. Run render_preview_video (using state from output_buf) -> outputs video to video_output
    # 4. Enable extraction buttons
    generate_btn.click(
        get_seed,
        inputs=[randomize_seed, seed],
        outputs=[seed],
        queue=True # Use queue for potentially long-running steps
    ).then(
        text_to_3d,
        inputs=[text_prompt, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps],
        outputs=[output_buf], # text_to_3d now ONLY outputs state
        api_name="text_to_3d" # Keep API name consistent if needed
    ).then(
        render_preview_video, # NEW step: Render video from state
        inputs=[output_buf],
        outputs=[video_output],
        api_name="render_preview_video" # Assign API name if you want to call this separately
    ).then(
        lambda: tuple([gr.Button(interactive=True), gr.Button(interactive=True)]), # Enable extraction buttons
        outputs=[extract_glb_btn, extract_gs_btn],
    )

    # Clear video and disable extraction buttons if prompt is cleared or generation restarted
    # (Consider adding logic to clear prompt on successful generation if desired)
    text_prompt.change( # Example: Clear video if prompt changes
         lambda: (None, gr.Button(interactive=False), gr.Button(interactive=False)),
         outputs=[video_output, extract_glb_btn, extract_gs_btn]
    )
    video_output.clear( # This might be redundant if text_prompt.change handles it
        lambda: tuple([gr.Button(interactive=False), gr.Button(interactive=False)]),
        outputs=[extract_glb_btn, extract_gs_btn],
    )

    # --- Extraction Handlers ---
    # GLB Extraction: Takes state from output_buf, outputs model and download path
    extract_glb_btn.click(
        extract_glb,
        inputs=[output_buf, mesh_simplify, texture_size],
        outputs=[model_output, download_glb], # Outputs to Model3D and DownloadButton path
        api_name="extract_glb"
    ).then(
        lambda: gr.Button(interactive=True), # Enable download button
        outputs=[download_glb],
    )
    
    # Gaussian Extraction: Takes state from output_buf, outputs model and download path
    extract_gs_btn.click(
        extract_gaussian,
        inputs=[output_buf],
        outputs=[model_output, download_gs], # Outputs to Model3D and DownloadButton path
        api_name="extract_gaussian"
    ).then(
        lambda: gr.Button(interactive=True), # Enable download button
        outputs=[download_gs],
    )

    # Clear model and disable download buttons if video/state is cleared
    model_output.clear(
        lambda: (gr.Button(interactive=False), gr.Button(interactive=False)),
        outputs=[download_glb, download_gs], # Disable both download buttons
    )
    
    # --- NEW API ENDPOINT DEFINITION ---
    # Define the combined function as an API endpoint. 
    # This is *separate* from the UI button clicks.
    # It directly calls the combined function.
    demo.load(
        None, # No function needed on load for this endpoint
        inputs=[
            text_prompt, # Map inputs from API request data based on order
            seed,
            ss_guidance_strength,
            ss_sampling_steps,
            slat_guidance_strength,
            slat_sampling_steps,
            mesh_simplify,
            texture_size
        ],
        outputs=None, # Output is handled by the function return for the API
        api_name="generate_and_extract_glb" # Assign the specific API name
    )

# --- Launch the Gradio app ---
if __name__ == "__main__":
    print("Loading Trellis pipeline...")
    # Consider adding error handling for pipeline loading
    try:
        pipeline = TrellisTextTo3DPipeline.from_pretrained("JeffreyXiang/TRELLIS-text-xlarge")
        pipeline.cuda()
        print("Pipeline loaded successfully.")
    except Exception as e:
        print(f"Error loading pipeline: {e}")
        # Optionally exit or provide a fallback UI
        sys.exit(1) 
        
    print("Launching Gradio demo...")
    # Enable queue for handling multiple users/requests
    # Set share=True if you need a public link (requires login for private spaces)
    demo.queue().launch()