import gradio as gr import spaces import os import shutil 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 # --- FastAPI / Threading Imports --- import threading import uvicorn import logging from fastapi import FastAPI, HTTPException from pydantic import BaseModel 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) # --- Global Pipeline Variable --- pipeline = None # --- Logging Setup --- logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) 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)) shutil.rmtree(user_dir) 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'], ) gs._xyz = torch.tensor(state['gaussian']['_xyz'], device='cuda') gs._features_dc = torch.tensor(state['gaussian']['_features_dc'], device='cuda') gs._scaling = torch.tensor(state['gaussian']['_scaling'], device='cuda') gs._rotation = torch.tensor(state['gaussian']['_rotation'], device='cuda') gs._opacity = torch.tensor(state['gaussian']['_opacity'], device='cuda') mesh = edict( vertices=torch.tensor(state['mesh']['vertices'], device='cuda'), faces=torch.tensor(state['mesh']['faces'], device='cuda'), ) 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, ) -> Tuple[dict, str]: """ Convert an text prompt to a 3D model. 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 information of the generated 3D model. str: The path to the video of the 3D model. """ # --- Determine user_dir robustly --- session_hash_str = str(req.session_hash) if hasattr(req, 'session_hash') and req.session_hash else f"api_call_{np.random.randint(10000)}" user_dir = os.path.join(TMP_DIR, session_hash_str) os.makedirs(user_dir, exist_ok=True) # Ensure directory exists 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, }, ) video = render_utils.render_video(outputs['gaussian'][0], num_frames=120)['color'] video_geo = render_utils.render_video(outputs['mesh'][0], num_frames=120)['normal'] video = [np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))] video_path = os.path.join(user_dir, 'sample.mp4') try: imageio.mimsave(video_path, video, fps=15) # Now the directory should exist except FileNotFoundError: print(f"ERROR: Directory {user_dir} still not found before mimsave!", file=sys.stderr) # Decide if we should raise or return an error state? # Returning a dummy path might hide the error, so let's raise for now raise state = pack_state(outputs['gaussian'][0], outputs['mesh'][0]) torch.cuda.empty_cache() return state, video_path @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. 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. """ # --- Determine user_dir robustly --- session_hash_str = str(req.session_hash) if hasattr(req, 'session_hash') and req.session_hash else f"api_call_{np.random.randint(10000)}" user_dir = os.path.join(TMP_DIR, session_hash_str) os.makedirs(user_dir, exist_ok=True) # Ensure directory exists gs, mesh = unpack_state(state) 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') try: glb.export(glb_path) # Now the directory should exist except FileNotFoundError: print(f"ERROR: Directory {user_dir} still not found before glb.export!", file=sys.stderr) raise torch.cuda.empty_cache() return glb_path, glb_path @spaces.GPU def extract_gaussian(state: dict, req: gr.Request) -> Tuple[str, str]: """ Extract a Gaussian file from the 3D model. Args: state (dict): The state of the generated 3D model. Returns: str: The path to the extracted Gaussian file. """ # --- Determine user_dir robustly --- session_hash_str = str(req.session_hash) if hasattr(req, 'session_hash') and req.session_hash else f"api_call_{np.random.randint(10000)}" user_dir = os.path.join(TMP_DIR, session_hash_str) os.makedirs(user_dir, exist_ok=True) # Ensure directory exists gs, _ = unpack_state(state) gaussian_path = os.path.join(user_dir, 'sample.ply') try: gs.save_ply(gaussian_path) # Now the directory should exist except FileNotFoundError: print(f"ERROR: Directory {user_dir} still not found before gs.save_ply!", file=sys.stderr) raise torch.cuda.empty_cache() return gaussian_path, gaussian_path # --- FastAPI App Setup --- api_app = FastAPI() class GenerateRequest(BaseModel): prompt: str seed: int = 0 # Default seed mesh_simplify: float = 0.95 # Default simplify factor texture_size: int = 1024 # Default texture size # Add other generation parameters if needed @api_app.post("/api/generate-sync") async def generate_sync_api(request_data: GenerateRequest): global pipeline # Access the globally initialized pipeline if pipeline is None: logger.error("API Error: Pipeline not initialized") raise HTTPException(status_code=503, detail="Pipeline not ready") prompt = request_data.prompt seed = request_data.seed mesh_simplify = request_data.mesh_simplify texture_size = request_data.texture_size # Extract other params if added to GenerateRequest logger.info(f"API /generate-sync received prompt: {prompt}") try: # --- Determine a unique temporary directory for this API call --- api_call_hash = f"api_sync_{np.random.randint(100000)}" user_dir = os.path.join(TMP_DIR, api_call_hash) os.makedirs(user_dir, exist_ok=True) logger.info(f"API using temp dir: {user_dir}") # --- Stage 1: Run the text-to-3D pipeline --- logger.info("API running pipeline...") # Use default values for parameters not exposed in the simple API for now outputs = pipeline.run( prompt, seed=seed, formats=["gaussian", "mesh"], sparse_structure_sampler_params={"steps": 25, "cfg_strength": 7.5}, slat_sampler_params={"steps": 25, "cfg_strength": 7.5}, ) gs = outputs['gaussian'][0] mesh = outputs['mesh'][0] logger.info("API pipeline finished.") torch.cuda.empty_cache() # --- Stage 2: Extract GLB --- logger.info("API extracting GLB...") glb = postprocessing_utils.to_glb(gs, mesh, simplify=mesh_simplify, texture_size=texture_size, verbose=False) glb_path = os.path.join(user_dir, 'generated_sync.glb') glb.export(glb_path) logger.info(f"API GLB exported to: {glb_path}") torch.cuda.empty_cache() # Return the absolute path within the container return {"status": "success", "glb_path": os.path.abspath(glb_path)} except Exception as e: logger.error(f"API /generate-sync error: {str(e)}", exc_info=True) # Clean up temp dir on error if it exists if os.path.exists(user_dir): try: shutil.rmtree(user_dir) except Exception as cleanup_e: logger.error(f"API Error cleaning up dir {user_dir}: {cleanup_e}") raise HTTPException(status_code=500, detail=f"Generation failed: {str(e)}") # Note: We don't automatically clean up the user_dir on success, # as the file needs to be accessible for download by the calling server. # A separate cleanup mechanism might be needed eventually. # --- Gradio Blocks Definition --- 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. * If you find the generated 3D asset satisfactory, click "Extract GLB" to extract the GLB 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(): 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(): video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300) model_output = gr.Model3D(label="Extracted GLB/Gaussian", height=300) with gr.Row(): download_glb = gr.DownloadButton(label="Download GLB", interactive=False) download_gs = gr.DownloadButton(label="Download Gaussian", interactive=False) output_buf = gr.State() # Handlers demo.load(start_session) demo.unload(end_session) generate_btn.click( get_seed, inputs=[randomize_seed, seed], outputs=[seed], ).then( text_to_3d, inputs=[text_prompt, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps], outputs=[output_buf, video_output], ).then( lambda: tuple([gr.Button(interactive=True), gr.Button(interactive=True)]), outputs=[extract_glb_btn, extract_gs_btn], ) video_output.clear( lambda: tuple([gr.Button(interactive=False), gr.Button(interactive=False)]), outputs=[extract_glb_btn, extract_gs_btn], ) extract_glb_btn.click( extract_glb, inputs=[output_buf, mesh_simplify, texture_size], outputs=[model_output, download_glb], ).then( lambda: gr.Button(interactive=True), outputs=[download_glb], ) extract_gs_btn.click( extract_gaussian, inputs=[output_buf], outputs=[model_output, download_gs], ).then( lambda: gr.Button(interactive=True), outputs=[download_gs], ) model_output.clear( lambda: gr.Button(interactive=False), outputs=[download_glb], ) # --- Functions to Run FastAPI in Background --- def run_api(): """Run the FastAPI server.""" uvicorn.run(api_app, host="0.0.0.0", port=8000) # Run on port 8000 def start_api_thread(): """Start the API server in a background thread.""" api_thread = threading.Thread(target=run_api, daemon=True) api_thread.start() logger.info("Started FastAPI server thread on port 8000") return api_thread # Launch the Gradio app and FastAPI server if __name__ == "__main__": logger.info("Initializing Trellis Pipeline...") # Make pipeline global so API endpoint can access it pipeline = TrellisTextTo3DPipeline.from_pretrained("JeffreyXiang/TRELLIS-text-xlarge") pipeline.cuda() logger.info("Trellis Pipeline Initialized.") # Start the background API server start_api_thread() # Launch the Gradio interface (blocking call) logger.info("Launching Gradio Demo...") demo.launch()