cpuai/Trellis.2.multiview
0
1import gradio as gr2from gradio_client import Client, handle_file3import spaces4from concurrent.futures import ThreadPoolExecutor5 6import os7os.environ["OPENCV_IO_ENABLE_OPENEXR"] = '1'8os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"9os.environ["ATTN_BACKEND"] = "flash_attn_3"10os.environ["FLEX_GEMM_AUTOTUNE_CACHE_PATH"] = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'autotune_cache.json')11os.environ["FLEX_GEMM_AUTOTUNER_VERBOSE"] = '1'12from datetime import datetime13import shutil14import cv215from typing import *16import torch17import numpy as np18from PIL import Image19import base6420import io21import tempfile22from trellis2.modules.sparse import SparseTensor23from trellis2.pipelines import Trellis2ImageTo3DPipeline24from trellis2.renderers import EnvMap25from trellis2.utils import render_utils26import o_voxel27 28# Patch postprocess module with local fix for cumesh.fill_holes() bug29import importlib.util30_local_postprocess = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'o-voxel', 'o_voxel', 'postprocess.py')31if os.path.exists(_local_postprocess):32 import sys33 _spec = importlib.util.spec_from_file_location('o_voxel.postprocess', _local_postprocess)34 _mod = importlib.util.module_from_spec(_spec)35 _spec.loader.exec_module(_mod)36 o_voxel.postprocess = _mod37 sys.modules['o_voxel.postprocess'] = _mod38 39 40MAX_SEED = np.iinfo(np.int32).max41TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'tmp')42MODES = [43 {"name": "Normal", "icon": "assets/app/normal.png", "render_key": "normal"},44 {"name": "Clay render", "icon": "assets/app/clay.png", "render_key": "clay"},45 {"name": "Base color", "icon": "assets/app/basecolor.png", "render_key": "base_color"},46 {"name": "HDRI forest", "icon": "assets/app/hdri_forest.png", "render_key": "shaded_forest"},47 {"name": "HDRI sunset", "icon": "assets/app/hdri_sunset.png", "render_key": "shaded_sunset"},48 {"name": "HDRI courtyard", "icon": "assets/app/hdri_courtyard.png", "render_key": "shaded_courtyard"},49]50STEPS = 851DEFAULT_MODE = 352DEFAULT_STEP = 353 54 55css = """56/* Overwrite Gradio Default Style */57.stepper-wrapper {58 padding: 0;59}60 61.stepper-container {62 padding: 0;63 align-items: center;64}65 66.step-button {67 flex-direction: row;68}69 70.step-connector {71 transform: none;72}73 74.step-number {75 width: 16px;76 height: 16px;77}78 79.step-label {80 position: relative;81 bottom: 0;82}83 84.wrap.center.full {85 inset: 0;86 height: 100%;87}88 89.wrap.center.full.translucent {90 background: var(--block-background-fill);91}92 93.meta-text-center {94 display: block !important;95 position: absolute !important;96 top: unset !important;97 bottom: 0 !important;98 right: 0 !important;99 transform: unset !important;100}101 102/* Previewer */103.previewer-container {104 position: relative;105 font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif;106 width: 100%;107 height: 722px;108 margin: 0 auto;109 padding: 20px;110 display: flex;111 flex-direction: column;112 align-items: center;113 justify-content: center;114}115 116.previewer-container .tips-icon {117 position: absolute;118 right: 10px;119 top: 10px;120 z-index: 10;121 border-radius: 10px;122 color: #fff;123 background-color: var(--color-accent);124 padding: 3px 6px;125 user-select: none;126}127 128.previewer-container .tips-text {129 position: absolute;130 right: 10px;131 top: 50px;132 color: #fff;133 background-color: var(--color-accent);134 border-radius: 10px;135 padding: 6px;136 text-align: left;137 max-width: 300px;138 z-index: 10;139 transition: all 0.3s;140 opacity: 0%;141 user-select: none;142}143 144.previewer-container .tips-text p {145 font-size: 14px;146 line-height: 1.2;147}148 149.tips-icon:hover + .tips-text {150 display: block;151 opacity: 100%;152}153 154/* Row 1: Display Modes */155.previewer-container .mode-row {156 width: 100%;157 display: flex;158 gap: 8px;159 justify-content: center;160 margin-bottom: 20px;161 flex-wrap: wrap;162}163.previewer-container .mode-btn {164 width: 24px;165 height: 24px;166 border-radius: 50%;167 cursor: pointer;168 opacity: 0.5;169 transition: all 0.2s;170 border: 2px solid var(--neutral-600, #555);171 object-fit: cover;172}173.previewer-container .mode-btn:hover { opacity: 0.9; transform: scale(1.1); }174.previewer-container .mode-btn.active {175 opacity: 1;176 border-color: var(--color-accent);177 transform: scale(1.1);178}179 180/* Row 2: Display Image */181.previewer-container .display-row {182 margin-bottom: 20px;183 min-height: 400px;184 width: 100%;185 flex-grow: 1;186 display: flex;187 justify-content: center;188 align-items: center;189}190.previewer-container .previewer-main-image {191 max-width: 100%;192 max-height: 100%;193 flex-grow: 1;194 object-fit: contain;195 display: none;196}197.previewer-container .previewer-main-image.visible {198 display: block;199}200 201/* Row 3: Custom HTML Slider */202.previewer-container .slider-row {203 width: 100%;204 display: flex;205 flex-direction: column;206 align-items: center;207 gap: 10px;208 padding: 0 10px;209}210 211.previewer-container input[type=range] {212 -webkit-appearance: none;213 width: 100%;214 max-width: 400px;215 background: transparent;216}217.previewer-container input[type=range]::-webkit-slider-runnable-track {218 width: 100%;219 height: 8px;220 cursor: pointer;221 background: var(--neutral-700, #404040);222 border-radius: 5px;223}224.previewer-container input[type=range]::-webkit-slider-thumb {225 height: 20px;226 width: 20px;227 border-radius: 50%;228 background: var(--color-accent);229 cursor: pointer;230 -webkit-appearance: none;231 margin-top: -6px;232 box-shadow: 0 2px 5px rgba(0,0,0,0.2);233 transition: transform 0.1s;234}235.previewer-container input[type=range]::-webkit-slider-thumb:hover {236 transform: scale(1.2);237}238 239/* Overwrite Previewer Block Style */240.gradio-container .padded:has(.previewer-container) {241 padding: 0 !important;242}243 244.gradio-container:has(.previewer-container) [data-testid="block-label"] {245 position: absolute;246 top: 0;247 left: 0;248}249"""250 251 252head = """253<script>254 function refreshView(mode, step) {255 // 1. Find current mode and step256 const allImgs = document.querySelectorAll('.previewer-main-image');257 for (let i = 0; i < allImgs.length; i++) {258 const img = allImgs[i];259 if (img.classList.contains('visible')) {260 const id = img.id;261 const [_, m, s] = id.split('-');262 if (mode === -1) mode = parseInt(m.slice(1));263 if (step === -1) step = parseInt(s.slice(1));264 break;265 }266 }267 268 // 2. Hide ALL images269 // We select all elements with class 'previewer-main-image'270 allImgs.forEach(img => img.classList.remove('visible'));271 272 // 3. Construct the specific ID for the current state273 // Format: view-m{mode}-s{step}274 const targetId = 'view-m' + mode + '-s' + step;275 const targetImg = document.getElementById(targetId);276 277 // 4. Show ONLY the target278 if (targetImg) {279 targetImg.classList.add('visible');280 }281 282 // 5. Update Button Highlights283 const allBtns = document.querySelectorAll('.mode-btn');284 allBtns.forEach((btn, idx) => {285 if (idx === mode) btn.classList.add('active');286 else btn.classList.remove('active');287 });288 }289 290 // --- Action: Switch Mode ---291 function selectMode(mode) {292 refreshView(mode, -1);293 }294 295 // --- Action: Slider Change ---296 function onSliderChange(val) {297 refreshView(-1, parseInt(val));298 }299</script>300"""301 302 303empty_html = f"""304<div class="previewer-container">305 <svg style=" opacity: .5; height: var(--size-5); color: var(--body-text-color);"306 xmlns="http://www.w3.org/2000/svg" width="100%" height="100%" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round" class="feather feather-image"><rect x="3" y="3" width="18" height="18" rx="2" ry="2"></rect><circle cx="8.5" cy="8.5" r="1.5"></circle><polyline points="21 15 16 10 5 21"></polyline></svg>307</div>308"""309 310 311def image_to_base64(image):312 buffered = io.BytesIO()313 image = image.convert("RGB")314 image.save(buffered, format="jpeg", quality=85)315 img_str = base64.b64encode(buffered.getvalue()).decode()316 return f"data:image/jpeg;base64,{img_str}"317 318 319def start_session(req: gr.Request):320 user_dir = os.path.join(TMP_DIR, str(req.session_hash))321 os.makedirs(user_dir, exist_ok=True)322 323 324def end_session(req: gr.Request):325 user_dir = os.path.join(TMP_DIR, str(req.session_hash))326 if os.path.exists(user_dir):327 shutil.rmtree(user_dir)328 329 330def remove_background(input: Image.Image) -> Image.Image:331 try:332 with tempfile.NamedTemporaryFile(suffix='.png') as f:333 input = input.convert('RGB')334 input.save(f.name)335 output = rmbg_client.predict(handle_file(f.name), api_name="/image")[0][0]336 output = Image.open(output)337 return output338 except Exception as e:339 raise gr.Error(f"Background removal failed: {e}. Please upload images with transparent backgrounds (RGBA), or try again later.")340 341 342def preprocess_image(input: Image.Image) -> Image.Image:343 """344 Preprocess the input image.345 """346 # if has alpha channel, use it directly; otherwise, remove background347 has_alpha = False348 if input.mode == 'RGBA':349 alpha = np.array(input)[:, :, 3]350 if not np.all(alpha == 255):351 has_alpha = True352 max_size = max(input.size)353 scale = min(1, 1024 / max_size)354 if scale < 1:355 input = input.resize((int(input.width * scale), int(input.height * scale)), Image.Resampling.LANCZOS)356 if has_alpha:357 output = input358 else:359 output = remove_background(input)360 output_np = np.array(output)361 alpha = output_np[:, :, 3]362 bbox = np.argwhere(alpha > 0.8 * 255)363 if bbox.size == 0:364 # No visible pixels, center the image in a square365 size = max(output.size)366 square = Image.new('RGB', (size, size), (0, 0, 0))367 output_rgb = output.convert('RGB') if output.mode == 'RGBA' else output368 square.paste(output_rgb, ((size - output.width) // 2, (size - output.height) // 2))369 return square370 bbox = np.min(bbox[:, 1]), np.min(bbox[:, 0]), np.max(bbox[:, 1]), np.max(bbox[:, 0])371 center = (bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2372 size = max(bbox[2] - bbox[0], bbox[3] - bbox[1])373 size = int(size * 1)374 bbox = center[0] - size // 2, center[1] - size // 2, center[0] + size // 2, center[1] + size // 2375 output = output.crop(bbox) # type: ignore376 output_np = np.array(output).astype(np.float32)377 rgb = output_np[:, :, :3]378 alpha = output_np[:, :, 3:4] / 255.0379 # Keep full RGB for visible pixels, zero out transparent background380 mask = (alpha > 0.05).astype(np.float32)381 rgb = rgb * mask382 output = Image.fromarray(rgb.astype(np.uint8))383 return output384 385 386def pack_state(latents: Tuple[SparseTensor, SparseTensor, int]) -> dict:387 shape_slat, tex_slat, res = latents388 return {389 'shape_slat_feats': shape_slat.feats.cpu().numpy(),390 'tex_slat_feats': tex_slat.feats.cpu().numpy(),391 'coords': shape_slat.coords.cpu().numpy(),392 'res': res,393 }394 395 396def unpack_state(state: dict) -> Tuple[SparseTensor, SparseTensor, int]:397 shape_slat = SparseTensor(398 feats=torch.from_numpy(state['shape_slat_feats']).cuda(),399 coords=torch.from_numpy(state['coords']).cuda(),400 )401 tex_slat = shape_slat.replace(torch.from_numpy(state['tex_slat_feats']).cuda())402 return shape_slat, tex_slat, state['res']403 404 405def get_seed(randomize_seed, seed):406 """407 Get the random seed.408 """409 return np.random.randint(0, MAX_SEED) if randomize_seed else seed410 411 412def prepare_multi_example() -> List[str]:413 """414 Prepare multi-image examples. Returns list of image paths.415 Shows only the first view as representative thumbnail.416 """417 multi_case = sorted(set([i.split('_')[0] for i in os.listdir("assets/example_multi_image")]))418 examples = []419 for case in multi_case:420 first_img = f'assets/example_multi_image/{case}_1.png'421 if os.path.exists(first_img):422 examples.append(first_img)423 return examples424 425 426def load_multi_example(image) -> List[Image.Image]:427 """Load all views for a multi-image case by matching the input image."""428 if image is None:429 return []430 431 # Convert to PIL Image if needed432 if isinstance(image, np.ndarray):433 image = Image.fromarray(image)434 435 # Convert to RGB for consistent comparison436 input_rgb = np.array(image.convert('RGB'))437 438 # Find matching case by comparing with first images439 example_dir = "assets/example_multi_image"440 case_names = sorted(set([f.rsplit('_', 1)[0] for f in os.listdir(example_dir) if f.endswith('.png')]))441 442 for case_name in case_names:443 first_img_path = f'{example_dir}/{case_name}_1.png'444 if os.path.exists(first_img_path):445 first_img = Image.open(first_img_path).convert('RGB')446 first_rgb = np.array(first_img)447 448 # Compare images (check if same shape and content)449 if input_rgb.shape == first_rgb.shape and np.array_equal(input_rgb, first_rgb):450 # Found match, load all views (without preprocessing - will be done on Generate)451 images = []452 for i in range(1, 7):453 img_path = f'{example_dir}/{case_name}_{i}.png'454 if os.path.exists(img_path):455 img = Image.open(img_path).convert('RGBA')456 images.append(img)457 if images:458 return images459 460 # No match found, return the single image461 return [image.convert('RGBA') if image.mode != 'RGBA' else image]462 463 464def split_image(image: Image.Image) -> List[Image.Image]:465 """466 Split a concatenated image into multiple views.467 """468 image = np.array(image)469 alpha = image[..., 3]470 alpha = np.any(alpha > 0, axis=0)471 start_pos = np.where(~alpha[:-1] & alpha[1:])[0].tolist()472 end_pos = np.where(alpha[:-1] & ~alpha[1:])[0].tolist()473 images = []474 for s, e in zip(start_pos, end_pos):475 images.append(Image.fromarray(image[:, s:e+1]))476 return [preprocess_image(image) for image in images]477 478 479@spaces.GPU(duration=120)480def image_to_3d(481 multiimages,482 seed,483 resolution,484 ss_guidance_strength,485 ss_guidance_rescale,486 ss_sampling_steps,487 ss_rescale_t,488 shape_slat_guidance_strength,489 shape_slat_guidance_rescale,490 shape_slat_sampling_steps,491 shape_slat_rescale_t,492 tex_slat_guidance_strength,493 tex_slat_guidance_rescale,494 tex_slat_sampling_steps,495 tex_slat_rescale_t,496 multiimage_algo,497 tex_multiimage_algo,498 req: gr.Request,499 progress=gr.Progress(track_tqdm=True),500):501 if not multiimages:502 raise gr.Error("Please upload images or select an example first.")503 504 # Preprocess images (background removal for images without alpha)505 images = [image[0] for image in multiimages]506 processed_images = [preprocess_image(img) for img in images]507 508 # --- Sampling ---509 outputs, latents = pipeline.run_multi_image(510 processed_images,511 seed=seed,512 preprocess_image=False,513 sparse_structure_sampler_params={514 "steps": ss_sampling_steps,515 "guidance_strength": ss_guidance_strength,516 "guidance_rescale": ss_guidance_rescale,517 "rescale_t": ss_rescale_t,518 },519 shape_slat_sampler_params={520 "steps": shape_slat_sampling_steps,521 "guidance_strength": shape_slat_guidance_strength,522 "guidance_rescale": shape_slat_guidance_rescale,523 "rescale_t": shape_slat_rescale_t,524 },525 tex_slat_sampler_params={526 "steps": tex_slat_sampling_steps,527 "guidance_strength": tex_slat_guidance_strength,528 "guidance_rescale": tex_slat_guidance_rescale,529 "rescale_t": tex_slat_rescale_t,530 },531 pipeline_type={532 "512": "512",533 "1024": "1024_cascade",534 "1536": "1536_cascade",535 }[resolution],536 return_latent=True,537 mode=multiimage_algo,538 tex_mode=tex_multiimage_algo,539 )540 mesh = outputs[0]541 mesh.simplify(16777216) # nvdiffrast limit542 images = render_utils.render_snapshot(mesh, resolution=1024, r=2, fov=36, nviews=STEPS, envmap=envmap)543 state = pack_state(latents)544 torch.cuda.empty_cache()545 546 # --- HTML Construction ---547 def encode_preview_image(args):548 m_idx, s_idx, render_key = args549 img_base64 = image_to_base64(Image.fromarray(images[render_key][s_idx]))550 return (m_idx, s_idx, img_base64)551 552 encode_tasks = [553 (m_idx, s_idx, mode['render_key'])554 for m_idx, mode in enumerate(MODES)555 for s_idx in range(STEPS)556 ]557 558 with ThreadPoolExecutor(max_workers=8) as executor:559 encoded_results = list(executor.map(encode_preview_image, encode_tasks))560 561 encoded_map = {(m, s): b64 for m, s, b64 in encoded_results}562 images_html = ""563 for m_idx, mode in enumerate(MODES):564 for s_idx in range(STEPS):565 unique_id = f"view-m{m_idx}-s{s_idx}"566 is_visible = (m_idx == DEFAULT_MODE and s_idx == DEFAULT_STEP)567 vis_class = "visible" if is_visible else ""568 img_base64 = encoded_map[(m_idx, s_idx)]569 570 images_html += f"""571 <img id="{unique_id}"572 class="previewer-main-image {vis_class}"573 src="{img_base64}"574 loading="eager">575 """576 577 btns_html = ""578 for idx, mode in enumerate(MODES):579 active_class = "active" if idx == DEFAULT_MODE else ""580 btns_html += f"""581 <img src="{mode['icon_base64']}"582 class="mode-btn {active_class}"583 onclick="selectMode({idx})"584 title="{mode['name']}">585 """586 587 full_html = f"""588 <div class="previewer-container">589 <div class="tips-wrapper">590 <div class="tips-icon">💡Tips</div>591 <div class="tips-text">592 <p>● <b>Render Mode</b> - Click on the circular buttons to switch between different render modes.</p>593 <p>● <b>View Angle</b> - Drag the slider to change the view angle.</p>594 </div>595 </div>596 597 <!-- Row 1: Viewport containing 48 static <img> tags -->598 <div class="display-row">599 {images_html}600 </div>601 602 <!-- Row 2 -->603 <div class="mode-row" id="btn-group">604 {btns_html}605 </div>606 607 <!-- Row 3: Slider -->608 <div class="slider-row">609 <input type="range" id="custom-slider" min="0" max="{STEPS - 1}" value="{DEFAULT_STEP}" step="1" oninput="onSliderChange(this.value)">610 </div>611 </div>612 """613 614 return state, full_html615 616 617@spaces.GPU(duration=120)618def extract_glb(619 state,620 decimation_target,621 texture_size,622 req: gr.Request,623 progress=gr.Progress(track_tqdm=True),624):625 """626 Extract a GLB file from the 3D model.627 628 Args:629 state (dict): The state of the generated 3D model.630 decimation_target (int): The target face count for decimation.631 texture_size (int): The texture resolution.632 633 Returns:634 Tuple[str, str]: The path to the extracted GLB file (for Model3D and DownloadButton).635 """636 user_dir = os.path.join(TMP_DIR, str(req.session_hash))637 shape_slat, tex_slat, res = unpack_state(state)638 mesh = pipeline.decode_latent(shape_slat, tex_slat, res)[0]639 mesh.simplify(16777216) # nvdiffrast limit640 glb = o_voxel.postprocess.to_glb(641 vertices=mesh.vertices,642 faces=mesh.faces,643 attr_volume=mesh.attrs,644 coords=mesh.coords,645 attr_layout=pipeline.pbr_attr_layout,646 grid_size=res,647 aabb=[[-0.5, -0.5, -0.5], [0.5, 0.5, 0.5]],648 decimation_target=decimation_target,649 texture_size=texture_size,650 remesh=True,651 remesh_band=1,652 remesh_project=0,653 use_tqdm=True,654 )655 now = datetime.now()656 timestamp = now.strftime("%Y-%m-%dT%H%M%S") + f".{now.microsecond // 1000:03d}"657 os.makedirs(user_dir, exist_ok=True)658 glb_path = os.path.join(user_dir, f'sample_{timestamp}.glb')659 glb.export(glb_path, extension_webp=False)660 torch.cuda.empty_cache()661 return glb_path, glb_path662 663 664with gr.Blocks(theme=gr.themes.Soft(primary_hue="orange", neutral_hue="slate"), css=css, head=head) as demo:665 gr.HTML("""666 <div style="display: flex; align-items: center; gap: 24px;">667 <a href="https://www.opsiclear.com" target="_blank" style="flex-shrink: 0; display: flex; align-items: center;">668 <img src="https://www.opsiclear.com/assets/logos/Logo_v2_compact_name.svg" alt="OpsiClear"669 style="width: 140px; height: auto; object-fit: contain;">670 </a>671 <div style="min-width: 0; border-left: 2px solid var(--border-color-primary); padding-left: 24px;">672 <h2 style="margin: 0 0 8px 0; font-size: 1.4rem; line-height: 1.3; font-weight: 700;">Multi-View to 3D with <a href="https://microsoft.github.io/TRELLIS.2" target="_blank" style="text-decoration: none; color: var(--color-accent);">TRELLIS.2</a></h2>673 <ul style="margin: 0; padding-left: 18px; font-size: 0.88rem; line-height: 1.7; color: var(--body-text-color-subdued, var(--body-text-color));">674 <li>Upload multiple images from different viewpoints to create a 3D asset with multi-image conditioning.</li>675 <li>Click an example below to load a pre-made multi-view set, or upload your own images.</li>676 <li>Click <b>Generate</b> to create the 3D model, then <b>Extract GLB</b> to export.</li>677 <li style="color: #e67300;"><b>Note:</b> Generation quality is highly sensitive to parameters. Adjust settings in Advanced Settings if results are unsatisfactory.</li>678 <li style="color: #cc3333;"><b>Non-Commercial:</b> This space uses models with licenses that <b>forbid commercial use</b> (BRIA RMBG-2.0: CC BY-NC 4.0, nvdiffrast/nvdiffrec: NVIDIA Source Code License).</li>679 </ul>680 </div>681 </div>682 """)683 684 with gr.Row():685 with gr.Column(scale=1, min_width=360):686 multiimage_prompt = gr.Gallery(label="Multi-View Images", format="png", type="pil", height=400, columns=3, interactive=True)687 remove_img_btn = gr.Button("Remove Selected Image", size="sm", variant="secondary")688 689 resolution = gr.Radio(["512", "1024", "1536"], label="Resolution", value="1024")690 seed = gr.Slider(0, MAX_SEED, label="Seed", value=0, step=1)691 randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)692 decimation_target = gr.Slider(100000, 500000, label="Decimation Target", value=300000, step=10000)693 texture_size = gr.Slider(1024, 4096, label="Texture Size", value=2048, step=1024)694 695 with gr.Accordion(label="Advanced Settings", open=False):696 gr.Markdown("Stage 1: Sparse Structure Generation")697 with gr.Row():698 ss_guidance_strength = gr.Slider(1.0, 10.0, label="Guidance Strength", value=7.5, step=0.1)699 ss_guidance_rescale = gr.Slider(0.0, 1.0, label="Guidance Rescale", value=0.7, step=0.01)700 ss_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=12, step=1)701 ss_rescale_t = gr.Slider(1.0, 6.0, label="Rescale T", value=5.0, step=0.1)702 gr.Markdown("Stage 2: Shape Generation")703 with gr.Row():704 shape_slat_guidance_strength = gr.Slider(1.0, 10.0, label="Guidance Strength", value=7.5, step=0.1)705 shape_slat_guidance_rescale = gr.Slider(0.0, 1.0, label="Guidance Rescale", value=0.5, step=0.01)706 shape_slat_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=12, step=1)707 shape_slat_rescale_t = gr.Slider(1.0, 6.0, label="Rescale T", value=3.0, step=0.1)708 gr.Markdown("Stage 3: Material Generation")709 with gr.Row():710 tex_slat_guidance_strength = gr.Slider(1.0, 10.0, label="Guidance Strength", value=1.0, step=0.1)711 tex_slat_guidance_rescale = gr.Slider(0.0, 1.0, label="Guidance Rescale", value=0.0, step=0.01)712 tex_slat_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=12, step=1)713 tex_slat_rescale_t = gr.Slider(1.0, 6.0, label="Rescale T", value=3.0, step=0.1)714 multiimage_algo = gr.Radio(["stochastic", "multidiffusion"], label="Structure Algorithm", value="stochastic")715 tex_multiimage_algo = gr.Radio(["stochastic", "multidiffusion"], label="Texture Algorithm", value="multidiffusion")716 717 with gr.Column(scale=10):718 preview_output = gr.HTML(empty_html, label="3D Asset Preview", show_label=True, container=True)719 720 with gr.Row():721 generate_btn = gr.Button("Generate", variant="primary")722 extract_btn = gr.Button("Extract GLB")723 724 glb_output = gr.Model3D(label="Extracted GLB", height=600, show_label=True, display_mode="solid", clear_color=(0.25, 0.25, 0.25, 1.0))725 download_btn = gr.DownloadButton(label="Download GLB")726 727 with gr.Accordion(label="Examples", open=True):728 example_image = gr.Image(visible=False) # Hidden component for examples729 examples_multi = gr.Examples(730 examples=prepare_multi_example(),731 inputs=[example_image],732 fn=load_multi_example,733 outputs=[multiimage_prompt],734 run_on_click=True,735 cache_examples=False,736 examples_per_page=50,737 )738 739 output_buf = gr.State()740 selected_img_idx = gr.State(value=None)741 742 743 # Handlers744 demo.load(start_session)745 demo.unload(end_session)746 747 def on_gallery_select(evt: gr.SelectData):748 return evt.index749 750 def remove_selected_image(images, idx):751 if images is None or idx is None or not images:752 return images, None753 images = list(images)754 if idx < len(images):755 images.pop(idx)756 return images, None757 758 multiimage_prompt.select(on_gallery_select, outputs=[selected_img_idx])759 remove_img_btn.click(760 remove_selected_image,761 inputs=[multiimage_prompt, selected_img_idx],762 outputs=[multiimage_prompt, selected_img_idx],763 )764 765 generate_btn.click(766 get_seed,767 inputs=[randomize_seed, seed],768 outputs=[seed],769 ).then(770 image_to_3d,771 inputs=[772 multiimage_prompt, seed, resolution,773 ss_guidance_strength, ss_guidance_rescale, ss_sampling_steps, ss_rescale_t,774 shape_slat_guidance_strength, shape_slat_guidance_rescale, shape_slat_sampling_steps, shape_slat_rescale_t,775 tex_slat_guidance_strength, tex_slat_guidance_rescale, tex_slat_sampling_steps, tex_slat_rescale_t,776 multiimage_algo, tex_multiimage_algo777 ],778 outputs=[output_buf, preview_output],779 )780 781 extract_btn.click(782 extract_glb,783 inputs=[output_buf, decimation_target, texture_size],784 outputs=[glb_output, download_btn],785 )786 787 788# Launch the Gradio app789if __name__ == "__main__":790 os.makedirs(TMP_DIR, exist_ok=True)791 792 # Construct ui components793 btn_img_base64_strs = {}794 for i in range(len(MODES)):795 icon = Image.open(MODES[i]['icon'])796 MODES[i]['icon_base64'] = image_to_base64(icon)797 798 rmbg_client = Client("briaai/BRIA-RMBG-2.0")799 pipeline = Trellis2ImageTo3DPipeline.from_pretrained('microsoft/TRELLIS.2-4B')800 pipeline.rembg_model = None801 pipeline.low_vram = False802 pipeline.cuda()803 804 envmap = {805 'forest': EnvMap(torch.tensor(806 cv2.cvtColor(cv2.imread('assets/hdri/forest.exr', cv2.IMREAD_UNCHANGED), cv2.COLOR_BGR2RGB),807 dtype=torch.float32, device='cuda'808 )),809 'sunset': EnvMap(torch.tensor(810 cv2.cvtColor(cv2.imread('assets/hdri/sunset.exr', cv2.IMREAD_UNCHANGED), cv2.COLOR_BGR2RGB),811 dtype=torch.float32, device='cuda'812 )),813 'courtyard': EnvMap(torch.tensor(814 cv2.cvtColor(cv2.imread('assets/hdri/courtyard.exr', cv2.IMREAD_UNCHANGED), cv2.COLOR_BGR2RGB),815 dtype=torch.float32, device='cuda'816 )),817 }818 819 demo.launch()820