"""Drape — snap an outfit, shop every piece. (Gradio app for Hugging Face Spaces)""" from __future__ import annotations import inspect import logging import random import traceback from pathlib import Path import gradio as gr import numpy as np from drape import __version__ from drape.config import settings from drape.index import CatalogIndex, Match from drape.taxonomy import GROUP_LABELS, GROUPS logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s") log = logging.getLogger("drape") ROOT = Path(__file__).parent MAX_SLOTS = 6 # --------------------------------------------------------------------------- # ZeroGPU: free Hugging Face accounts can only host Gradio apps on ZeroGPU # hardware, which refuses to start unless at least one function is decorated # with @spaces.GPU. Drape runs comfortably on CPU, so this stub is never called # and visitors never spend GPU quota. # --------------------------------------------------------------------------- try: import spaces # type: ignore @spaces.GPU def _zerogpu_placeholder(): # pragma: no cover return None except ImportError: # local runs / CI pass # --------------------------------------------------------------------------- # Startup: load the catalog and the models once. # --------------------------------------------------------------------------- ENGINE = None INDEX: CatalogIndex | None = None STARTUP_ERROR: str | None = None def _load_engine(): global ENGINE, INDEX, STARTUP_ERROR try: if settings.index_dir: log.info("Loading catalog index from %s", settings.index_dir) INDEX = CatalogIndex.load(settings.index_dir) else: log.info("Loading catalog index from the Hub: %s", settings.index_repo) INDEX = CatalogIndex.from_hub(settings.index_repo) log.info("Catalog: %d items %s", len(INDEX), INDEX.group_counts()) from drape.models import OpenClipEmbedder, SegformerClothesSegmenter from drape.pipeline import ShopTheLook embedder = OpenClipEmbedder(settings.embed_model, device="cpu") if INDEX.manifest.get("embed_model") and INDEX.manifest["embed_model"] != embedder.name: raise RuntimeError( f"Index was built with {INDEX.manifest['embed_model']} but the app loads {embedder.name}. " "Rebuild the index or set DRAPE_EMBED_MODEL to match." ) if embedder.dim != INDEX.dim: raise RuntimeError(f"Embedding size mismatch: model {embedder.dim} vs index {INDEX.dim}") segmenter = SegformerClothesSegmenter(settings.seg_model, device="cpu") ENGINE = ShopTheLook(embedder, segmenter, INDEX, max_side=settings.max_side) ENGINE._group_prototypes() # warm up log.info("Drape ready.") except Exception as exc: # keep the UI up and explain what is missing STARTUP_ERROR = f"{type(exc).__name__}: {exc}" log.error("Startup failed:\n%s", traceback.format_exc()) _load_engine() # --------------------------------------------------------------------------- # Helpers # --------------------------------------------------------------------------- def _gallery(matches: list[Match], prefix: str = "") -> list[tuple]: return [(INDEX.thumbnail(m.row), f"{prefix}{m.caption()}") for m in matches] def _require_engine(): if ENGINE is None: raise gr.Error( "Drape is not ready: " + (STARTUP_ERROR or "still loading, try again in a minute.") ) def _empty_slots(): out = [] for _ in range(MAX_SLOTS): out += [gr.update(visible=False), gr.update(value=""), gr.update(value=[])] return out # --------------------------------------------------------------------------- # Handlers # --------------------------------------------------------------------------- def on_analyze(image, k): _require_engine() if image is None: raise gr.Error("Upload a photo first, or press “Surprise me”.") try: result = ENGINE.analyze(image, k=int(k)) except Exception as exc: log.error(traceback.format_exc()) raise gr.Error(f"Could not analyse this photo: {exc}") from exc garments = result.garments[:MAX_SLOTS] if result.mode == "outfit": found = ", ".join(g.name.lower() for g in garments) status = f"**Found {len(garments)} piece{'s' if len(garments) != 1 else ''}:** {found}." else: status = ( f"**No person detected**, so I treated the photo as a single product " f"(looks like **{garments[0].name.lower()}**)." ) status += f" {result.timings['total_s']:.1f}s on CPU" slots = [] for i in range(MAX_SLOTS): if i < len(garments): g = garments[i] shelves = " / ".join(GROUP_LABELS[s].lower() for s in g.kind.search_groups) title = f"### {i + 1} · {g.name}\nsearched in {shelves}" slots += [gr.update(visible=True), gr.update(value=title), gr.update(value=_gallery(g.matches))] else: slots += [gr.update(visible=False), gr.update(value=""), gr.update(value=[])] state = { "garments": [ {"name": g.name, "groups": list(g.kind.search_groups), "emb": g.embedding.tolist()} for g in garments ], "look": result.look_embedding.tolist(), "present": result.present_groups, } choices = [f"{i + 1} · {g.name}" for i, g in enumerate(garments)] piece = gr.update(choices=choices, value=choices[0] if choices else None, interactive=bool(choices)) looks = _complete(state, "") return [result.overlay, status, state, piece, gr.update(value=[]), looks, *slots] def _complete(state, style_text): if not state: return gr.update(value=[]) groups = ENGINE.complete_look( np.asarray(state["look"], dtype=np.float32), state["present"], k_per_group=4, style_text=style_text ) items = [] for group, matches in groups.items(): items += _gallery(matches, prefix=f"{GROUP_LABELS[group]} · ") return gr.update(value=items) def on_complete(state, style_text): _require_engine() if not state: raise gr.Error("Analyse a photo first.") return _complete(state, style_text) def on_refine(state, piece, text, alpha, k): _require_engine() if not state or not piece: raise gr.Error("Analyse a photo first, then pick a piece to refine.") idx = int(piece.split("·")[0].strip()) - 1 g = state["garments"][idx] matches = ENGINE.refine( np.asarray(g["emb"], dtype=np.float32), tuple(g["groups"]), text, alpha=float(alpha), k=int(k) ) return gr.update(value=_gallery(matches)) def on_text_search(query, group_label, k): _require_engine() if not (query or "").strip(): raise gr.Error("Describe what you are looking for, e.g. “red satin slip dress”.") group = None if group_label == "Everything" else _label_to_group[group_label] return gr.update(value=_gallery(ENGINE.text_search(query, k=int(k), group=group))) def on_surprise(): _require_engine() rows = INDEX.rows_for(("dress", "top", "outerwear", "bottom", "shoes", "bag")) return INDEX.thumbnail(int(random.choice(rows))) _label_to_group = {GROUP_LABELS[g]: g for g in GROUPS} # --------------------------------------------------------------------------- # UI # --------------------------------------------------------------------------- THEME = gr.themes.Base( primary_hue=gr.themes.colors.orange, neutral_hue=gr.themes.colors.stone, font=[gr.themes.GoogleFont("Inter"), "ui-sans-serif", "system-ui", "sans-serif"], radius_size=gr.themes.sizes.radius_sm, ).set( button_primary_background_fill="#1c1917", button_primary_background_fill_hover="#44403c", button_primary_text_color="#fafaf9", button_primary_background_fill_dark="#fafaf9", button_primary_background_fill_hover_dark="#e7e5e4", button_primary_text_color_dark="#1c1917", ) CSS = """ @import url('https://fonts.googleapis.com/css2?family=Cormorant+Garamond:ital,wght@0,500;0,600;1,500&display=swap'); .gradio-container { max-width: 1180px !important; margin: 0 auto !important; } #drape-hero { padding: 28px 4px 8px; border-bottom: 1px solid var(--border-color-primary); margin-bottom: 8px; } #drape-hero h1 { font-family: 'Cormorant Garamond', Georgia, serif; font-size: 3.4rem; font-weight: 600; letter-spacing: 0.02em; margin: 0; line-height: 1; } #drape-hero h1 em { font-style: italic; font-weight: 500; color: #b45309; } #drape-hero p { margin: 10px 0 0; font-size: 1.05rem; opacity: 0.8; } .drape-muted { opacity: 0.6; font-size: 0.9em; } .drape-section h3, .drape-section h2 { font-family: 'Cormorant Garamond', Georgia, serif; font-size: 1.6rem; font-weight: 600; margin: 4px 0; } .drape-slot { border-top: 1px solid var(--border-color-primary); padding-top: 6px; } footer { opacity: 0.6; } """ def _hero_md() -> str: if INDEX is not None: stats = f"{len(INDEX):,} products · Marqo-FashionSigLIP embeddings · SegFormer garment parsing · runs on CPU" else: stats = "Catalog not loaded yet" return ( "
Upload any outfit photo. Drape finds each piece, matches it to real products, " "and suggests what completes the look.
" f"{stats}
Drape v{__version__} · built by Imene Guizani · " "catalog images from the public Polyvore dataset, used for research demonstration only.
" ) analyze_outputs = [overlay_out, status_md, state, piece_dd, refine_gal, complete_gal, *slot_components] analyze_btn.click(on_analyze, [image_in, k_slider], analyze_outputs) surprise_btn.click(on_surprise, None, image_in).then(on_analyze, [image_in, k_slider], analyze_outputs) refine_btn.click(on_refine, [state, piece_dd, refine_txt, alpha_slider, k_slider], refine_gal) refine_txt.submit(on_refine, [state, piece_dd, refine_txt, alpha_slider, k_slider], refine_gal) complete_btn.click(on_complete, [state, style_txt], complete_gal) style_txt.submit(on_complete, [state, style_txt], complete_gal) q_btn.click(on_text_search, [q_txt, q_group, q_k], q_gal) q_txt.submit(on_text_search, [q_txt, q_group, q_k], q_gal) return demo # Gradio 6 moved theme/css from Blocks() to launch(); support both. _LAUNCH_TAKES_STYLE = "css" in inspect.signature(gr.Blocks.launch).parameters def _blocks_style_kwargs() -> dict: return {} if _LAUNCH_TAKES_STYLE else {"theme": THEME, "css": CSS} def _launch_style_kwargs() -> dict: return {"theme": THEME, "css": CSS} if _LAUNCH_TAKES_STYLE else {} demo = build_ui() if __name__ == "__main__": demo.queue(default_concurrency_limit=2).launch(ssr_mode=False, **_launch_style_kwargs())