"""LightOnOCR-3 demo: drop a page, pick a model and a mode, watch the transcription stream in. The interface is one custom component (ui/): a sidebar for the document, the model, the mode and the examples, and a stage with the page on the left, every layout block drawn on it, and the text on the right, block by block, hover and click synced between the two. Gradio is the backend: the component sends custom events (open a file, change page, run, stop) and receives a JSON value, streamed while the model writes, so the boxes appear while the page is being read. Files: ocr.py (the variants and the two backends), grounding.py (the output format), render.py (PDF or image to page image), ui/ (the interface: template, CSS, JS). """ from __future__ import annotations import hashlib import io import time from pathlib import Path import gradio as gr from PIL import Image import ocr from grounding import parse_blocks, strip_thinking, to_markdown from render import DEFAULT_LONGEST_EDGE, page_count, render_page ROOT = Path(__file__).parent UI = ROOT / "ui" UPLOADS = Path(gr.utils.get_upload_folder()).resolve() # where the component's upload() puts files CACHE = UPLOADS / "lightonocr3" # page images for the viewer, served by Gradio like the uploads COLLECTION = "https://huggingface.co/collections/lightonai/lightonocr-3-6ac64536c01ee676d9bf46a7" PAPER = "https://arxiv.org/abs/2601.14251" BLOG = "https://huggingface.co/blog/lightonai/lightonocr-3" WEBSITE = "https://lighton.ai/" # Click a thumbnail to run that page with the current settings: (file, caption). EXAMPLES = [ ("examples/paper-page-3.png", "LightOnOCR paper, page 3"), ("examples/manual-figure.png", "Technical manual with a figure"), ("examples/annual-report-table.png", "Annual report, dense table"), ("examples/docs-table.png", "Documentation page"), ("examples/arxiv-two-columns.png", "Two-column paper"), ("examples/thesis-math.png", "Thesis page, equations"), ("examples/handwritten-math.png", "Handwritten notes, math"), ("examples/handwritten-diagrams.png", "Handwritten notes, diagrams"), ("examples/lecture-notes.png", "Lecture notes, highlighted"), ("examples/social-post.png", "Social media post"), ("examples/telegram.jpg", "1906 telegram, photo"), ("examples/library-card.jpg", "Library card, photo"), ] EXAMPLES = [e for e in EXAMPLES if (ROOT / e[0]).is_file()] gr.set_static_paths(paths=[ROOT / "examples"]) # thumbnails are plain tags def thumb(path: str) -> str: small = ROOT / "examples" / "thumbs" / (Path(path).stem + ".jpg") return f"gradio_api/file={small.relative_to(ROOT)}" if small.is_file() else f"gradio_api/file={path}" # ---- The component's value ----------------------------------------------------------------------------- def cache_image(image: Image.Image) -> str: """Save the page as a JPEG in Gradio's cache and return its URL. Sent once per page, not with every stream update.""" CACHE.mkdir(parents=True, exist_ok=True) buf = io.BytesIO() image.save(buf, "JPEG", quality=88) path = CACHE / (hashlib.sha1(buf.getvalue()).hexdigest()[:20] + ".jpg") if not path.exists(): path.write_bytes(buf.getvalue()) return f"gradio_api/file={path}" def page_name(doc: dict) -> str: stem = Path(doc.get("name") or doc.get("path") or "page").stem return f"{stem}-page-{doc['page']}" if doc.get("pages", 1) > 1 else stem def state_value(doc: dict | None = None, image: Image.Image | None = None, image_url: str | None = None, mode: str = "grounding", raw: str = "", status: str = "idle", message: str = "", model: str | None = None, example: int | None = None) -> dict: """What ui/app.js renders: the document, its page, the parsed blocks (grounding) and the Markdown, plus a status line.""" doc = doc or {} value = {"path": doc.get("path"), "name": doc.get("name"), "pages": doc.get("pages", 1), "page": doc.get("page", 1), "image": image_url, "width": image.width if image else None, "height": image.height if image else None, "mode": mode, "model": model, "raw": raw, "markdown": raw, "blocks": [], "status": status, "message": message, "filename": page_name(doc) if doc else "page", "example": example} if mode == "grounding": blocks = parse_blocks(raw) value["blocks"] = [b.to_dict() for b in blocks] value["markdown"] = to_markdown(blocks) return value EMPTY = state_value() def resolve(path: str | None) -> Path: """A path the component may ask for: something it uploaded, or an example. Anything else is refused.""" p = Path(str(path or "")).resolve() if not p.is_file() or not (p.is_relative_to(UPLOADS) or p.is_relative_to((ROOT / "examples").resolve())): raise gr.Error("Please upload the file again.") return p def open_doc(path: str | None, name: str | None, page: int, longest_edge: int) -> tuple[dict, Image.Image]: p = resolve(path) pages = page_count(p) page = min(max(int(page or 1), 1), pages) image = render_page(p, page, int(longest_edge or DEFAULT_LONGEST_EDGE)) return {"path": str(p), "name": name or p.name, "pages": pages, "page": page}, image def ready(doc: dict, image: Image.Image, mode: str, example: int | None = None) -> dict: where = f"page {doc['page']} of {doc['pages']} · " if doc["pages"] > 1 else "" return state_value(doc, image, cache_image(image), mode, status="ready", example=example, message=f"{doc['name']} · {where}{image.width}×{image.height} px · ready") # ---- Events from the component --------------------------------------------------------------------------- def data(evt: gr.EventData) -> dict: return evt._data if isinstance(evt._data, dict) else {} def on_open(evt: gr.EventData): d = data(evt) doc, image = open_doc(d.get("path"), d.get("name"), 1, d.get("longest_edge")) return ready(doc, image, d.get("mode") or "grounding") def on_page(evt: gr.EventData): d = data(evt) doc, image = open_doc(d.get("path"), None, d.get("page", 1), d.get("longest_edge")) return ready(doc, image, d.get("mode") or "grounding") def on_run(evt: gr.EventData): """Render the page, stream the model over it, and yield the value as the text comes in.""" d = data(evt) mode = d.get("mode") if d.get("mode") in ocr.MODES else "grounding" model_key = d.get("model") if d.get("model") in ocr.ENABLED else ocr.DEFAULT_VARIANT temperature = min(max(float(d.get("temperature", 0.2)), 0.0), 1.0) max_tokens = int(d.get("max_tokens", ocr.DEFAULT_MAX_TOKENS)) longest_edge = int(d.get("longest_edge", DEFAULT_LONGEST_EDGE)) example = d.get("example") if example is not None: example = int(example) if not 0 <= example < len(EXAMPLES): raise gr.Error("Unknown example.") path, name = str(ROOT / EXAMPLES[example][0]), EXAMPLES[example][1] doc, image = open_doc(path, name, 1, longest_edge) else: doc, image = open_doc(d.get("path"), d.get("name"), d.get("page", 1), longest_edge) variant = ocr.VARIANTS[model_key] shown = dict(doc=doc, image=image, image_url=cache_image(image), mode=mode, model=variant.name, example=example) size = f"{image.width}×{image.height} px" yield state_value(**shown, status="running", message=f"{variant.name} · {size} · waiting for the first token…") started = time.monotonic() first_token: float | None = None last_shown = 0.0 text = "" try: for update in ocr.stream(model_key, image, mode, temperature, max_tokens): now = time.monotonic() text = strip_thinking(update.text) if update.error: yield state_value(**shown, raw=text, status="error", message=update.error) return if text and first_token is None: first_token = now - started if update.done: elapsed = now - started value = state_value(**shown, raw=text, status="done") parts = [variant.name, ocr.backend_name(model_key), size] if mode == "grounding": parts.append(f"{len(value['blocks'])} block{'s' if len(value['blocks']) != 1 else ''}") if update.tokens: parts.append(f"{update.tokens:,} tokens") if first_token is not None and elapsed - first_token > 0.5: parts.append(f"{update.tokens / (elapsed - first_token):.0f} tok/s") parts.append(f"{elapsed:.1f} s") value["message"] = " · ".join(parts) yield value elif now - last_shown >= 0.3: # the component re-renders on every update: a few per second is plenty last_shown = now yield state_value(**shown, raw=text, status="running", message=f"{variant.name} · {size} · generating… {now - started:.0f} s") except gr.Error: raise # ZeroGPU quota messages and the like: Gradio shows them except Exception as e: yield state_value(**shown, raw=text, status="error", message=f"{type(e).__name__}: {e}") # ---- The app: one component ------------------------------------------------------------------------------ TEMPLATE = (UI / "app.html").read_text(encoding="utf-8") for key, text in {"__LOGO_LIGHT__": (ROOT / "static" / "lighton.svg").read_text(encoding="utf-8"), "__LOGO_DARK__": (ROOT / "static" / "lighton-dark.svg").read_text(encoding="utf-8"), "__COLLECTION__": COLLECTION, "__PAPER__": PAPER, "__BLOG__": BLOG, "__WEBSITE__": WEBSITE}.items(): TEMPLATE = TEMPLATE.replace(key, text) HEAD = """ """ # The page frame and the design tokens. `.dark` is the class Gradio puts on ; ui/app.js toggles it too. CSS = """ body { --lo-font: Inter, ui-sans-serif, system-ui, -apple-system, "Segoe UI", Roboto, Helvetica, Arial, sans-serif; --lo-mono: ui-monospace, SFMono-Regular, Menlo, Consolas, monospace; --lo-bg: #f3f5f8; --lo-surface: #ffffff; --lo-surface-2: #f6f8fa; --lo-ink: #0f172a; --lo-muted: #64748b; --lo-line: #e5e9f0; --lo-line-strong: #cbd3df; --lo-accent: #0a8db0; --lo-accent-soft: rgba(10, 141, 176, .1); --lo-danger: #dc2626; --lo-ok: #16a34a; background: var(--lo-bg) !important; } body.dark { --lo-bg: #0b0f15; --lo-surface: #121820; --lo-surface-2: #182029; --lo-ink: #e6ebf2; --lo-muted: #8e9aab; --lo-line: #232d3a; --lo-line-strong: #34404f; --lo-accent: #2eb5d1; --lo-accent-soft: rgba(46, 181, 209, .14); --lo-danger: #f87171; --lo-ok: #4ade80; } .dark .logo-light { display: none !important } .dark .logo-dark { display: inline !important } .gradio-container { max-width: none !important; padding: 0 !important; background: transparent !important } .gradio-container .main.app { padding: 0 !important } #lo-app { border: 0 !important; border-radius: 0 !important } footer { display: none !important } @keyframes lo-spin { to { transform: rotate(360deg) } } @keyframes lo-live { 50% { box-shadow: 0 0 0 3px color-mix(in srgb, var(--c) 45%, transparent) } } @keyframes lo-blink { 50% { opacity: 0 } } @keyframes lo-slide { 0% { transform: translateX(-100%) } 100% { transform: translateX(340%) } } """ with gr.Blocks(title="LightOnOCR-3", fill_width=True) as demo: app = gr.HTML( EMPTY, html_template=TEMPLATE, css_template=(UI / "app.css").read_text(encoding="utf-8"), js_on_load=(UI / "app.js").read_text(encoding="utf-8"), head=HEAD, apply_default_css=False, padding=False, elem_id="lo-app", # props read by ui/app.js models=[{"key": k, "name": ocr.VARIANTS[k].name, "summary": ocr.VARIANTS[k].summary} for k in ocr.ENABLED], examples=[{"thumb": thumb(p), "caption": c} for p, c in EXAMPLES], defaults={"model": ocr.DEFAULT_VARIANT, "mode": "grounding", "temperature": 0.2, "max_tokens": ocr.DEFAULT_MAX_TOKENS, "longest_edge": DEFAULT_LONGEST_EDGE}, ) app.lo_open(on_open, None, [app], show_progress="hidden") app.lo_page(on_page, None, [app], show_progress="hidden") running = app.lo_run(on_run, None, [app], show_progress="hidden") app.lo_stop(lambda: None, None, None, cancels=[running], show_progress="hidden") if __name__ == "__main__": demo.queue().launch(css=CSS, ssr_mode=False, footer_links=[])