import spaces # ZeroGPU: must come before torch / CUDA-touching imports import json import re import time import gradio as gr import torch from PIL import Image from transformers import AutoModelForImageTextToText, AutoProcessor MODEL_ID = "typhoon-ai/typhoon-ocr2-2b" # The model card allows up to 16,384 new tokens, but a page that long decodes # for ~9 minutes at the ~30 tok/s measured on this Space — far past any ZeroGPU # per-call reservation. It also documents a repetition-loop failure mode # ("repeating text until it reaches the token limit ... most on dense # multi-column pages"), where truncation is the graceful outcome: parse_elements() # below keeps every complete element. The densest example page emits ~1,230 # tokens, so 2,048 tokens gives ~65% headroom over real pages while keeping the # worst case inside the GPU reservation (~71 s at 30 tok/s). Questions want a # short answer and get a tighter cap so their reservation can stay small. MAX_NEW_TOKENS = 2048 MAX_NEW_TOKENS_QA = 1024 # --------------------------------------------------------------------------- # Model + processor, loaded at module scope and moved to CUDA eagerly. # (The model card uses device_map="auto"; on ZeroGPU .to("cuda") is the rule.) # --------------------------------------------------------------------------- model = AutoModelForImageTextToText.from_pretrained(MODEL_ID, dtype=torch.bfloat16).to("cuda") processor = AutoProcessor.from_pretrained(MODEL_ID) # --------------------------------------------------------------------------- # Task modes # --------------------------------------------------------------------------- TASK_PARSE = "Full-page parsing (layout -> Markdown)" TASK_KIE = "Key information extraction (JSON)" TASK_QA = "Ask a question about the document" DEFAULT_TASK = TASK_PARSE # --------------------------------------------------------------------------- # Prompts and helpers below are copied from the model card (Usage section). # --------------------------------------------------------------------------- PARSING_PROMPT = """- Extract layout information from the provided PDF image. - For each layout element, output its bbox, category, and the text content within the bbox. - Bbox format: [x1, y1, x2, y2]. - Allowed layout categories: ['header', 'title', 'text', 'figure', 'table', 'formula', 'figure_caption', 'table_caption', 'formula_caption', 'figure_footnote', 'table_footnote', 'page_footnote', 'footer']. - Text extraction and formatting: 1) For 'figure', the text field must be an empty string. 2) For 'formula', format text as LaTeX. 3) For 'table', format text as HTML. 4) For all other categories (e.g., text, title), format text as Markdown. - The output text must be exactly the original text from the image, with no translation or rewriting. - Sort all layout elements in human reading order. - Final output must be a single JSON object.""" KIE_PROMPT = ( "Perform open-schema key information extraction (WYSIWYG) on this image: " "one JSON object, printed labels as keys, values copied exactly, \"\" for blank " "fields, arrays for repeated rows. Return only JSON." ) # Prompts from the model card's example gallery (verbatim, in Thai). SNACK_PROMPT = "จากรูปภาพกรุณาดึงข้อมูล เปอร์เซ็นต์ของส่วนประกอบแต่ละอย่าง จากรูปภาพเป็นรูปแบบ JSON เป็นภาษาไทย เฉพาะอันที่มีเปอร์เซ็นต์\n\nโดยมี key ตามนี้\nมะพร้าว\nน้ำตาล\nเกลือสมุทร\nกลิ่นคาราเมล\nตอบเฉพาะข้อความตามภาพตรง ๆ อย่างกระชับ โดยไม่ใส่คำอธิบายเพิ่มเติม" PM25_PROMPT = "ตามข้อมูลในภาพกรุณาจับคู่รายการระดับความเข้มข้น PM2.5 (จำนวนจำกัด) เกี่ยวข้องดังต่อไปนี้: 'ดีมาก', 'ดี', 'ปานกลาง', 'เริ่มมีผลกระทบต่อสุขภาพ' และ 'มีผลกระทบต่อสุขภาพ' กับค่าที่จะถูกจับคู่ได้แก่ '7 จังหวัด', '0 จังหวัด', '28 จังหวัด', '16 จังหวัด' และ '25 จังหวัด' เลือกรายละเอียดจากเนื้อหาในเอกสารแล้วให้คำตอบกลับมาในรูปแบบ JSON โดยตรง:\nตอบเฉพาะข้อความตามภาพตรง ๆ อย่างกระชับ โดยไม่ใส่คำอธิบายเพิ่มเติม" HANDWRITTEN_PROMPT = "ข้อที่ 4 ในภาพนี้ เป็นข้อความที่เขียนด้วยมือว่าอะไรบ้าง ตอบเฉพาะข้อความตามภาพตรง ๆ อย่างกระชับ โดยไม่ใส่คำอธิบายเพิ่มเติม" REGION_PROMPT = "แบ่งความยาวและความสูงของรูปภาพออกเป็น 1000 ส่วน แล้วช่วยดึงข้อความที่อยู่ในพิกัด [786, 615, 931, 675] ของรูปภาพออกมาให้หน่อย ตอบเฉพาะข้อความตามภาพตรง ๆ อย่างกระชับ โดยไม่ใส่คำอธิบายเพิ่มเติม" def parse_elements(raw: str) -> list: """Parse the model output: a JSON list of elements inside a ```json fence.""" m = re.search(r"```json\n(.*?)(?:\n```|$)", raw, re.DOTALL) text = (m.group(1) if m else raw).strip() try: return json.loads(text) except json.JSONDecodeError: # Output stopped at the token limit: drop the last, incomplete element. cut = text.rfind('{"bbox":') return json.loads(text[:cut].rstrip().rstrip(",") + "]") def to_markdown(elements: list, keep_header_footer: bool = False) -> str: skip = set() if keep_header_footer else {"header", "footer", "page_footnote"} return "\n\n".join( e["text"] for e in elements if e.get("text") and e["category"] not in skip ) def upscale_min_long(image: Image.Image, min_long: int = 1600) -> Image.Image: """Enlarge the image so that its long side is at least `min_long` pixels, keeping the aspect ratio. Images already large enough are unchanged.""" scale = min_long / max(image.size) if scale <= 1: return image return image.resize( (round(image.width * scale), round(image.height * scale)), Image.LANCZOS ) def _resolve_prompt(task: str, question: str) -> str: if task == TASK_PARSE: return PARSING_PROMPT if task == TASK_KIE: return KIE_PROMPT return question.strip() def _duration(image, task=TASK_PARSE, question="", *args, **kwargs): """Per-task GPU reservation (seconds). Sized from durations measured on this Space against the model authors' own example pages: full-page parsing 41 s, key-info extraction 17 s, question answering 2 s. Each value covers that mode's worst case at the measured ~30 tok/s (2,048 tokens -> ~71 s; 1,024 -> ~37 s) so a long decodes finishes inside its reservation instead of aborting, while the shorter modes keep a small reservation for queue priority and quota. Extras are swallowed: Gradio injects `progress=` positionally. """ if task == TASK_QA: return 45 return 90 @spaces.GPU(duration=_duration) def run_ocr( image, task: str = DEFAULT_TASK, question: str = "", keep_header_footer: bool = False, progress=gr.Progress(track_tqdm=True), ): """Run Typhoon OCR V2 2B on a document image. Args: image: the document page / photo to read. task: one of "Full-page parsing (layout -> Markdown)", "Key information extraction (JSON)" or "Ask a question about the document". question: the question to ask (used only in question mode). keep_header_footer: in parsing mode, also keep header/footer/page-footnote elements in the Markdown output. Returns: Tuple of (markdown, raw_output, seconds): the formatted result, the raw model output, and the wall-clock seconds the generation took. """ t0 = time.perf_counter() prompt = _resolve_prompt(task, question) if not prompt: raise gr.Error("Enter a question for the document, or pick another task.") pil_image = Image.open(image).convert("RGB") if isinstance(image, str) else image if pil_image is None: raise gr.Error("Upload a document image first.") pil_image = upscale_min_long(pil_image.convert("RGB")) messages = [{"role": "user", "content": [ {"type": "image", "image": pil_image}, {"type": "text", "text": prompt}, ]}] inputs = processor.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt", enable_thinking=False, processor_kwargs={"images_kwargs": {"min_pixels": 2048, "max_pixels": 16777216}}, ).to(model.device) max_new = MAX_NEW_TOKENS_QA if task == TASK_QA else MAX_NEW_TOKENS with torch.inference_mode(): out = model.generate(**inputs, max_new_tokens=max_new, do_sample=False) raw = processor.decode( out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True ) seconds = round(time.perf_counter() - t0, 1) if task == TASK_PARSE: try: elements = parse_elements(raw) md = to_markdown(elements, keep_header_footer=keep_header_footer) if not md.strip(): md = "_No text elements found on the page._" except (json.JSONDecodeError, KeyError, IndexError, TypeError): md = "_Could not parse the model output as layout JSON — see the raw output._" else: md = raw.strip() return md, raw, seconds # --------------------------------------------------------------------------- # Example gallery (images and prompts from the model card) # --------------------------------------------------------------------------- EXAMPLES = [ ("assets/parsing_exam_input.jpg", TASK_PARSE, "", "Full-page parsing — Thai exam paper"), ("assets/parsing_book_input.jpg", TASK_PARSE, "", "Full-page parsing — Thai training handout"), ("assets/kie_receipt.jpg", TASK_KIE, "", "Key info — restaurant bill"), ("assets/kie_infographic.jpg", TASK_KIE, "", "Key info — rice-seed infographic"), ("assets/kie_snack.jpg", TASK_QA, SNACK_PROMPT, "Ingredients % — snack bag"), ("assets/kim_pm25.jpg", TASK_QA, PM25_PROMPT, "KIE mapping — PM2.5 advisory table"), ("assets/handwritten.jpg", TASK_QA, HANDWRITTEN_PROMPT, "Handwriting — lecture notes"), ("assets/region_sign.jpg", TASK_QA, REGION_PROMPT, "Region reading — street sign"), ] CSS = """ #col-container { max-width: 1200px; margin: 0 auto; } .dark .gradio-container { color: var(--body-text-color); } """ with gr.Blocks(title="Typhoon OCR V2 2B") as demo: with gr.Column(elem_id="col-container"): gr.Markdown( """# Typhoon OCR V2 2B Thai/English OCR and document parsing: one image in, structured text out — full-page layout parsing (bboxes → Markdown/HTML/LaTeX), open-schema key-information extraction as JSON, or free-form questions about the document. Model: [typhoon-ai/typhoon-ocr2-2b](https://huggingface.co/typhoon-ai/typhoon-ocr2-2b) · 2.2B params · Apache-2.0 weights (use subject to OpenTyphoon Terms)""" ) with gr.Row(): with gr.Column(scale=1): image_in = gr.Image( type="pil", label="Document image", sources=["upload", "clipboard"], height=420, ) task = gr.Radio( [TASK_PARSE, TASK_KIE, TASK_QA], value=DEFAULT_TASK, label="Task", info=( "Parsing uses the model's fixed layout prompt; key-info extraction " "pulls all fields as JSON; question mode asks anything." ), ) question = gr.Textbox( label="Question (question mode only)", placeholder="e.g. ข้อที่ 4 ในภาพนี้เขียนว่าอะไรบ้าง?", lines=2, ) with gr.Accordion("Advanced", open=False): keep_header_footer = gr.Checkbox( value=False, label="Keep header/footer/page-footnote in Markdown", info="Parsing mode only. The model card's evaluation drops them.", ) run_btn = gr.Button("Run OCR", variant="primary") with gr.Column(scale=1): md_out = gr.Markdown(label="Result") with gr.Accordion("Raw model output", open=False): raw_out = gr.Code(language="markdown", label="Raw output") time_out = gr.Number(label="Seconds", precision=1, interactive=False) gr.Examples( examples=[[e[0], e[1], e[2]] for e in EXAMPLES], inputs=[image_in, task, question], outputs=[md_out, raw_out, time_out], fn=run_ocr, example_labels=[e[3] for e in EXAMPLES], cache_examples=True, cache_mode="lazy", api_visibility="undocumented", ) run_btn.click( run_ocr, inputs=[image_in, task, question, keep_header_footer], outputs=[md_out, raw_out, time_out], api_name="run_ocr", ) demo.launch(mcp_server=True, theme=gr.themes.Citrus(), css=CSS)