[Task] image-to-text EP coverage investigation

Overview

Image-to-text models generate textual descriptions or transcriptions from images — covering OCR (TrOCR, manga-ocr), document understanding (Donut, Nougat), and image captioning (BLIP, ViT-GPT2). 9 of 11 models pass on at least one EP.

This issue focuses on the 2 models with severe EP gaps — donut and nougat currently pass only on VitisAI and fail on both QNN and OV. These are encoder-decoder architectures with custom preprocessing, and the failure pattern is distinct from the rest of the task's models.

Agent Scenarios

• Document parsing agent: convert scanned PDFs, receipts, or forms into structured JSON using Donut's end-to-end document understanding, without OCR preprocessing
• Scientific paper extraction agent: use Nougat to convert academic PDFs (including LaTeX math) into structured markdown for downstream summarization or RAG ingestion
• OCR pipeline agent: transcribe printed or handwritten text from images (TrOCR, manga-ocr) as a preprocessing step for document search and indexing

EP Coverage Status

Reference: microsoft/trocr-large-printed, microsoft/trocr-base-printed, nlpconnect/vit-gpt2-image-captioning, microsoft/trocr-base-handwritten, kha-white/manga-ocr-base, microsoft/trocr-large-handwritten, Salesforce/blip-image-captioning-base already pass all 3 EPs.

Acceptance Criteria

• naver-clova-ix/donut-base passes all 3 EPs (investigate QNN + OV failure)
• facebook/nougat-base passes all 3 EPs (investigate QNN + OV failure)