session log 2026 04 14

Session Log — 2026-04-14

Summary

Three main tracks completed: mecab_furigana_rs integration, Umi-OCR evaluation, and PaddleOCR-VL-For-Manga GGUF evaluation. Plus analysis of replacing Ollama with direct llama.cpp.

---

1. Integrated mecab_furigana_rs into lenzu

Goal: Replace 704 lines of inline MeCab logic in furigana.rs with the standalone mecab-furigana-rs crate (v0.1.0).

Changes:

• lenzu/Cargo.toml — added mecab-furigana-rs = "0.1.0"
• lenzu/src/furigana.rs — rewritten from 704 lines to ~130 lines (thin wrapper)
  • Removed #[cfg(target_os = "linux")] gate — crate handles cross-platform gracefully
  • annotate() and compare_and_maybe_overwrite() signatures preserved
  • All 4 tests pass
• README.md — added mecab-furigana-rs to Related Crates section
• lenzu/README.md — updated tech deps table to link the crate instead of raw MeCab

---

2. Umi-OCR Evaluation (PaddleOCR-based)

Goal: Test Umi-OCR via Docker HTTP API against 3 standard test PNGs.

Prototype: prototypes/umi-ocr-eval

Results: 0/3 PASS (pre-rescale oversized images — yokogaki 711×389, tategaki/tegaki 2760×1504)

Image	Orientation	Expected	Got	Score	Time (ms)
yokogaki	Horizontal	データを正確に読み取る	デー々を正確に読み取る	0.961	1084
tategaki	Vertical	『言語モデルのテスト』	き転でげんの テスイ	0.554, 0.253	2360
tegaki	Calligraphy	手書きの文字サンプル	チ書きの丈字サンプル	0.903	2116

Verdict: Not viable for manga OCR. Vertical text completely broken (maintainer acknowledges this in issue #434). Even horizontal text has character-level errors. Images were rescaled on 2026-04-15 — re-evaluation with smaller inputs pending.

---

3. PaddleOCR-VL-For-Manga GGUF Evaluation

Goal: Test PaddleOCR-VL-For-Manga via llama.cpp server against the same 3 PNGs.

Prototype: prototypes/paddleocr-vl-manga

Results: 2/3 PASS (CPU-only, no CUDA toolkit — BF16 on Xeon E5-2670v3; pre-rescale images)

Image	Orientation	Expected	Got	Tokens	Time (ms)
yokogaki	Horizontal	データを正確に読み取る	データを正確に読み取る	13	4645
tategaki	Vertical	『言語モデルのテスト』	今回は「言語モデルの テスト」 『言語モデルの テスト』	23	107418
tegaki	Calligraphy	手書きの文字サンプル	手書きの文字サンプル	11	104009

Verdict: Significantly better than Umi-OCR. Yokogaki and tegaki exact match. Tategaki has correct text present but with hallucinated prefix — potentially fixable with max_tokens tuning or post-processing. CPU times are slow (100+ seconds) due to BF16 without CUDA — expect ~1-5s with GPU. Images were rescaled on 2026-04-15 — re-evaluation with smaller inputs pending.

Why llama-server, not Ollama?

Ollama can load plain GGUF text models but vision models need a multimodal projector (--mmproj), which Ollama's Modelfile has no keyword for. The official MedAIBase/PaddleOCR-VL:0.9b on Ollama was published without the projector — text completion only, no image input.

---

4. HuggingFace Model Links Cataloged

Documented in docs/model-evaluation.md under "To evaluate: alternative OCR models":

Model	Link	Type
manga-ocr (original)	https://huggingface.co/mayocream/manga-ocr/tree/main	PyTorch weights
manga-ocr ONNX full	https://huggingface.co/xingliao/manga-ocr-onnx-full/tree/main	ONNX export
PaddleOCR-VL-For-Manga (GGUF)	https://huggingface.co/adambarbato/PaddleOCR-VL-For-Manga-GGUF/tree/main	GGUF quantized
PaddleOCR-VL-For-Manga (base)	https://huggingface.co/jzhang533/PaddleOCR-VL-For-Manga/tree/main	Original weights
PaddleOCRv5 Det For Manga	https://huggingface.co/bluolightning/PaddleOCRv5-Server-Det-For-Manga/tree/main	Detection model

Recommended upgrade path (from external advice):

1. Quick fix: xingliao/manga-ocr-onnx-full — quantized variant, keep ort backend
2. Best speed/accuracy: adambarbato/PaddleOCR-VL-For-Manga-GGUF — llama-cpp-rs backend
3. Lightweight: bluolightning/PaddleOCRv5-Server-Det-For-Manga — PaddleOCR in Rust

---

5. Ollama-to-llama.cpp Migration Analysis

Full analysis at docs/planning-ollama-to-llamacpp.md.

Key findings:

• lenzu already uses OpenAI-compatible /v1/chat/completions — same API as llama-server
• Streaming, logprobs, vision payloads all compatible
• Main challenge: model switching (Ollama does it transparently; llama-server needs multi-port or model reload strategy)
• Recommended path: llama-cpp-2 Rust crate for in-process inference (no server overhead)
• VRAM budget: PaddleOCR-VL (1.8 GB) + qwen2.5:3b (1.8 GB) fits in 8 GB with room to spare

---

6. Infrastructure Changes

scripts/setup.sh

• Added --skip-paddleocr flag
• Added nvidia-cuda-toolkit to apt dependencies
• Added PaddleOCR-VL GGUF download section (~1.8 GB, uses curl with progress)

.gitignore

• Added prototypes/paddleocr-vl-manga/model/ to prevent committing GGUF files

---

7. Test Image Size Concern (resolved 2026-04-15)

The tategaki and tegaki PNGs were 2760x1504 (4.6-4.7 MB) — full compositor screenshots, not realistic manga bubble crops. This was fixed on 2026-04-15: images rescaled to manga-bubble-realistic sizes (yokogaki 360×197, tategaki 480×262, tegaki 480×262, sample-texts 640×349). With the rescaled images:

• manga-ocr-rs: 3/3 correct text (tategaki bracket-style 「」vs『』 only diff)
• DBNet+manga-ocr pipeline: 3/3 exact match with high confidence (95-99% OCR)
• Inference time per crop: ~1.0-1.5 s (was 25-40 s for oversized images)

---

Files Changed

File	Change
lenzu/Cargo.toml	Added mecab-furigana-rs = "0.1.0"
lenzu/src/furigana.rs	Rewritten as thin wrapper (704 → ~130 lines)
README.md	Added mecab-furigana-rs to Related Crates
lenzu/README.md	Updated tech deps table
.gitignore	Added prototypes/paddleocr-vl-manga/model/
scripts/setup.sh	Added CUDA toolkit, PaddleOCR-VL download, --skip-paddleocr
docs/model-evaluation.md	Added HuggingFace model links and upgrade path
docs/planning-ollama-to-llamacpp.md	New — full Ollama replacement analysis
prototypes/umi-ocr-eval	New — Umi-OCR Docker prototype + results
prototypes/paddleocr-vl-manga	New — PaddleOCR-VL GGUF prototype + results