Implementation and measured chapter results.
MAGI extraction with persistent character IDs, essential-only text dialogue analysis and character-name linking.
No scene segmentation or model reading-order correction. Text inference uses think:false.
Start from an image folder with python run.py run <folder> --story <story> --chapter <chapter>.
Continue with python run.py resume --run-dir <run-directory>. Completed steps and targets are reused.
Legacy MAGI-only extraction can be reanalyzed only with an explicit config disabling layout. It cannot be reused as Koharu extraction; start from the original images for a new v5 run.
New runs use three presentation groups with independent resumable checkpoints:
layout → extract → bubble_geometry (the extract CLI command completes this group).normalize → scan → analyze.export → translate.Normalize filters whole boxes of unrelated noise with versioned rules. Translate produces separate natural and localized Vietnamese JSON/reviews, with resumable checkpoints and manual sentence locks. analyze resolves text and binds scanned names after each batch, with checkpoints for text, panel decisions, bank updates and participant memory. Saved v2/v3/v4 runs retain their original flow and analysis contracts.
Koharu is mandatory/default for new runs. Locally, install requirements-visual-layout.txt in a separate Transformers 5 environment and set layout.python to that interpreter in your config. Use that interpreter to run scripts/setup_visual_models.py --only layout. Missing/corrupt weights or an incompatible worker fail before creating a run or touching its bank. MAGI remains in its Transformers 4 environment. New defaults: character matching 0.75, character detection 0.35. Use a fresh bank; legacy MAGI geometry IDs are not imported into Koharu runs.
On Runpod, manga setup prepares both independently signed environments on the persistent runtime volume; manga prefetch prepares the pinned Koharu weights plus MAGI/NER/Ollama caches; manga doctor --inference checks both workers and assets. The managed config selects the actual visual interpreter and model path. The Docker image remains code-only. These source changes still require a code sync or a new image before they take effect on a pod.
bubble_geometry is heuristic visual evidence, not speech classification. New runs save speech_style_policy: semantic-speech-v1; analyze produces speech_style (normal, shout, whisper, thought, caption, unknown), evidence, source and review status. Those fields plus bubble/source-box provenance reach export and translation requests. Unknown/conflicting styles need review; nonessential text is explicitly not_analyzed. Merged texts retain each source's bubble metadata without inventing aggregate geometry.
Inpaint is disabled by default and was not part of this refactor's validation. If explicitly enabled after preparing AOT assets, --until inpaint runs its prerequisites and reconstruction branch. Inpaint does not render translated text. Koharu layout failures block extraction; AOT failures leave inpaint partial while analysis can continue.
New runs save speaker_policy: individual-v1: speakers are supported character IDs, others or narrator. Narrator requires narration; speaker groups is suspended. Saved v3 runs without this field retain their previous contract. Speaker policy and compatibility.
Use python run.py translate --run-dir <run-directory> to translate a committed legacy export directly. This does not refilter or rewrite the source run. Vietnamese translation, glossary, manual edits and future graph context.
The run's review.md shows per-step runtime and the recorded total across retries. Legacy steps without timing remain unknown. Regression tests and the bilingual chapter pilot.
Open <run-directory>/review.md to check compact tables for each step and links to images. Reviews are generated automatically; use python run.py review --run-dir <run-directory> to create them for an existing run without model inference.
scripts/experiments/ and docs/archive/.Use python -B -m unittest discover -s tests -v for the offline regression suite.
Implementation and measured chapter results.
MAGI extraction with persistent character IDs, essential-only text dialogue analysis and character-name linking.
No scene segmentation or model reading-order correction. Text inference uses think:false.
Start from an image folder with python run.py run <folder> --story <story> --chapter <chapter>.
Continue with python run.py resume --run-dir <run-directory>. Completed steps and targets are reused.
Legacy MAGI-only extraction can be reanalyzed only with an explicit config disabling layout. It cannot be reused as Koharu extraction; start from the original images for a new v5 run.
New runs use three presentation groups with independent resumable checkpoints:
layout → extract → bubble_geometry (the extract CLI command completes this group).normalize → scan → analyze.export → translate.Normalize filters whole boxes of unrelated noise with versioned rules. Translate produces separate natural and localized Vietnamese JSON/reviews, with resumable checkpoints and manual sentence locks. analyze resolves text and binds scanned names after each batch, with checkpoints for text, panel decisions, bank updates and participant memory. Saved v2/v3/v4 runs retain their original flow and analysis contracts.
Koharu is mandatory/default for new runs. Locally, install requirements-visual-layout.txt in a separate Transformers 5 environment and set layout.python to that interpreter in your config. Use that interpreter to run scripts/setup_visual_models.py --only layout. Missing/corrupt weights or an incompatible worker fail before creating a run or touching its bank. MAGI remains in its Transformers 4 environment. New defaults: character matching 0.75, character detection 0.35. Use a fresh bank; legacy MAGI geometry IDs are not imported into Koharu runs.
On Runpod, manga setup prepares both independently signed environments on the persistent runtime volume; manga prefetch prepares the pinned Koharu weights plus MAGI/NER/Ollama caches; manga doctor --inference checks both workers and assets. The managed config selects the actual visual interpreter and model path. The Docker image remains code-only. These source changes still require a code sync or a new image before they take effect on a pod.
bubble_geometry is heuristic visual evidence, not speech classification. New runs save speech_style_policy: semantic-speech-v1; analyze produces speech_style (normal, shout, whisper, thought, caption, unknown), evidence, source and review status. Those fields plus bubble/source-box provenance reach export and translation requests. Unknown/conflicting styles need review; nonessential text is explicitly not_analyzed. Merged texts retain each source's bubble metadata without inventing aggregate geometry.
Inpaint is disabled by default and was not part of this refactor's validation. If explicitly enabled after preparing AOT assets, --until inpaint runs its prerequisites and reconstruction branch. Inpaint does not render translated text. Koharu layout failures block extraction; AOT failures leave inpaint partial while analysis can continue.
New runs save speaker_policy: individual-v1: speakers are supported character IDs, others or narrator. Narrator requires narration; speaker groups is suspended. Saved v3 runs without this field retain their previous contract. Speaker policy and compatibility.
Use python run.py translate --run-dir <run-directory> to translate a committed legacy export directly. This does not refilter or rewrite the source run. Vietnamese translation, glossary, manual edits and future graph context.
The run's review.md shows per-step runtime and the recorded total across retries. Legacy steps without timing remain unknown. Regression tests and the bilingual chapter pilot.
Open <run-directory>/review.md to check compact tables for each step and links to images. Reviews are generated automatically; use python run.py review --run-dir <run-directory> to create them for an existing run without model inference.
scripts/experiments/ and docs/archive/.Use python -B -m unittest discover -s tests -v for the offline regression suite.