mule-image

dtoro/mule-image

Fork 0

Commit Graph

Author	SHA1	Message	Date
dtoro	94c07b1d0d	feat: upgrade to SigLIP2 ViT-B/16 for semantic search Replace OpenCLIP ViT-B/32 (512-d, ~78% recall) with SigLIP2 ViT-B/16 (768-d, ~84% recall) as the default embedding model for significantly better image-text retrieval quality. - New SigLIP2Embedder class with 384px input and SigLIP normalization - ONNX export pipeline for SigLIP2 visual + textual encoders - Migration 0010: resize embeddings.vector from 512 to 768 dimensions - Config-driven model selection: "siglip2_vitb16" (default) or "openclip_vitb32" (legacy) — both models can coexist - Content classifier follows the configured embedder family - Existing embeddings cleared on migration; vision backfill regenerates Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	2026-04-12 22:09:16 +02:00
dtoro	fa9b21856f	feat: replace face pipeline with InsightFace, add content classifier Face detection/recognition: - Replace YuNet + SFace with InsightFace buffalo_l (RetinaFace + ArcFace) - 512-d ArcFace embeddings (was 128-d SFace), migration 0006 resizes column - Remove YOLO person-bbox workaround — RetinaFace is accurate enough - Detection threshold 0.65 cleanly separates real faces (0.72+) from false positives on dogs/paintings (0.56-0.61) Content-type classification: - CLIP zero-shot classifier using native PyTorch text encoder + ONNX image encoder for high-quality text-image similarity - Categories: photograph, screenshot, document, receipt, meme, artwork - Writes Tag(kind=content_type) per photo via photo_tags - Margin-based confidence: top-1 vs top-2 score difference - New ClassifierSettings in config (enabled, min_confidence) - Wired into vision_fanout pipeline Tested: 6 real faces from 4 photos (zero false positives), 11/13 photos classified (8 photograph, 2 artwork, 1 meme). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	2026-04-10 13:49:02 +02:00
dtoro	9282a5c734	feat: add vision pipeline scaffolding with ONNX backend Introduce the app/services/vision/ module with ABC interfaces, ONNX Runtime backend, model registry, and per-task implementations: - OpenCLIP ViT-B/32 embedder (image + text, 512-d) - RapidOCR engine (PP-OCRv4 via ONNX, no PaddlePaddle) - YOLOv8n object detector (raw ONNX, no ultralytics runtime) - YuNet + SFace face processor (Apache 2.0, opencv_zoo, 128-d) - DBSCAN face clustering helper Add VisionSettings to config (mulita.yml + Pydantic), bootstrap_models.py for first-boot weight downloads, models_data Docker volume, and ROCm backend stub for future GPU acceleration. No Celery tasks wired yet — models load but nothing invokes them. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	2026-04-10 09:00:06 +02:00

Author

SHA1

Message

Date

dtoro

94c07b1d0d

feat: upgrade to SigLIP2 ViT-B/16 for semantic search

Replace OpenCLIP ViT-B/32 (512-d, ~78% recall) with SigLIP2 ViT-B/16
(768-d, ~84% recall) as the default embedding model for significantly
better image-text retrieval quality.

- New SigLIP2Embedder class with 384px input and SigLIP normalization
- ONNX export pipeline for SigLIP2 visual + textual encoders
- Migration 0010: resize embeddings.vector from 512 to 768 dimensions
- Config-driven model selection: "siglip2_vitb16" (default) or
  "openclip_vitb32" (legacy) — both models can coexist
- Content classifier follows the configured embedder family
- Existing embeddings cleared on migration; vision backfill regenerates

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-04-12 22:09:16 +02:00

dtoro

fa9b21856f

feat: replace face pipeline with InsightFace, add content classifier

Face detection/recognition:
- Replace YuNet + SFace with InsightFace buffalo_l (RetinaFace + ArcFace)
- 512-d ArcFace embeddings (was 128-d SFace), migration 0006 resizes column
- Remove YOLO person-bbox workaround — RetinaFace is accurate enough
- Detection threshold 0.65 cleanly separates real faces (0.72+) from
  false positives on dogs/paintings (0.56-0.61)

Content-type classification:
- CLIP zero-shot classifier using native PyTorch text encoder + ONNX
  image encoder for high-quality text-image similarity
- Categories: photograph, screenshot, document, receipt, meme, artwork
- Writes Tag(kind=content_type) per photo via photo_tags
- Margin-based confidence: top-1 vs top-2 score difference
- New ClassifierSettings in config (enabled, min_confidence)
- Wired into vision_fanout pipeline

Tested: 6 real faces from 4 photos (zero false positives), 11/13 photos
classified (8 photograph, 2 artwork, 1 meme).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-04-10 13:49:02 +02:00

dtoro

9282a5c734

feat: add vision pipeline scaffolding with ONNX backend

Introduce the app/services/vision/ module with ABC interfaces, ONNX
Runtime backend, model registry, and per-task implementations:
- OpenCLIP ViT-B/32 embedder (image + text, 512-d)
- RapidOCR engine (PP-OCRv4 via ONNX, no PaddlePaddle)
- YOLOv8n object detector (raw ONNX, no ultralytics runtime)
- YuNet + SFace face processor (Apache 2.0, opencv_zoo, 128-d)
- DBSCAN face clustering helper

Add VisionSettings to config (mulita.yml + Pydantic), bootstrap_models.py
for first-boot weight downloads, models_data Docker volume, and ROCm
backend stub for future GPU acceleration.

No Celery tasks wired yet — models load but nothing invokes them.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-04-10 09:00:06 +02:00

3 Commits