BrunoThe Git-native API client. Local-first, open source.
DAAAM
A local AI search engine for all your content
6 followers
A local AI search engine for all your content
6 followers
Existing DAMs are photo-first, cloud-dependent, or give you tags like "beach." DAAAM is built for video. A custom vision model runs on your Mac and captions every frame: shot type, mood, composition, scene, lighting. Audio is transcribed with word-level timestamps fused to the timeline. Search by what you see or what was said. Connect to Claude via MCP for AI-driven storyboarding. Export to Final Cut, Premiere, or Resolve. Perpetual license. Your footage never leaves your machine.
How does the custom vision model hold up with lower resolution or heavily compressed footage where fine details like mood or shot type might be ambiguous to detect?
@ravzanv5o
Great question!
For macro-level details like mood, shot type, lighting, and composition, the local vision model holds up incredibly well even on highly compressed or low-resolution footage. Because modern multimodal models are trained on a wide variety of real-world data, they analyze global color palettes, contrast patterns, and spatial layouts rather than micro-textures to determine the 'feel' and structure of a scene.
To help the model handle degraded or flat-colored footage, we do a few key things under the hood:
Dynamic Range & LUTs: For flat Log profiles (like Apple Log, S-Log, etc.) or HDR, our ingest pipeline automatically tone-maps the frames to Rec.709. This ensures the model sees accurate contrast and lighting cues rather than a washed-out, low-contrast grey image.
Smart Filtering for Fine Details: For micro-details like face indexing, we run strict quality gates. If a face is too small (under 60px) or excessively blurry, it is automatically filtered out so noisy data doesn't clutter your catalog.
Built-in System Honesty: The model is instructed to never 'guess' or hallucinate details. If compression, motion smear, or low light makes a detail ambiguous, it flags these sources of noise in a confidence_note so downstream indexing knows it’s uncertain.
In short: search queries like 'warm backlit wide establishing shot' work beautifully even on degraded footage, while the system gracefully flags or filters out fine details (like faces or small text) if the quality drops too low!
The frame-level captioning on a local Mac is genuinely impressive. Searched a messy Premiere project for "low angle, golden hour" and it just worked.