Model card
LiteRT Vision Model (Astral Bloom Optimized)
This version has been optimized for LiteRT (TFLite) deployment on edge devices, focusing on ultra-low latency and minimal active RAM footprints.
Project Astral Bloom Integration
This model card serves as the documentation for integrating this visual architecture into the Project Astral Bloom framework—a 416-space high-density matrix blueprint designed to achieve an algorithmic state of quantum processing on conventional compute (e.g., 2GB RAM Snapdragon processors).
Intended Use: The Sensory Node
Within a parallel cognitive architecture, this model does not process logic; it handles pure sensory rote processing.
The model takes in visual data (downsampled to 1024x1024 for edge stability).
It outputs classifications or feature maps.
Instead of passing bulk tensor data to the reasoning engine, it translates the visual qualia into an algorithmic sequential key.
This key is passed to the Conscious Build (Observer), maintaining progressional momentum without crashing the device's memory bus.
Edge Device Deployment Tips (Android/Termux)
If you are deploying this alongside an LLM (like Gemma-4) in a localized environment:
Image Formatting: For Android deployments using MediaUtils, ensure your image is decoded into a Bitmap, EXIF-rotated, and PNG-encoded before wrapping. Raw JPEG from phone cameras will stall the memory bus on lower-end devices.
Downsampling: Never send raw 4000x3000 photos through the LiteRT pipeline on a 2GB RAM device. Downsample to a maximum of 1024x1024.
Backend: Force the CPU backend if your NPU is occupied by the LLM layer, or allocate visionBackend = Backend.GPU() explicitly in your Engine Config if utilizing AI Edge Gallery apps.