62 lines
2.5 KiB
Markdown
62 lines
2.5 KiB
Markdown
# AI Features
|
|
|
|
## Overview
|
|
|
|
Majoor Assets Manager includes optional AI-assisted discovery and enrichment workflows.
|
|
|
|
Examples include:
|
|
- semantic search from natural language queries
|
|
- image similarity search
|
|
- automatic tag suggestion
|
|
- caption enrichment
|
|
- prompt alignment and quality scoring
|
|
|
|
These features depend on optional models and packages and are not required for the base extension.
|
|
|
|
## VRAM Controls
|
|
|
|
For ComfyUI systems where generation VRAM is more important than Majoor AI discovery:
|
|
|
|
- turn off **Enable AI semantic search**
|
|
- keep **Index vectors during scans** off
|
|
- keep **Generate AI captions during indexing** off
|
|
- set **Vector indexing concurrency** to `1`
|
|
- enable **Unload AI models after use**
|
|
- use **Memory purge** in Index Status when you want to immediately release Majoor SigLIP/X-CLIP/Florence caches and ask ComfyUI to unload loaded models
|
|
|
|
Automatic unload-after-use targets Majoor AI models and torch CUDA cache. The manual Memory purge action also asks ComfyUI to unload loaded models when the queue is idle.
|
|
|
|
## Privacy Model For AI Features
|
|
|
|
The current documentation clarifies an important point: AI inference is intended to run locally on the user's machine.
|
|
|
|
In practice, that means:
|
|
- prompts and images are not sent to a hosted cloud inference service for semantic search, similarity, captions, or prompt alignment
|
|
- internet access is mainly needed for the initial model download from HuggingFace
|
|
- once models are cached locally, AI features can work offline
|
|
|
|
This is the main reason the project now has a dedicated privacy and offline guide.
|
|
|
|
## When To Use AI Features
|
|
|
|
AI features are most useful when:
|
|
- the library is too large for manual browsing alone
|
|
- filenames are inconsistent or unhelpful
|
|
- you want thematic grouping across many unrelated folders
|
|
- you want better retrieval from prompt or visual intent
|
|
|
|
## Token Clarification
|
|
|
|
Two token concepts appear in the docs and should not be confused:
|
|
|
|
- HuggingFace token: optional, used for model download reliability and rate limits
|
|
- Majoor API token: used to secure remote write access to the local backend
|
|
|
|
Neither token is described as a hosted inference key for uploading user prompts or assets.
|
|
|
|
## Canonical Docs
|
|
|
|
- [AI Features Guide](https://github.com/MajoorWaldi/ComfyUI-Majoor-AssetsManager/blob/main/docs/AI_FEATURES.md)
|
|
- [Privacy and Offline Guide](https://github.com/MajoorWaldi/ComfyUI-Majoor-AssetsManager/blob/main/docs/PRIVACY_OFFLINE.md)
|
|
- [Security Environment Variables](https://github.com/MajoorWaldi/ComfyUI-Majoor-AssetsManager/blob/main/docs/SECURITY_ENV_VARS.md)
|