darkshapes/MIR

MIR (Machine Intelligence Resource)

A naming schema for AIGC/ML work.

The MIR classification format seeks to standardize and complete a hyperlinked network of model information, improving accessibility and reproducibility across the AI community.
The work is inspired by:

• AIR-URN project by CivitAI
• Spandrel library's super-resolution registry

Example:

mir : model . transformer . clip-l : stable-diffusion-xl

 mir : model .    lora      .    hyper    :   flux-1
  ↑      ↑         ↑               ↑            ↑
 [URI]:[Domain].[Architecture].[Series]:[Compatibility]

Code for this project can be found at darkshapes/MIR on GitHub

Definitions:

Like other URI schema, the order of the identifiers roughly indicates their specificity from left (broad) to right (narrow)

Domains

• dev: Varying local neural network layers, in-training, pre-release, items under evaluation, likely in unexpected formats
  
• model: Static local neural network layers. Publicly released machine learning models with an identifier in the database
  
• operations: Varying global neural network attributes, algorithms, optimizations and procedures on models
  
• info: Static global neural network attributes, metadata with an identifier in the database
  

Architecture

Broad and general terms for system architectures.

• dit: Diffusion transformer, typically Vision Synthesis
• unet: Unet diffusion structure
• art : Autoregressive transformer, typically LLMs
• lora: Low-Rank Adapter (may work with dit or transformer)
• vae: Variational Autoencoder etc

Series

Foundational network and technique types.

Compatibility

Implementation details based on version-breaking changes, configuration inconsistencies, or other conflicting indicators that have practical application.

Goals

• Standard identification scheme for ALL fields of ML-related development
• Simplification of code for model-related logistics
• Rapid retrieval of resources and metadata
• Efficient and reliable compatibility checks
• Organized hyperparameter management

Why not use `diffusion`/`sgm`, `ldm`/`text`/hf.co folder-structure/brand or trade name/preprint paper/development house/algorithm

• The format here isnt finalized, but overlapping resource definitions or complicated categories that are difficult to narrow have been pruned
• Likewise, definitions that are too specific have also been trimmed
• HF.CO become inconsistent across folders/files and often the metadata enforcement of many important developments is neglected
• Development credit often shared, Paper heredity tree, super complicated
• Algorithms (esp application) are less common knowledge, vague, and I'm too smooth-brain.
• Overall an attempt at impartiality and neutrality with regards to brand/territory origins

Why `unet`, `dit`, `lora` over alternatives

• UNET/DiT/Transformer are shared enough to be genre-ish but not too narrowly specific
• Very similar technical process on this level
• Functional and efficient for random lookups
• Short to type

Roadmap

• Decide on @ or : delimiters (like @8cfg for an indistinguishable 8 step lora that requires cfg)
• crucial spec element, or an optional, MIR app-determined feature?
• Proof of concept generative model registry
• Ensure compatability/integration/cross-pollenation with OECD AI Classifications
• Ensure compatability/integration/cross-pollenation with NIST AI 200-1 NIST Trustworthy and Responsible AI

massive thank you to @silveroxides for phenomenal work collecting pristine state dicts and related information