exdysa/mir

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

MIR (Machine Intelligence Resource)

MIR is a naming standard, a proposed schema for AIGC/ML work.
In its current incarnation, it looks like this:

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

 uri : model .    lora      .    hyper       :   flux-1
  ↑      ↑         ↑               ↑               ↑
 mir:[domain].[architecture].[implementation]:[compatibility]

The solution is provided as a remedy to patch the fractionalization of modelspec standards between development houses (such as models released independently or indifferently to HF.CO ) and to archive metadata which would otherwise remain incomplete.

This work was inspired by the CivitAi AIR-URN project
and by the super-resolution registry code from the Spandrel library.

Goals

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

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

• Exact frameworks (SGM/LDM/RectifiedFlow) includes too few
• Diffusion/Transformer are too broad, share and overlap resources
• Multimodal models complicate content terms (Text/Image/Vision/etc)
• HF.CO names do all of this & become inconsistent across folders/files
• Development credit often shared (ex RunwayML with Stable Diffusion)
• Paper heredity would be a neat tree, but it complicates retrieval
• Algorithms (esp application) are less common knowledge, vague, and I'm too smooth-brain.
• Impartiality

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

Roadmap

• Decide on @ (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