Sharing a broad, practical and universal approach to formatting model information. Would like to see any community solution to be brand, system, modality, language and format independent by abstracting compatibility into units that apply to classification as much as they do code application.
link This is something borne out of app building struggles and prior example (most notably the work of AIR from CivitAI and the registry from Spandrel/ChaiNNer). The idea is that the details of any one dimension (ie: domain, architecture, implementation) are sub-attributes of that dimension. Ex, the state-dict entries of SDXL would be stored beneath the implementation header, perhaps as a file or in a mapping based on the class hierarchy. I am in the process of building around this, maybe it can inspire discussion or further innovation, or perhaps it should evolve to be closer to OMI-MS in some respects? Posting here because of access restriction to standards channel.
Sharing a broad, practical and universal approach to formatting model information. Would like to see any community solution to be brand, system, modality, language and format independent by abstracting compatibility into units that apply to classification as much as they do code application.
link This is something borne out of app building struggles and prior example (most notably the work of AIR from CivitAI and the registry from Spandrel/ChaiNNer). The idea is that the details of any one dimension (ie: domain, architecture, implementation) are sub-attributes of that dimension. Ex, the state-dict entries of SDXL would be stored beneath the implementation header, perhaps as a file or in a mapping based on the class hierarchy. I am in the process of building around this, maybe it can inspire discussion or further innovation, or perhaps it should evolve to be closer to OMI-MS in some respects? Posting here because of access restriction to standards channel.