Models deployed via COG (i.e user deploys his own container/cog based model, or uses some of the ones we deployed (notably SDXL)) hava a more specific input/output. The actual input and output data can be opaque (i.e object), but the structure (i.e where are the stats, webhooks etc), is specified.
Ideally SDXL should inherit from CogBaseModel (or CustomModel, although CustomModel doesn't capture any of the cog specifics), so the CogBaseModel can return output + stats in a predictable interface.
I can work on that in the coming days.
TODO: Link to some example input output, now the text is a bit unspecified.
ichernev
commented
6 months ago
Models deployed via COG (i.e user deploys his own container/cog based model, or uses some of the ones we deployed (notably SDXL)) hava a more specific input/output. The actual input and output data can be opaque (i.e object), but the structure (i.e where are the stats, webhooks etc), is specified.
Ideally SDXL should inherit from CogBaseModel (or CustomModel, although CustomModel doesn't capture any of the cog specifics), so the CogBaseModel can return output + stats in a predictable interface.
I can work on that in the coming days.
TODO: Link to some example input output, now the text is a bit unspecified.