I have deployed stable diffusion 1.5 from Vertex AI’s model garden. I have an endpoint on which I am successfully running online inference with a request payload like the following:
{ “instances”: [{ “prompt”: interesting_prompt }] }
Stable diffusion should be able to accept many parameters, including width/height, sampling steps, n_samples/batch size, etc. But I’ve been unable to figure out how to pass these parameters in a Vertex AI payload. It makes the most sense to me to pass them as key/value pairs alongside the prompt in the same object in instances, but no key names besides prompt that I’ve tried have changed the output at all. I haven’t found any documentation to support these use cases.
How can I pass these additional parameters for online inference?
See https://github.com/CompVis/stable-diffusion/blob/main/scripts/txt2img.py for the parameters that stable diffusion 1.5 text-to-image supports.
I can’t for the life of me find any documentation specifying how to pass these parameters to SD as deployed on a Vertex AI endpoint, and gcloud ai models describe [modelName, --project- --region etc] returns an empty predictSchemata object, so no guidance from an instance schema either.
Will SD as deployed from the model garden onto a vertex ai endpoint really not support passing any of these parameters?? Must I suffer through deploying vanilla SD1.5 as a custom model to use it normally with basic text2img parameters, even while SD1.5 already sits in the model garden? How can this be?? Who is responsible for stable diffusion’s castration!? What PM allowed this madness???
Hopefully I am missing something simple and obvious. Please show me my error.