Minimal txt2img example

[akx]

• ⚠️ This PR is still a draft since it builds on and as such includes Un-hardcode "cuda" as default device name #49 and Late-import scipy (only required when using LinearMultistepSampler) #50 (and Late-import scipy (only required when using LinearMultistepSampler) #50 may be superseded by the LPIPS vendoring in Pre release 1.0 #59).

---

This PR adds a simple, minimal txt2img command line example tool that also works on Apple Silicon.

By running env SGM_DEFAULT_DEVICE=mps python txt2img.py on this branch you'll eventually get a familiar sight in outputs/:

It also adds the minimal set of dependencies I identified to be enough to run the demo (as originally discussed in #17).

[bghira]

can you please ensure you test a batch size greater than 1 through the base and refiner?

[akx]

@bghira I'm holding off on this PR until the 1.0 stuff stabilizes. This PR will probably be reworked to use the new higher-level helper APIs from #57 (which are essentially a superset of what I had already written here).

I did test with a batch size of 2, that did work fine.

[bghira]

@akx the batch size issue is prompt-dependent, eg. happy child flying a kite on a sunny day introduces many watermark artifacts with batch_size > 1, but not at batch_size = 1. and the default astronaut prompt, does not cause any issues.

the reason so far as I can tell is due to some issue in the stride used for the refiner steps. but i've mostly managed to reproduce the issue, not identify the actual cause or solution other than refining images separately.

edit: and the issue is independent of MPS or CUDA being used.