Describe what this pull request is trying to achieve.

This adds support for stable-diffusion-2-1-unclip checkpoints that are used for generating image variations. (See also.)

It works in the same way as the current support for the SD2.0 depth model, in that you run it from the img2img tab, it extracts information from the input image (in this case, CLIP or OpenCLIP embeddings), and feeds those into the model in addition to the text prompt. Normally you would do this with denoising strength set to 1.0, since you don't actually want the normal img2img behaviour to have any influence on the generated image.

One thing I did not implement is any way to use this functionality but starting from random noise like txt2img does - which would probably generate more varied variations. This would be good for future work.

Additional notes and description of your changes

Key changes:

• Autoloading the config. This is straightforward, except one complication: v2-1-stable-unclip-l-inference.yaml requires and hardcodes the path of a supplementary file at checkpoints/karlo_models/ViT-L-14_stats.th . I opted to hotpatch the config to point at models/karlo/ViT-L-14_stats.th instead; and to check the file into the repo itself, since it's very small - thoughts on this approach?
• unclip_image_conditioning() - the key feature of this change. Adapted from Stability's script.
• Changes all over sd_samplers_compvis and sd_samplers_kdiffusion to account for the fact that these models expect the conditioning to be provided differently (e.g. it needs to be in c_adm instead of c_concat). Not the cleanest code, but it works.

Environment this was tested in

Windows, NVIDIA GTX 1660 6GB

Screenshots or videos of your changes