fix: inject negative embeddings correctly

[tazlin]

See #69

also reworks test to confirm this behavior.

[db0]

All the expected images are the same which is not right.

[tazlin]

I have changed the tests to expect the same image on output. The purpose of this is to make sure the inject behavior is working in a predictable way.

[db0]

Ah I see. If you think it helps with noticing issues, sure, but normally the comparison of the images would be automated. The pedant in me wants to mention that you could avoid uploading 3 references by pointing them all the same image :D

[tazlin]

I can just as soon do that. All I really want separate are the test output images anyway.

Although, I do think it would highlight the fact to a manual review that they are expected to be the same with the duplicate images (as in, on a test failure).