Verifying the work
After the not-so-many lines of code, we finally have all the logic ready, so let’s give the code a test.
In the simple version of a long prompt encoder, we still get a cat with some patterns in the body instead of pure white, as we gave in the prompt. Now, let’s add weight to the white keyword to see whether anything happens:
prompt = "photo, cute cat running on the grass" * 10
prompt = prompt + ",pure (white:1.5) cat" * 10
neg_prompt = "low resolution, bad anatomy"
prompt_embeds, prompt_neg_embeds = get_weighted_text_embeddings(
pipe, prompt = prompt, neg_prompt = neg_prompt
)
image = pipe(
prompt_embeds = prompt_embeds,
negative_prompt_embeds = prompt_neg_embeds,
generator = torch.Generator("cuda").manual_seed(1)
).images[0]
image
Our new embedding function magically enabled us to generate a pure white cat...
