Home Artists Posts Import Register

Downloads

Content

Instruct Pix2Pix is working, is simiar to img2img, but it will take the original image a lot more in consideration.

Gonna try to finish the build tomorrow, the complex stuff is working, just need to fix now a lot of little stuff.

Comments

Anonymous

Is there a difference in how the prompt is formatted? Are the describing terms "turn him into" weighted in some way or can you just describe the targeted final image by saying "a cyborg" and pretty much get the same result?

cool1

In a help section on an Instruct Pix2Pix page it says "Rephrasing the instruction sometimes improves results (e.g., "turn him into a dog" vs. "make him a dog" vs. "as a dog"). In the readme file of it it says it was trained on " (1) an input image, (2) an editing instruction, and (3) an output edited image" (a dataset of 454,445 examples" of those according to that page) so I assume that a prompt that's more like an edit instruction (like was in the training) might be better, though we can try out different ways when we get the GUI release.

Anonymous

I looked through the LORA options on civitia and I gotta be honest. I dont see much there that appeals to me. A wide variety of anime and photoreal based on sex positions/acts (meh) and specific famous people (boring), that I will never use in my art makes up the bulk of options. I saw a few that might be valuable but underwhelmed with options around style modifications. That said, perhaps the variety will grow over time. My 2 cents.

Anonymous

You can extract a LORA from a dreambooth model. It will not be 100% the same quality, but it will be 90% of the way there, while also being ~100, 200MB. This allows you to then use the LORA on top of other models. It is very worth putting time into.