[bug]: When using the RunwayML inpainting model, negative prompts do not work at all, which is not documented

[n00mkrad]

Is there an existing issue for this?

• [X] I have searched the existing issues

OS

Windows

GPU

cuda

VRAM

24 GB

What happened?

When trying to use a negative prompt (unconditioning) in combination with the RunwayML inpainting model, the words instead get picked up as part of the regular prompt.

To reproduce, run this for example:

portrait photo [monochrome]

SD 1.5 will never show a monochrome image, while SD 1.5 inpainting will always do as it interprets it as part of the prompt.

I assume this is an architectural limitation, however I did not find any mention of this in the documentation.

Screenshots

No response

Additional context

No response

Contact Details

No response

[damian0815]

@lstein any idea? i know the inpainting model runs a different code path, maybe somewhere in the hybrid conditioning the negative words are being mixed up

[yoniker]

I was about to open an issue exactly about that just now :) Running 1.5 inpainting on a linux machine (so it's probably not the environment).

[n00mkrad]

I did some testing; this always happens with that model, doesn't matter if it's txt2img, img2img, or masked inpainting.

[n00mkrad]

Weighting also does not work - None of the prompt features do, it seems.

[haqthat]

I am having this exact same issue. Thought I was doing something wrong, tried [] and weighting, neither worked.

[yoniker]

@lstein I think that you might be busy with the 2.x models, is this a trivial fix or a time-consuming fix?

[mrbaca14]

I have the same issue confirmed. Negative prompts with the inpainting 1.5 model do not work with the [brackets] for negative prompts. Confirmed negative prompts work as intended with all OTHER models.