Do negative prompt has the limit on its length?

[aoirusann]

From code, it seems both of cond and uncond are processed in a same way: calling prompt_parser.get_learned_conditioning:

# modules/processing.py
with devices.autocast():
    uc = prompt_parser.get_learned_conditioning(shared.sd_model, len(prompts) * [p.negative_prompt], p.steps)
    c = prompt_parser.get_multicond_learned_conditioning(shared.sd_model, prompts, p.steps)

So though there is no token counter for negative prompt, it should also have token limit?
Am I right?

If this is true, maybe I can add a token counter for negative prompt......

[ClashSAN]

currently if you go over 77, the token limit automatically increases. I assume this would apply to negative prompt.
Note about auto prompt extension: some tradeoffs may have been made to allow for this, though I'm not sure yet what these are specifically. If someone would provide these details, It'll be nice to drop them in the wiki.

[aoirusann]

I see...
I'm also not sure yet what those tradeoffs are specifically. Maybe I need some time to learn about Diffusion Model...
But it seems a token counter can also be helpful for negative prompt. I'll take a try.

[B34STW4RS]

If I use a negative prompt thats a bit too long, speed automatically halves, taking away a token or two usually rectifies this, but I can keep adding positive prompts without any noticeable speed impact which is quite strange.

[aoirusann]

#2572 Metion the PR. I'm working for this discussion.