Prompt weight info and parentheses passed to text_encoder

[david-rokeby]

As far as I can tell, if you send a prompt like

(A lovely beach-side cottage in the southern seas)0.5

the entire prompt including the parentheses and the 0.5 weight are passed into the tokenizer, and those additional tokens are passed along to the text_encoder.

Am I reading this correctly or have I missed something?
Is this the appropriate thing to do? I would have assumed that this additional weighting syntactical stuff should be removed before passing to the encoder.

[damian0815]

hmm, that should'n't be happening. can you post the complete sequence of calls? if you're doing compel("prompt") then the prompt is being parsed into a tree structure first, which is used to guide the assembly of the embedding tensor (eg applying weights/blends), and this is then returned.

[david-rokeby]

Hi Damian, Here is a test that shows the behaviour: (this is on Mac OS… similar on Ubuntu) compel 1.1.0 pypi_0 pypi pyparsing 3.0.9 pypi_0 pypi python 3.10.10 hc0d8a6c_2 diffusers 0.14.0 pypi_0 pypi from compel import Compel from diffusers import StableDiffusionPipeline pipeline = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5") compel = Compel(tokenizer=pipeline.tokenizer, text_encoder=pipeline.text_encoder) prompt = '(a medieval painting of the Virgin Mary riding a rocket through outer space.)0.7. (I love you.)0.8' conditioning = compel.build_conditioning_tensor(prompt) I get the following from conjunction = self.parse_prompt_string(text) Conjunction:[FlattenedPrompt:[Fragment:'(a medieval painting of the Virgin Mary riding a rocket through outer space.)0.7. (I love ***@***.***] | type AND | weights [1.0] | loras [] Without the periods I get: Conjunction:[FlattenedPrompt:[Fragment:'a medieval painting of the Virgin Mary riding a rocket through outer ***@***.***, Fragment:'I love ***@***.***] | type AND | weights [1.0] | loras []

…

On Apr 13, 2023, at 11:06 AM, Damian Stewart ***@***.***> wrote: hmm, that should'n't be happening. can you post the complete sequence of calls? if you're doing compel("prompt") then the prompt is being parsed into a tree structure first, which is used to guide the assembly of the embedding tensor (eg applying weights/blends). — Reply to this email directly, view it on GitHub <#22 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/APRNCD5OC6FDDIPTWTBI4XDXBAJAPANCNFSM6AAAAAAW226KOY>. You are receiving this because you authored the thread.

[damian0815]

thanks, that seems like a parser bug, i'll take a look

[damian0815]

yes, that's a bug. fixed in compel v1.1.1 (just released)