how do we use this

[xalteropsx]

@tolgacangoz bro how do we use custom_pipeline="lpw_stable_diffusion" #6976 with compel // i do have custom lpw but i want to know if diffuser have builtin with compel support

this is also make me think if it not possible #3171 but i saw recently some update on lpw pipeline no wonder if it work or not

[xalteropsx]

@tolgacangoz nvm notice here damian0815/compel#54

[asomoza]

Don't use a mix of them, they do exactly the same thing, the lpw pipeline has its own prompt weighting which is what compel does, you don't need to use compel if you use the lpw pipeline.

The lpw pipelines use the automatic1111 style of weighting like this:

a cat with a (red:1.5) ball

where compel uses it like this:

a cat with a red++ ball

[xalteropsx]

may be i download an model from_pretrained to test lpw thanks alot bro
if so the case then look like lpw more dominant then compel but that make really hard my whole thing depend .fromsinglefile lol >.< sorry for troubling but really awesome help

[asomoza]

yeah, the lpw pipelines do support more than 77 tokens, the lpw means long prompt weighting but this doesn't work with from_single_file and the custom_pipeline argument. You have to either download the pipeline locally and use it directly or with from_pretrained.

If you need to use it with a single file checkpoint, as I suggested you can save it locally and import it like this:

from lpw_stable_diffusion_xl import SDXLLongPromptWeightingPipeline

pipe = SDXLLongPromptWeightingPipeline.from_single_file(
    model_path,
    torch_dtype=torch.float16,
    use_safetensors=True,
)

and no problem, glad to help.

[xalteropsx]

ah i havent aware of this thing lol this really rock do have sd one as well ?

[asomoza]

yep

[xalteropsx]

awesome bro it really save me alot of time to switching from single and downloading model again thanks u million may god bless u >.< once i succeed hope to donate something