Why different inputs text_l and text_g in CLIPTextEncodeSDXL?

[ifsheldon]

I notice that there are two inputs text_g and text_l to CLIPTextEncodeSDXL . I skimmed through the SDXL technical report and I think these two are for OpenCLIP ViT-bigG and CLIP ViT-L.

I don't know why Stability wants two CLIPs, but I think the input to the two CLIPs can be the same. So, @comfyanonymous perhaps can you tell us the motivation of allowing the two CLIPs to have different inputs? Did you find interesting usage?

https://github.com/comfyanonymous/ComfyUI/blame/4a77fcd6ab01d69e18c384faa29ae1c3d02237f3/comfy_extras/nodes_clip_sdxl.py#L41C5-L41C5

Thanks!

[yuhuihu]

I also have the same question. Here's som info I founded: https://www.youtube.com/watch?v=zucLSuCvqM8

[okolenmi]

I tested it on my own. It looks like G is good in details/concept and L is good in image content.

L is like "Charcter with long green hair on white background"
G is like "Angry face expression, crossed hands"

G is bad in things like character traits (age, clothes, etc.) L is good in this things. G is better than L in complicated concept things. You can add characters in L and then change what they doing in G.

[mcandre]

In practice, lots of users are manually copying and pasting the same prompt into both parameters.

Computers exist to automate. Let's create an option to automatically sync these fields.

[asagi4]

@mcandre if you just use the standard CLIPTextEncode node instead of the SDXL-specific one, it will use the same text for both TEs. Or you can just use a primitive text node and connect it to both inputs.