Feature Request #48
Description
Thank you all for doing such a great job creating useful custom nodes for comfyui. I'm especially fond of your wildcard processor node and use it to run any new model I'm evaluating through a gauntlet of complex prompts stored as wildcard files and parsed with a different node into positive, negative, CLIP-G and CLIP-L segments. Your wildcard processor has saved me countless hours being able to automate this work AND replicate it every single time.
I'm running into a bit of an obstacle with long clips. I was not aware of any problem until I recently started exploring SDXL workflows and found that my favorite clip text encoder was not performing as expected. In short, I write very detailed prompts sometimes according to Automatic1111 in the neighborhood of 2300 tokens. These are working great in SD1.5 with my favorite clip text encoder but in sdxl i get an error message (not handy at the moment but glad to supply) essentially that max() argument can't be void. This only happens when the encoder is set to the new way of handling tokens in A1111 parsing, which lets them extend into the several thousand (don't know the exact number rn). Anyway, switching from "new" parsing method to "old" parsing method removes the max() error but limits to approx 75 tokens. Unfortunately the potential of SDXL to shine is being limited in my creative world by my inability to find a suitable long clip text encoder for SDXL. I wondered if you could consider that project OR refer me to a working node (I have searched and am not able to find any that both support SDXL AND long clips). the feature set I'm looking for is support for clip g and clip l segments, parsing methods including comfy, comfy++, A1111, full, compel, and fixed attention, also token normalization with some choices about length, mean, and length+mean. I don't currently use multi-regional prompting but that's part of the broken node I'm struggling with and if that could be included as well, it would be great. The broken node I'm struggling with also supports the width/height, target width/height, crop width/height, and A score fields if desired. The reason I'm writing you is it doesn't look like that node has been maintained in some time now and I know you all are an active group, a competent group, and do a nice job resolving issues.