Klein FinalCut
4k Workflow
Universal Formats
FP16
FP8 with FP32 sensitive blocks
Comfy Quants
Scaled_FP8
INT8_Convrot
Description
FAQ
Comments (8)
Another great release, follows the prompt very well with default clip. Thanks a bunch for sharing. ♥
Thanks
This is great, follows prompts for a some things way better than others and does a lot less of the weird stuff they do, also the only Klein 9b model of the ones I've tried that will consistently generate video game/animated characters realistically without having to throw in 3 or 4 extra weighted realism related words in the prompt. Thanks for sharing
Thanks
Thanks! this model fixes most of the problems that made me discourage F2k. For some reason most Klein9b models on my 4060 generate 4x slower when I use lora. In the case of your model, when I omit the sage node and set dtype to none, the generation time is only 10-20% longer with lora than without. this is wonderful.
I checked many models and only realdream also works this way, but I definitely prefer the output from your model.
THANKS! x2
Thank you
hey brother, i dont know if you figured it out but i had issues with Sage Attention speeds on my 3090 and it was because i had the Sage Node placed right after the Model Loader Node. Once i placed it after my Power Lora Node, BOOM! fast as hell.
@TheAsh3n omg... you may be right, I always put sage before lora loader. Thanks for the answer, but the problem no longer concerns me. I bought rtx5070 and all problems disappeared :) lora or not lora, sage or not sage, it always goes like a storm :)