Flux... only smol
Added 2024-08-18 11:25:58 +0000 UTCLoads more unets have dropped - ideal for those with smol GPUs! There is even a Q2 :)
Can be 50% slower than the standard dev model FP8 weight type, but it's a small price to pay. Wonder how many times I can mention small?
Anyway, here are all the options at the moment:
https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main
You will need the GGUF UNET loader custom load though:
https://github.com/city96/ComfyUI-GGUF
Looks like some may even work with loras too!
Comments
GGUF is just 1 optional node from https://github.com/city96/ComfyUI-GGUF - simply replace the existing model loading node in any workflow with one to load GGUF instead :)
Nerdy Rodent
2024-08-24 11:54:07 +0000 UTCWheres the json workflow for this one
Theron Govender
2024-08-23 23:29:16 +0000 UTCSo many to try now!
Nerdy Rodent
2024-08-21 18:31:40 +0000 UTCThe Q4 K S version is now my go to model for Flux, as it works with all of the LoRAs I've tried. Q8 didn't seem to work with LoRAs.
Nibmeister
2024-08-21 00:46:24 +0000 UTCThe controlnets still aren't really up to SD 1.5 quality yet, but give it a couple of months and all the bits should be there!
Nerdy Rodent
2024-08-20 18:42:19 +0000 UTCFantastic. Any chance you might revisit your Reposer workflows with Flux in mind? Particularly interested in the face + pose + clothes one
Stephen Hilsdon
2024-08-20 13:42:09 +0000 UTCI get a lot of odd warnings loading the Q8 version of Flex Dev GGUF but it loads really fast and achieves 6s/it on my 8GB VRAM 3070TI.
Nibmeister
2024-08-18 15:44:42 +0000 UTC
