IMPORTANT
INT5 model requires my custom node - https://github.com/bakapotatolord/ComfyUI-PotatoForge-INTx
Clone it in ComfyUI/custom_nodes folder or it won't work. INT8 Mixed model should work as is, no need for custom node.
Quantization of Krea 2 Turbo primarily meant for low VRAM/RAM usage, expect degradation in quality compared to INT8 Convrot only quants or BF16
INT5 is highly experimental, my goal was to find a middle ground between INT4 and my very own INT6. All block layers are in int5 convrot, it's not mixed.
Once again, used a simple T2I workflow, no second stage or upscaling done as I'm too lazy for that.
Tested on:
NVIDIA GTX 1660 Super 6 GB
32 GB System RAM
Test settings:
8 steps
CFG 1
Euler / Simple
696 × 1048 resolution (0.7 on resolution selector)
Approximately 9 s/it on a GTX 1660 Super
Quantized using my toolkit
Description
Requires this custom node - https://github.com/bakapotatolord/ComfyUI-PotatoForge-INTx
FAQ
Comments (4)
Would love to see 1:1 seed comparisons and timings on your rig. The int8 runs well about 25% less time than the standard int8...
I don't want to re-jig my whole setup to work new nodes in if its not more savings and minor quality drop.
I will try to post comparison images, but do note there will be drift between images cause of quantization.
INT5 by itself is just storage savings (compared to INT8) and better quality than INT4, I'm hoping to have a mixed model with INT4, INT5, INT6, INT8 that should hopefully be better than INT4-INT8 mixed at same size.
What do you mean by standard int8 btw? I'm surprised about that 25% less time lol
@BakaPotatoLord Krea 2 Turbo int 8, from Civ, runs 36 seconds +/- on a 1600x1200 image on my 3070ti (8gb) - your int8 mixed runs 28 seconds +/- and its likely just due to memory swapping
@xFennec777 That's pretty surprising. I have to assume it's the size of model (my first one being a mix of int4 and int8) and the amount of RAM you have. If it goes to pagefile, then performance takes quite a hit.