Sorry, where is the Turbo version of Kroma 0.3?
Won't there be a Turbo version anymore?
I think its in development
I think its in development
I don't think that's the case
silveroxides/Kroma-Quant
have some sort of Kroma 0.3 turbo fusion.
The mcard say Turbo model has the Turbo delta pre-merged; which implies that with the base version, you can merge the Turbo delta yourself OR use any existing Krea 2 Turbo LoRA.
The mcard say Turbo model has the Turbo delta pre-merged; which implies that with the base version, you can merge the Turbo delta yourself OR use any existing Krea 2 Turbo LoRA.
I figured that but the base being 50 ish GB makes it a problem, mainly cause it's F32 so having a BF16 would be great even if it's just base
I agree, F32 is huge. It's not really meant for direct inference. For generating images, it's better to quantize it yourself or just use a quantized model. You can still add any Turbo LoRA to those.
"kroma-v0.3-tdm-4steps-artifact.safetensors" so what is this then? Is it a turbo lora? If not, is there a specific Turbo lora people are using that is good?
there is bound to be a int8 quant somewhere soon. And possibly a krea2 turbo lora works. since we're seeing relatively rapid 0.x numbers it might be a bit messy for a while.
lodestones is historically not very good at documenting what versions do what of what. like different chroma versions. BUT very good a t making them...
btw, don't dismiss 0.1 because he number is lower. Haven't tested 0.3 yet (for this reason) but the 0.1 lora on top of normal krea2 in8 or fp8 turbo, works very well for what i want.
And better than0.2. It is basically my default currently, even for non-spicy (photorealistic) content. It makes super sanitized krea2 actually useful.
Using ostris' krea2 text encode edit node (but not his lora or model patcher which give me weird artrifacts) there is even a relatively useful reference image function for people, styles and so on. (although not editing).
not on 50gb (not just vram, swapping on my 32gb would work ok), but that is just a enormous space just to store for one model. That almost half the size of a Stalker 2 patch :P.
but deffo one of these days a int8+turbo lora. i hope krea2 turbo loras still work. The "problem" is, i'm more or less happy with 0.1 :) it unlocks krea2 from most restrictive model ever to full nsfw and all in between.
Likely this one: Stabhappy/kroma-v0.3-base-INT8-convrot It's 14GB. that easily does with the krea2 turbo lora on top ( krea2_turbo_lora_rank_64_bf16 ) , at least vram wise. And their 0.2 version worked well.
There are some smallified models by silveroxide with very fancy names and a very complicated explanation for these names, but those for 0.2 didn't work for me, since they felt like fully censored Krea2.
it's up
thank you!
Do you mind if I do a int8 convrot version of it, for those that are VRAM poor?