Int8_convrot quants in ComfyUI format of https://huggingface.co/lodestones/Kroma.
lodestones indicates that Kroma is offered under the MIT license — https://opensource.org/license/mit — but does not currently provide a copyright notice.
Kroma is a finetune of Krea 2, subject to the Krea 2 Community License: https://www.krea.ai/krea-2-licensing
I did not train and do not claim ownership of this model. It is trained and owned by lodestones — https://huggingface.co/lodestones/Kroma — I am just sharing int8 convrot quants for ComfyUI.
Description
This is the quant of the version merged with the Kroma Turbo diff after training, and should be used with the same settings as Krea 2 Turbo.
FAQ
Comments (14)
Not good
I'm sorry you feel that way. If you provide a little more information about why you feel it is “not good” maybe I can provide some information or do something else to help,
For what if Lora already here?
The v0.1 version of Kroma was released by the creator as a LoRA + .diff (bundled in one .safetensors) applied to the Krea 2 base model. v0.2 has only been released as a full diffusion model. If someone else has produced a v0.2 LoRA, I haven't seen it (but the full int8 model is still probably better for 30xx cards even if a LoRA is available.)
@StrugglingSorceror Ok I see
Model is interesting, some results was amazing, but have issues with stability and anathomy - sometimes fingers is bad. But model have strong sides - he can push to limits female curves.
It is the second early release of a model going through a heavy training regimen, so the current weaknesses should improve in later versions. But—with some caveats—I find it a generally strong general purpose model as it is.
It is very early checkpoint and largely why we have not done an official uploaded it to civ yet.
Does this INT8 ConvRot version retain the same recommended settings as the original Kroma v0.2 it was converted from — 8–12 steps, CFG 1.0–1.5, Shift 1.15?
It should work well with the same settings; I personally usually use the that range of steps in the default ComfyUI Krea 2 template which doesn't use Shift, with CFG 1.0.
They started too many branch projects, the Flux Klein version was terminated before they had time to perfect it. I don't know how the ZImage version is developing right now, and now they have started the Krea2 project again.
Chroma is great, but it still has performance and compatibility issues. I hope this project can reach Chroma's level and improve even further.
This one’s pretty fun so far. I’ve had good results with euler_a/sigmoid_offset. A few gens go a bit anime, but the realism is getting there. Seeing as that this is experimental, I’ll be watching for updates because it’s just going to get better.
Pretty decent with euler ancestral/beta 15 steps.
(v0.2) Let's GOOOOO!!
Very surprised how good it is already. Bye bye LoRas! Some images still need refinements but considering this is the early version I am hyped again, this surely will be the right successor of Chroma.
Ty for sharing this version