V10 (Krea 2) is OUT!
Download is available on CivitAI now. I've removed the gofile and google drive links since they are no longer needed.
🚨 Attention:
There was a bug with the fp8 version caused by a faulty fp16-to-fp8 conversion script, which resulted in blurry images lacking detail. If you encountered this issue, please re-download the file.
There was a bug with the fp8 version caused by a faulty fp16-to-fp8 conversion script, which resulted in blurry images lacking detail. If you encountered this issue, please re-download the file.
Sorry for the inconvenience!
Preview pics can be found on Civitai.
Available Model Variants:
- Raw: fp16 / fp8 / int8_convrot
- Turbo: fp16 / fp8 / int8_convrot / Q4 GGUF (reuploaded the working version) / Q2 GGUF
Hardware Recommendations:
- 8 GB VRAM: Your best bet is the Turbo Q4 GGUF variant.
- 12 GB VRAM (or if you don't mind slightly longer generation times): Go with either Turbo fp8 or Turbo int8_convrot.
- 24 GB VRAM: You can comfortably run the full Turbo fp16 variant.
Use Turbo for everyday inference (generation) and Raw if you plan on training or fine-tuning.
And
Thank you for the model. Please tell us more about available concepts, params, features etc in the v10
Jul 10 21:05
coyotte
And, yes, I will make an article on Civit with all the info soon
Jul 10 21:09 (changed)
coyotte
🚨 Attention: There was a bug with the fp8 version that caused it to produce blurry images lacking detail. If you encountered this issue, please re-download the file.
Sorry for the inconvenience!
Jul 11 01:12 (changed)
1
Infinite Cetra
FYI Google drive links still include models that produce blurry results except for the turbo. The turbo works great though.
Jul 12 08:43 (changed)
coyotte
Infinite Cetra, are you talking about the raw version? if so, that's not a bug - Raw versions are only meant for training (if you want to use them for generations, you need a lot of steps)
Jul 12 11:23 
1
CheekyBeekeeper
Any LoRA training suggestions to optimize for your workflows?
Jul 25 06:40
po vaj
Are Krea2 fine tunes at a point where its useful downloading them yet? I know it takes a long time for checkpoint fine tunes to converge. It took SDXL over a year to finally see good quality model fine tunes. Looking forward to a fine tune model that pushes Krea2 past that synthetic digital art plastic look.
Jul 28 05:06 (changed)
Wittaya Piriyasurawong
Hi, coyotte. Not sure what happen. I use v8 without any error. But for the v10, I got error at the negative prompt with turbo fp16 version. Please kindly check.
Jul 31 06:36
1
coyotteReplying to Wittaya Piriyasurawong
Wittaya Piriyasurawong, Since this is not a Unet SDXL model, but Diffusion Transformers-based one, you need to use an external text encoder, this one is correct
https://huggingface.co/Comfy-Org/Qwen3-VL/resolve/main/text_encoders/qwen3vl_4b_fp8_scaled.safetensors
Jul 31 19:06
1
Wittaya PiriyasurawongReplying to coyotte
coyotte, I try by use your recommended workflows. Now, I understand for your explanation. Thank you for your support very much.
Aug 01 06:20
1
Ciddan Junk
Overall the model is truly excellent, but too many shots are ruined by random cum showing up totally unprompted. Especially when using FaceDetailer or refinement samplers.
Jul 31 17:21
coyotteReplying to Nick Zane
Nick Zane, I've linked the workflow in my reply to Ciddan Junk, and had explained that issue in my reply to Wittaya
Aug 01 22:48 (changed)
1
Nick ZaneReplying to coyotte
coyotte, amazing mon ami thank you very much managed to get it to work, would be amazing if you could share some workflows and the screenshots of them working/in action with the modules to maintain face consistency. anyways thanks again
Aug 02 13:26
DisrupttheNorm
Tried to install the turbo fp8 model and my invoke keeps crashing. Should I be using a different UI or should it be working?
Aug 04 00:17
coyotte
Shmowen, haven't used Invoke since 2023, so I can't tell for sure unfortunately. They might not even support Krea 2 architecture yet.
Aug 04 11:48
DisrupttheNormReplying to coyotte
coyotte, I switched to ComfyUI and am having a much better experience. I also grabbed your workflow. Thank you for doing all this work! Now I just need to figure out how to train specific branches of this model with my own images. I did download the "Raw" fp8 version after trying the Turbo. I am very new to this.
Aug 04 21:29
Travis Appel
how do I get this on comfy?
Aug 06 05:28
Travis AppelReplying to coyotte
coyotte, so this is only for local, not for a runpod by chance. - Nevermind I figured out how to download to the pod :) Is this a checkpoint?
Aug 06 14:46 (changed)
coyotteReplying to Travis Appel
Travis Appel, That depends on your 'checkpoint' definition. This is a diffusion model. It's not like SDXL, where you had Unet (Diffusion Model) + Text Encoder + VAE all-in-one. You need to add each manually.
Aug 06 19:03
Uncle Ne Dobre
Discord link is expired in DM :)
Aug 16 23:09