DaSiWa - Krea2
๐ค Why I Made This
Pushing Krea2 to its limits with my style:
๐ฅ Easy usage
๐ Enhanced details
โ๏ธ Less prompting hassle
๐ Enhanced Quality
๐ Touched by Darkness
๐ Better understanding of anime style composition
๐Workflow
My optimized workflow.
๐LoRA's
But: This checkpoint is not meant to replace all LoRAs, it is meant to:
Perform better overall at his own
As easy as possible to use
With LoRAs to be more awesome
โ ๏ธ Read the corresponding announcements.
๐ข Make sure to check it out for in-depth information and a complex comparison!
๐ ๏ธ Recommended Settings
Turbo
CFG 1
Euler/linear_quadratic | er_sde/simple
8 Steps
1536x1536
RAW
CFG 3.5
Euler/linear_quadratic | er_sde/simple
52+ Steps
1536x1536
Closing words
๐คฉ I want to thank all the fantastic other creators who made super nice LoRAs and concepts to play with! Support that awesome creators by using their LoRAs and post to their gallery and share the meta-data!
โ ๏ธ I made all this with permissions or open-source resources (the time it is incorporated).
I share as much insights as I can without compromising my work. I'm doing this for fun as my hobby and just do not want my hobby to be destroyed.
More details can be obtained in the corresponding announcements!
If you would like to contribute in my awesome (๐) checkpoint or willing to share resources I'll gladly give credit! Just contact me!
โ All credits / resources are mentioned inside the announcements! - Since different versions may have different resources.
YOU are responsible for outputs as always! If you make ToS violating content and I get aware I WILL report this.
Disclaimer
This models are shared without warranties and with the condition that it is used in a lawful and responsible way. I do not support or take responsibility for illegal, harmful, or harassing uses. By downloading or using it, you accept that you are solely responsible for how it is used.
Description
V2
Unchained
More NSFW
More Realistic Details
More Anime Details
FAQ
Comments (30)
THANKS! very cool model. Always up to date
Just heard about Krea2 earlier today, went to check it out and OF COURSE DaSiWa already has a cutting edge checkpoint out for it, and of course it works flawlessly. Very impressive. Having great results with a two-pass workflow. Surely DaSiWa will release a workflow that works even better than mine soon ;)
Thank you! Hehehe! My Image workflow already works for it ๐
I tried to use CuteDisaster in Forge Neo and got a convrot_w4a4 error.
Make sure if you use INT8 ConvRot, your setup supports this
@Darksidewalkerย I have a RTX 4090, which should be good enough, but so far i've struggled to get INT8 working in Forge Neo despite following instructions i've found.
Hang on, it's working now. I updated Forge Neo to the latest version, which oddly enough removed the INT8 option from the "diffusion in low bits" dropdown, but it seems CuteDisaster now works with the Automatic option now. The only other thing I did was add --sage to my launch parameters. Regardless it seems to work now.
How do I activate the 'Turbo' function? It seems to take a long time to generate images.
I am using CPU only. When I use Moody Krea 2 Mix (uncensored) v4, I can make a 512x512 image in 10 minutes. Using yours (DasiwaKrea2TurboRaw_cutedisasterV2Turbo.safetensors), it takes MANY hours.
Turbo means 8 step inference, it's always active
@Darksidewalkerย I just tried the NT4 version, and it was fast (it looks really good, too :) ). The INT8 version is super slow.
@Davnes007ย The "nf4" labeled version is INT4 ConvRot. Maybe the INT8 CR was slow because you are low on resources.
Hi, can you do FP8 'turbo' version?
Why?
INT8+ConvRot maintains the same quality as FP8 is more compatible and faster. I see no benefit for an extra version.
@Darksidewalkerย amd gpu gots trouble while using int8 model with loras, fp8 version is necessary in some kind of degree.
@bnx005514ย any references to that claim?
@Darksidewalkerย yes, int8 crashes on 9070xt (rocm 7.2.1):
Windows fatal exception: access violation Exception Code: 0xC0000005 File "comfy_kitchen\backends\eager\quantization.py", line 753 in fast_int8_mm File "comfy_kitchen\backends\eager\quantization.py", line 994 in int8_linear torch_hip.dll + 0x26DC36D, ?_int_mm@cuda@at@@YA...
i can send you a full log in DM.
@Darksidewalkerย What GPU are you using? just curious about that claim because on my RTX 4080 Super FP8 runs twice as fast as Int8 . this is repeatable on every model I've tried it on. I'm just wondering if your results are different because your GPU is a different model (3 series? 5? amd?)
@mmrift466ย Int8 is especially good for AMD, if this fails this must be a setup problem. The claim it will not work with Loras are not documented anywhere... So I'm curious if this is not just a claim from a faulty setup.
I did not say this. I'm really sure Int8 CR should run great on almost any setup, that's the whole purpose.
@Darksidewalkerย Sorry, the claim I meant was "INT8+ConvRot maintains the same quality as FP8 is more compatible and faster. I see no benefit for an extra version." when someone asked for an FP8 version. I'm saying my experience, across every int8+convrot model I've tried, is it runs 2x SLOWER than FP8 on my RTX-4080 Super. I will concede that I could be missing something on my system that handles Int8+convrot so if anyone has any insight there, I'm listening, but it's just been my experience that the RTX-40 series likes Fp8 best.
@mmrift466ย Did you update comfyui to latest? Updated torch and drivers? Int8 CR should run 20-60% faster on a 40xx card.
v2 gives more unwanted nudity than first one
Unchained vs not unchained
Also just prompt clothes, easy.
@Darksidewalker, I have the same experience. Clothes are prompted, and yet nudity is forced. Base model with SNOFS lora doesn't do that
@vamorandย than use that? Besides it's not forced since there are enough examples showing that's not. The bias is NSFW, since the checkpoint is tuned to that.
@Darksidewalkerย 100% prompt issues)
trying this in forge neo and I'm getting an error "You do not have Qwen3 state dict!"
I am on the latest version, I have the text encoder properly placed (it is the 5gb file with the correct name). I'm stumped.
Sorry I know you aren't a forge neo dev, I just dont know why it doesn't seem to like your model.
Best model of all I have tried for Krea 2. Hope you will improve it even more!
My god... by far the best Krea 2 model I've used...
Don't use the highly rated if you can get this far better one...
Does understand a lot of anime characters and animal parts anatomy, listens pretty well to camera prompts and understands ethnicity far better than other models.
You are a legend.
็็ๅพๅๆฌขไฝ่ ็ๆจกๅ๏ผไฝๆไฝฟ็จ็sd-webui-forge-neo-v3.3้กน็ฎ็ฎๅๅฏนint8็ฑปๅ็ๆจกๅ่ฟๆ ๆณ็จณๅฎ่ฟ่ก๏ผๆๅพ fp8็ฒพๅบฆ็ๆจกๅ
Absolutely the best Krea2 model I've tried. What I love the most about it is that it retains my character loras's breasts. Other models seem to be replacing the breasts with the breasts they were trained with.