Update:
- V3 is about half RAW still runs at 8-12 steps.
- Added V3 FP8.
- Added V3 Int4 ConvRot SR (stochastic rounding).
⚠️ IMPORTANT — Applies to v1.0, v1.1,v2, v3 ⚠️
❌ Do NOT use the INT8 Fast custom node.
This checkpoint is already in native ComfyUI INT8 (ConvRot) format.
✅ Use the standard ComfyUI Load Diffusion Model node.
Requirements
- ComfyUI v0.27+
- ComfyUI Kitchen
Krea2 Center v1: similar, experimenting with some block merging.
Krea2 Center v0.5 explores whether Krea2 can become more expressive through carefully balanced LoRA merging while preserving the model's realism, flexibility, and overall character. The goal isn't to replace Krea, but to gently extend its emotional range across portraits, fantasy, sci-fi, and stylized scenes.
Added Int8: about 25% faster on 50 series Nvidia don't have a 40 or 30 series to test.
Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing.
Description
INT8 ConvRot Learned (Experimental)
Uses AdaRound + Prodigy Adaptive learned quantization.
Compared to the previous INT8 ConvRot release, this version preserves significantly more fine detail and is visually much closer to the BF16 model in my testing, while retaining the same memory savings.
This is an experimental learned INT8 ConvRot. It aims to better preserve fine details and structural accuracy compared to the previous INT8 ConvRot. The improvements are often subtle and may be most noticeable in complex scenes, faces, clothing, armor, and other highly detailed subjects.
FAQ
Comments (13)
Are these merges or closer to the base model? I feel like all these checkpoints I tried for krea lose the original knowledge and likeness.
v3 checkpoint is sooo good, but sadly it produces the ugliest woman i'ever seen in krea lol and ignores any loras trying to fix it, so no go for me.
Version 3.0 is decent, leaning more towards sexy figures with good control over body shape. However, its drawback is that the facial features are somewhat monotonous, making it seem like they all depict the same person. Close-ups of sensitive areas are captured well.
The v3 human model is perfect, but the only drawback is that there are too many uniform faces. Hopefully, this will be greatly improved in the next version.
one of the best models; however, it doesn't seem to handle tops or short clothing very well XD.
It even produces NSFW content unintentionally.
good to know I will try to fix that next version
It gives me ERROR: KeyError: 'convrot_w4a4' when loading diffusion INT4. I am using an Ampere Card, RTX3060, which supports int 4). I am already with Comfyui Latest version (0.27) and Comfy Kitchen. The version INT8 works.
0.27 is not latest and int4 convrot is not in that you need at lest 0.27.1 ideally maybe 0.28
Good chepoint. Unlocked without applying gazillion "bypass" loras, not overfitted like other civitai merges. But seed variety is significantly worse than base k2 turbo.
>applying gazillion "bypass" loras
Those ruin the model... just use the "refusal reducer" lora and be done.
@brnfd24434343d dont use either ;) and yes many bypass / rebalance loras mess up the models
@brnfd24434343d The key word from the OP was "without applying" i.e. not using bypass loras to create good images, just saying that variety is lacking without changing the prompt, random seed does little to give new interesting images.
Yep. Bypass loras have their drawbacks. Character likeness is worse, skin details are lost. And "text refusal" lora gives image a "dirty" look, hard to explain just some leftover noise, esp. when combined with some other loras.
This model just gives better experience right out of the box.