krea anime V2
A well-balanced model that combines flexibility, precision, and a high level of detail; in terms of quality, it’s hot on the heels of Krea 2 Anime UWU v1 + Qwen Image 2.1.
It is recommended to use the text encoder proposed in 1.7.
8-12 step
euler\simple (I haven't tested it on others)
image resolution 1–1.3–1.5–2-3 MP
A mix of precisions: although the model file is about 800 MB larger, it actually contains only 1.7 million more parameters than the standard Krea2 Turbo. However, the largest fine-tuned matrix—comprising 343 million parameters is included within the model container in bf16 format, and a couple of million in fp32
Krea2 anime UWU V2 .safetensors
Total parameters: 12,821,780,812
--- Breakdown by format ---
torch.uint8 format: 16,128 parameters
Torch.int8 format: 12,155,092,992 parameters
Format torch.float32: 323,263,052 parameters
Format torch.bfloat16: 343,408,640 parameters
Meanwhile, the base Krea 2 Turbo model in BF16 format has the following configuration:
krea2Turbo_v10_bf16.safetensors
Total parameters: 12,820,073,036
--- Breakdown by format ---
BF16 format: 12,498,501,632 parameters
F32 format: 321,571,404 parameters
have fun)
krea anime 1.7
The following TE is used for inference
It’s also attached to the card, even though the author said it wasn’t necessary to credit them. The encoder is genuinely good.
As for mine... um... well, it looks like this...
Yeah, it’s still taken apart, but at least it works.
krea 2 anime
The model was trained on a high-quality dataset of over 3,000 images; approximately 350 of these consist of my own drawings and AI-generated images that I edited using drawing software.
Qween VL annotation + manual corrections
I consider this model one of the strongest and most powerful I have ever released.
The download fee is nominal. The hardware required to train the Krea 2 model on a large dataset is expensive; even though I can afford to train it locally—my PC costs around $5,000–$6,000 (and I’m currently upgrading it)—it still places a massive strain on the system. I need the funds to enable generation using this model on Civitai and possibly to purchase some cosmetics on the platform. I have previously released a large number of models completely for free, and I hope you understand my decision in this instance.
I was really impressed by this model's performance—specifically its speed and flexibility. I had originally planned a large-scale fine-tuning project for Z-Image, but I changed my mind and decided to focus on Krea 2 instead. At the moment, I have a "raw" dataset of 350,000 images (about 20,000 have been processed so far; manual filtering weeds out roughly 60–70% of the material, leaving a final set of around 100,000 images), as well as a high-standard, carefully curated regularization set. Additionally, I plan to compile a supplementary set of about 20,000 images featuring game characters (primarily from HoYoverse titles) if I determine they are underrepresented in the main dataset, given that the base Krea 2 model doesn't recognize them very well—and is unfamiliar with many of them entirely.
int 8 model
euller\simple 8 step
Image resolution 1.0–1.3–1.5 MP
Description
All examples use a standard workflow that doesn't save the prompt directly for display, but instead stores it within the workflow metadata. Consequently, all the prompts are contained in the files' metadata.