Skip to main content
Back

ricecake Fine-Tunes Krea 2 on 196K Anime Images Across All 12.8B Parameters

A solo developer fully fine-tuned Krea 2's 12.8B base model on 196k anime illustrations in about 8 hours, then transferred the result to the fast 8-step Turbo checkpoint.

Model, Full Access
  • Krea-2-Anime-tuned-Raw: full 12.8B fine-tune of Krea 2 Raw on 196k anime illustrations.
  • Trained in about 8.3 hours on 8 AMD MI355X GPUs using open-source JAX code.
  • The Turbo variant adds the Raw fine-tune's weight delta to Krea 2 Turbo, with no extra training.
  • Recommended hybrid sampler: Raw handles 2 steps with CFG, then Turbo finishes 6. About 2 minutes per 2K image on 24 GB.
  • Won 47:3 against base Raw in the author's own blind tests. Weak on male, multi-character, and non-anime images.
  • Free under the Krea 2 Community License, which requires an enterprise deal above $1M revenue or 50 seats.

Developer ricecake has released Krea-2-Anime-tuned-Raw, a full-parameter fine-tune of Krea 2 Raw trained on 196,099 anime illustrations. Training updated all 12.8 billion parameters, allowing the checkpoint to serve as a base for inference and additional LoRAs. The release also includes an eight-step Turbo variant, a custom prompt generator, and the JAX training code.

Developers get a domain-specific Raw model for further training and a faster Turbo model for generation. Ricecake built and published the entire stack, making the implementation open to inspection and reuse. The Raw repository had recorded more than 3,600 downloads at the time of reporting.

Core release artifacts
Artifact Purpose Link
Krea-2-Anime-tuned-Raw Full fine-tune and base for downstream LoRAs Raw model
Krea-2-Anime-tuned-Turbo Eight-step inference with the anime weight delta Turbo model
Krea-2-JAX Full-parameter training implementation JAX training code

Raw keeps the training headroom

Krea 2 ships as two checkpoints with the same diffusion-transformer architecture and different training goals. K2 Raw is an undistilled base intended for fine-tuning and post-training research. K2 Turbo uses guidance and timestep distillation to generate an image in eight denoising steps.

Image diffusion begins with noise and refines it over a sequence of steps. Classifier-free guidance, or CFG, strengthens prompt adherence by comparing conditional and unconditional predictions, which usually requires two model evaluations per step. Turbo’s distillation approximates that behavior without CFG during inference, reducing the work required to produce an image.

Krea recommends training on Raw and generating with Turbo. The Raw checkpoint preserves more flexibility for full fine-tuning, while Turbo provides the faster deployment target. Community demand for specialized Krea models is already established: fidel1234/Krea2, an Asian-character LoRA for Raw, has surpassed one million downloads.

196,099 images, 66 accelerator-hours

  • Source data: 196,099 illustrations from Danbooru and Gelbooru, restricted to general and sensitive ratings.