Krea 2 Anime-tuned

An anime full fine-tune of Krea 2 (all 12.8B DiT parameters) on ~196k curated anime illustrations. This page has two versions that work best together:

• Turbo: Krea 2 Turbo plus the full weight delta of the Raw fine-tune. Use it alone with 8 steps and no CFG.

• Raw: the fine-tuned Krea 2 Raw (EMA weights). Use it alone with CFG at ~1 MP, as the first stage of the hybrid sampler, or as a base for LoRAs.

Recommended: the hybrid sampler. Raw generates the first 2 of 8 steps with CFG, then Turbo finishes the last 6 without CFG. All sample images were made this way at 2K.

Files (ComfyUI)

• This model (fp8_scaled, 13.1 GB per version) → models/diffusion_models/. BF16 files and diffusers weights are on Hugging Face: Turbo · Raw.

• Text encoder qwen3vl_4b_fp8_scaled.safetensors → models/text_encoders/, from Comfy-Org/Krea-2.

• VAE qwen_image_vae.safetensors → models/vae/, from Comfy-Org/Krea-2.

• Example workflows (hybrid / Turbo / Raw): example_workflows.

Sampler settings

• Hybrid (both versions): one 8-step euler / simple schedule. Apply ModelSamplingFlux with base_shift = max_shift = 1.15 to both models. KSamplerAdvanced #1 (Raw): steps 0 → 2, cfg 2.5, empty negative prompt, add_noise enable, return_with_leftover_noise enable. KSamplerAdvanced #2 (Turbo): steps 2 → 8, cfg 1.0, add_noise disable, same seed. On a 24 GB L4 with both fp8 files, 2K takes ~2 min per image including model swaps.

• Turbo alone: 8 steps, euler / simple, cfg 1.0, fixed shift 1.15 at every resolution (ModelSamplingFlux 1.15 / 1.15). Supports 1K to 2K.

• Raw alone: 28 steps, euler / simple, cfg 5.5, empty negative prompt. Use ModelSamplingFlux with base_shift 0.5 / max_shift 0.90625, and set width / height to the target size. Best around 1 MP.

Prompting: capgen

The training captions are long natural-language descriptions in one consistent style, for example:
anime illustration of a girl, upper body portrait, long pale pink hair with blue tips, ..., clean line art, soft shading.
Short tag lists work, but prompts in this style work best.

capgen turns a few Danbooru tags into a full caption in this style. It needs no LLM, runs on CPU, and takes milliseconds per caption. Ways to use it:

• ComfyUI custom node with tag autocomplete. Search "capgen" in ComfyUI-Manager. The example workflows already include it.

• Browser demo, which runs locally in your browser.

• Python library.

Tip: capgen often adds a white background and leans toward Danbooru's sensitive rating (about PG-12 to PG-15). To exclude something, add -tag, e.g. 1girl, -white background, -cleavage. The sample images used this.

The samples were rendered on the author's TPU serving stack with the hybrid sampler, so the same prompt and seed will not give identical pixels in ComfyUI.

Training

• Data: 196,099 anime illustrations from Danbooru and Gelbooru. Only general and sensitive ratings, at least 1 MP, at least one girl, no male focus, and watermarked or text-heavy images removed. They are the top ~200k of ~906k candidates, ranked by a preference model trained on the author's pairwise picks (anime-aesthetic-nafcsd), so they reflect one person's taste.

• Captions: VLM-written natural-language captions in a single style.

• Recipe: full fine-tune of Krea 2 Raw, 9,192 steps. 4 epochs at 512 px, then 2 epochs at 1024 px, with EMA 0.999. Trained in JAX on 8× AMD MI355X in ~8.3 h. Code: Krea-2-JAX.

• Blind A/B tests by the author (50 prompts each): Raw-tuned beat base Krea 2 Raw 47 : 3, and Turbo-tuned (at 0.7× delta) beat base Krea 2 Turbo 46 : 4.

Limitations

• The training data is almost all single-girl illustrations. Boys, multi-character scenes and non-anime styles are weak.

• Raw alone at 2K can sometimes produce face artifacts. For 2K, use Turbo or the hybrid sampler.

• Niche Danbooru tags that never appeared in the captions may not be understood.

• Training used only general and sensitive images, but outputs can still be suggestive.

License

Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing.
This model is a modified version of the Krea Model (fine-tuned by ricecake). It is not an official Krea product and is not endorsed by Krea.

By downloading or using this model you agree to be bound by the Krea 2 Community License Agreement (copy and NOTICE file also in the Hugging Face repo) and its Acceptable Use Policy. Commercial use is allowed only under the license's revenue threshold (§2.3). If you deploy the model, you must use content filters (§4.2) and disclose AI-generated outputs where required (§4.3).