FFF_ANIMA-29B v1

Full fine-tune of OOO_ANIMA-29B v11 base (Anima 2.9B lineage).

OOO_ANIMA-29B v11 base の Full Fine-Tune です(Anima 2.9B 系)。

Files

File	Format	Size	SHA256
FFF_ANIMA-29B-v1.safetensors	bf16	5.84 GB	63920103cd6d2b1ef5859c1715d3a81443f0fb42d84c07686231c6cd8bb18742
FFF_ANIMA-29B-v1_int8_convrot.safetensors	INT8 ConvRot	3.16 GB	86431a4c3aa9dad494335fd18503123cf5f0c0d53249a4c9bc3fd42e348651c6

Required separately / 別に必要なもの:

Part	File	Where to get it
VAE	OOO_Anima_VAE_v3.5.safetensors	oron1208/OOO_ANIMA-29B-V10 → VAE/
VAE (alternative)	qwen_image_vae.safetensors	circlestone-labs/Anima → split_files/vae/
Text encoder	qwen_3_06b_base.safetensors	circlestone-labs/Anima → split_files/text_encoders/

Recommended prompt

Trigger next to the quality tags / trigger を品質タグの隣に置く形:

<your tags>, @fyfnt, masterpiece, best quality, high quality, score_9, newest, <rating>

Or trigger at the head / または先頭に置く形:

@fyfnt, <your tags>, masterpiece, best quality, high quality, score_9, newest, <rating>

Use one or the other; the trigger goes in once. The official Anima card recommends the quality tags as a prefix at the head of the prompt, but this model was trained with them at the end of the caption, so the end is the intended position (the head still works). <rating> is one of safe / sensitive / nsfw / explicit; the samples below use safe, some weighted as (safe:2.0).

どちらか一方を使います(trigger は 1 回)。公式の Anima は品質タグをプロンプトの先頭に置く流儀ですが、このモデルは caption の末尾に置いて学習しているので、末尾が想定の位置です(先頭でも動きます)。<rating> には safe / sensitive / nsfw / explicit のいずれかを入れます。下の生成例は safe で、一部は (safe:2.0) と強調しています。

Recommended negative prompt

worst quality, low quality, score_1, score_2, score_3, early, old, realistic, 3d, photorealistic, sketch, blush, embarrassed, empty eyes, ringed eyes, light particles, artifacts, copyright name, japanese text, page number, bad anatomy, bad hands, border

Optional. Add any of these as needed, for example when safe alone leaves more skin than you want / 必要に応じて適宜足してください(例: safe だけでは露出が下がりきらないとき):

cleavage, underwear, lingerie, bra, panties, see-through clothes

Recommended settings

Sampler	Steps	CFG	Shift	Size
ER SDE (or Euler)	25-30	4	3	1536x1536

1536x1536 is the training resolution of the base model and the safest choice. Other aspect ratios of about the same pixel count work as well: 1344x1792 for portrait, 1792x1344 for landscape. The samples below use those two. Going much above that area makes the style unstable.

1536x1536 が base モデルの学習解像度で、一番安全です。画素数がほぼ同じ比率なら同様に使えます(縦 1344x1792、横 1792x1344)。下の生成例はこの 2 つです。これより大きくすると画風が不安定になります。

Samples

Each sample started from the recommended prompt, negative prompt and settings above and was then adjusted per image. Treat them as starting points rather than a fixed recipe.

生成例はいずれも上の推奨プロンプト・ネガティブ・設定を土台に、画像ごとに手を入れて生成しています。固定のレシピというより出発点として見てください。

Lineage

Anima (CircleStone Labs) → Anima-2.9B (Gazingstars123) → OOO_ANIMA-29B v11 base (oron1208380) → FFF_ANIMA-29B v1

License: CircleStone Labs Non-Commercial License v1.2 (LICENSE.md)

Also on Civitai: https://civitai.com/models/2935633

---

English

One trigger word brings up a consistent soft-shading style, and faces do not blush unless you ask for it.

Which file: the bf16 file is the reference weights, for further training, merging, and generation when VRAM allows. The INT8 ConvRot file is for everyday generation in ComfyUI or Forge Neo; it was quantized from the bf16 file with convert_to_quant (--int8 --convrot --anima, ComfyUI quant metadata) and its output is practically identical. Both contain the DiT only (net.*, 40 blocks, LLM adapter included), so the VAE and text encoder listed above are needed as with OOO_ANIMA-29B.

Usage: the trigger @fyfnt can sit at the head of the prompt or next to the quality tags at the end, whichever suits the way you write prompts. Both were checked over the same 15 prompts and seeds: the style, colors and faces come out the same, while small details such as the outfit shift with the wording. The negative prompt and settings are starting points rather than requirements. The trigger of the upstream OOO_ANIMA refined model has no effect here.

Blush: about 20% of images on a fixed 30-image benchmark when nothing in the prompt asks for it (the base model on the same prompts: 13%). Write blush in the prompt when you want it.

Safety tags safe / sensitive / nsfw / explicit behave as in the base model. As with the base, safe on its own is not a guarantee: swimsuits or underwear still turn up occasionally. The optional negative tags above are the ones that worked here.

Training: full fine-tune of OOO_ANIMA-29B v11 base with kohya_ss sd-scripts.

License: this model is a Derivative under the CircleStone Labs Non-Commercial License v1.2, inherited from the models above. Non-commercial use only: no paid generation services, no selling access, no revenue from hosting the model. You may share and modify it, but keep the license file and this attribution with any redistribution. Generated images are yours. Please do not re-upload this model to generation services that charge users; re-uploads can be checked against the SHA256 above.

---

日本語

trigger を 1 語置くだけで柔らかい塗りの画風が出て、指示しない限り赤面しません。

どちらのファイルを使うか: bf16 が基準の重みで、追加学習・マージ・VRAM に余裕があるときの生成に使います。INT8 ConvRot は ComfyUI / Forge Neo での普段の生成用で、bf16 から convert_to_quant(--int8 --convrot --anima、ComfyUI 量子化メタデータ付き)で量子化したものです。出てくる絵はほぼ同じです。どちらも DiT 本体だけ(net.*、40 block、LLM adapter 込み)なので、OOO_ANIMA-29B と同じく上の VAE と text encoder が別に要ります。

使い方: trigger @fyfnt は、プロンプトの先頭でも、末尾の品質タグの隣でも構いません。普段の書き方に合う方でどうぞ。同じ 15 プロンプト・同じ seed で両方を比べてあり、画風と色と顔立ちは変わりません。服などの細部は語順で少し動きます。ネガティブと設定は出発点で、必須ではありません。上流の OOO_ANIMA refined の trigger はこのモデルでは効きません。

赤面: 固定 30 枚の評価で、プロンプトに何も書かないときの赤面率は約 20% です(同じプロンプトの base は 13%)。blush と書けば赤面します。

rating タグ safe / sensitive / nsfw / explicit は base と同じ意味で効きます。ただし base と同様に safe は絶対ではなく、ときどき水着や下着が出ます。気になるときは上のオプションの語をネガに足すと抑えられます。

学習: OOO_ANIMA-29B v11 base を kohya_ss sd-scripts で Full Fine-Tune しました。

ライセンス: このモデルは上流から継承した CircleStone Labs Non-Commercial License v1.2 の Derivative です。非商用に限ります(有料の生成サービスでの提供、アクセスの販売、ホスティングによる収益は不可)。再配布と改変はできますが、ライセンスファイルとこの帰属表示を一緒に付けてください。生成した画像は利用者のものです。利用者に課金する生成サービスへの転載はしないでください。転載は上の SHA256 で照合できます。