Takashi Takeuchi · modern style (Tsukihime Remake / FGO) | Anima

Style LoRA for Anima (2B) trained on Takashi Takeuchi's modern official art, roughly 2015 onward, mainly from Tsukihime Remake and FGO official art.

Quick how to use

• Trigger: @tkmstyle. (with the dot), right after the quality tags.

• Strength: 1.0

• Sampler: er_sde, simple scheduler, 30 steps, CFG 4.

• Prompt: masterpiece, best quality, score_9, score_8, score_7, year 2025, newest, highres, absurdres, official art. @tkmstyle. followed by your tags (character, series, outfit, pose, framing) and one or two plain-English sentences about the scene and the light. The captions were written in natural language, so that sentence matters.

• Negative: worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, bad anatomy, bad hands, missing fingers, extra fingers

My personal setup

After a lot of testing, this is what I use and what all the showcase images were made with. Each one has its full ComfyUI workflow embedded, so you can drag it into ComfyUI.

• Anima-Turbo v1.1 instead of Base, plus the official anima-rl-v0.1 LoRA at 1.0 (Anima-Official-LoRAs).

• euler, simple, 12 steps, CFG 1, about 1 MP.

• NegPiP (CLIPNegPip node from ComfyUI-ppm), because the negative prompt does nothing at CFG 1. I add this at the end of the positive: (worst quality:-1.5), (low quality:-1.5), (bad anatomy:-1.5), (bad hands:-1.5), (extra fingers:-1.5), (extra limbs:-1.5)

• Hires: 4x IllustrationJaNai V3 denoise, resize to 1.5x (1248x1824), Anima Tile & Repair LLLite v2.0 at 0.8 (end 0.9), then the same turbo for 5 steps at denoise 0.40. About a minute per image on an 8 GB laptop GPU.

Training

129 hand-picked official illustrations with long natural-language captions, trained on Anima Base v1.0. Rank 32, 20 epochs, low-noise-biased timesteps and the adaLN modulation layers trained as well, which keeps the base model's lighting instead of washing it out.

This is a style LoRA, not a character LoRA: it works with any character. Since it was trained on official art of around 40 Type-Moon characters, their likeness (face, hair, outfits) may come out boosted compared to the base model, especially for the most frequent ones: Arcueid Brunestud (best girl shall always be first), Artoria Pendragon (and her Alter), Mash Kyrielight, Ciel, Jeanne d'Arc and Ryougi Shiki.

Have fun!