General description
Style LoRA for CircleStone Anima, based on illustrations by dodobo.
Soft, sketchy anime portraits: loose hatched lines on hair and clothes, heavy dark lashes, matte eyes, a light blush on pale skin, a muted palette and plain light or dark backgrounds.
At its best on character portraits: upper body and cowboy shots of known characters, quiet expressions, a flower, a cup of tea, a phone selfie. Scenes with backgrounds work too (a snowy forest, a garden, a city at night).
This is a style, not a character: name the character you actually want.
Not an official model by the artist
──────────────────────────────────────────────────────────
About this version — Anima
Base model: Anima (anima-base-v1.0)
Trigger: @dodobo (at the very start of the prompt, followed by a comma, then regular Danbooru tags)
No style tags needed.
LoRA strength: 1.0
Hires fix is part of the recipe, not an option (see below).
Example prompt
@dodobo, safe, masterpiece, best quality, score_7, 1girl, solo, mikasa ackerman, short hair, black hair, grey eyes, red scarf, brown jacket, white shirt, upper body, looking at viewer, snow, snowing, outdoors, winter
Recommended settings
Sampler: DPM++ 2M SDE, scheduler simple, 25 steps, CFG 5
Size: 832×1248 (2:3)
Hires fix: lanczos up to 1216×1824 (or 1.5×), 20 steps, denoise 0.35, with the LoRA still loaded at 1.0.
The first pass sets the composition; the hires pass is where the artist's line work, lashes and hatching appear. If the hires pass runs without the LoRA, Anima redraws the faces in its own style (glossy eyes, no hatching), so keep the LoRA on both passes. Going above 1.5× tends to break details.
Negative prompt:
worst quality, low quality, score_1, score_2, score_3, jpeg artifacts, logo, watermark, signature, twitter username, bad anatomy, bad hands, missing finger, extra digits, fewer digits, nsfw, nude, explicit
Tips
Give clothes and hair a colour. The artist dresses almost everyone in black, so without a colour tag outfits tend to come out dark (a red bodysuit or a brown jacket can turn black). red bodysuit, brown jacket, white sundress fixes it.
Chainsaw Man characters: the artist draws a lot of Reze and Makima, and Makima's ringed eyes can spread to others. For Power add yellow eyes, cross-shaped pupils, blonde hair; for Denji add blonde hair (faint rings may still show).
Portraits are the home ground. The artist mostly paints upper-body portraits facing the viewer, and the LoRA leans that way. Full-body action and strict side views work, but follow the prompt less strictly and look closer to base Anima.
──────────────────────────────────────────────────────────
Known issues
Faces sometimes get a warm tan shadow, mostly with sunset or afternoon light in the prompt.
Large flat areas come out smoother than in the original paintings; the style lives mostly in the lines and the eyes.
Trained on Anima Base v1.0 only. Behaviour on Anima Turbo and third-party merges is unknown.
Showcase images: first pass 832×1248, hires to 1216×1824 at denoise 0.35, LoRA 1.0 on both passes. Full prompts are in the image metadata.
──────────────────────────────────────────────────────────
Training
187 images; duplicates and redundant alternative versions removed, every image with a watermark or signature excluded entirely (no cropping or inpainting); portrait, square and landscape mixed
Short captions: trigger + rating + character names + people count, framing, hair and eye colour, a few key items, background type; names checked by hand; nothing about style, palette or lighting, no quality tags
Anima LoRA, rank 64 / alpha 64, incl. AdaLN modulation; DiT trained, text encoder and LLM adapter frozen · AdamW8bit, LR 1e-4, constant with 5% warmup · batch 1 · 1024 res, buckets 512–1536 · sigmoid timesteps, scale 1.3 · caption dropout 5% · 12 epochs (7 + 5 continued from epoch 7)
Release file: exact 50/50 merge of the weight deltas of epochs 8 and 12, compressed from rank 128 back to rank 64 by SVD (99.9% of the delta kept). Epoch 12 alone had the most detail but pushed poses towards front-facing portraits; the merge keeps most of the detail and follows poses better
Picked by blind comparisons of epochs 1–12 and five merges on 8 portrait scenes and 4 pose scenes, then hires checks, image by image at full size