General description
Style LoRA for CircleStone Anima, based on the fan comics of MayngoMay.
Comedic anime-screencap look: flat cel shading, a bold clean outline, simplified details and exaggerated reaction faces.
Empty shocked eyes, deadpan stares, smug half-closed eyes, wavy-mouthed crying. Ask for an emotion and the LoRA turns it into a punchline.
Trained on Re:Zero comics, so Subaru, Emilia, Rem, Ram, Beatrice and Echidna feel right at home, but it works on any character the base model knows.
It can also draw multi-panel vertical comic pages and the artist's yellow outlined subtitles on request (see Tips).
This is a style, not a character: name the character you actually want. Every Re:Zero character in the dataset was named by hand in the captions, so an ordinary black-haired boy stays an ordinary boy and doesn't turn into Subaru.
Not an official model by the artist. Original artist
────────────────────────────────────────────────────────────
Anima
Base model: Anima (anima-base-v1.0)
Trigger: @mngm (at the very start of the prompt, followed by a comma, then regular Danbooru tags)
No style tags needed.
LoRA strength: 1.0 (going higher barely changes the picture)
Example prompt
@mngm, safe, masterpiece, best quality, score_7, 1boy, natsuki subaru, black hair, track jacket, shocked, wide-eyed, empty eyes, open mouth, sweat, mansion hallway, upper body
Recommended settings
Sampler: DPM++ 2M SDE, scheduler simple, 25 steps, CFG 5
Size: 832×1248 (2:3); landscape 1216×832 for single-panel reaction shots
Hires fix: 1.5×, lanczos, 20 steps, denoise 0.35
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, text, english text, speech bubble, subtitled, sound effects, comic, nsfw, nude, explicit
Tips
Prompt the emotion. The funny faces come from expression tags: shocked, wide-eyed, empty eyes, expressionless, half-closed eyes, smug, crying, streaming tears, wavy mouth, anger vein, puffy cheeks, sweat.
Comic pages: remove the text tags from the negative and add comic, 3koma, multiple panels, vertical panels, speech bubble, english text at 832×1536. You can describe each panel (top panel: ..., middle panel: ..., bottom panel: ...). The lettering comes out as gibberish, so re-letter it in an editor.
Subtitles: subtitled, english text, yellow text on a landscape shot gives the artist's "anime subtitle" captions.
────────────────────────────────────────────────────────────
Known issues
In two-character scenes the second character sometimes drops out. Describe both with their hair colour, or use a comic layout.
Not meant for child characters: it can draw them older than they are.
Trained on Anima Base v1.0 only. Behaviour on Anima Turbo and third-party merges is unknown.
Showcase images use hires fix 0.35. Full prompts are in the image metadata.
────────────────────────────────────────────────────────────
Training
219 panels cut from ~110 comic pages along the panel gutters; text-only panels, duplicates and badly cut crops dropped
Character names checked by hand on every panel (the WD tagger missed Subaru on ~20 panels and doesn't know Echidna): Subaru on 72 panels, Echidna on 25, plus Reinhard, Otto, Garfiel and others
Short captions: trigger + rating (safe / sensitive / explicit) + character names + people count, framing, hair colour + text tags (english text, subtitled, speech bubble, sound effects, comic) so the lettering binds to tags you can put in the negative; nothing about style, colours or lighting, no quality tags
Anima LoRA, rank 32 / alpha 32, incl. AdaLN modulation; DiT trained, text encoder and LLM adapter frozen · AdamW8bit, LR 1e-4, constant with 5% warmup · epoch 12 of 12 (2628 steps), batch 1 · 1024 res, buckets 512–1536 · sigmoid timesteps, scale 1.3 · caption dropout 5%
Epoch 12 picked by comparing epochs 8–12 on generic characters (leak test), comedic reaction shots and Echidna, 2 seeds each, image by image