General description
Style LoRA based on illustrations by leorlontyan.
Warm golden backlight, a soft yellow-green rim outline around every figure, olive and brown palettes, sparkles, big expressive eyes, wide grins and blushes.
Available for four bases — pick the version that matches your model:
Anima v2 — trigger @lrlt
Illustrious — trigger @lrlt
MOSS-Image — trigger @lrlt. (new — natural-language prompts)
Krea 2 — trigger @lrlt. (new — natural-language prompts, restyles the whole frame)
This is a style, not a character — name the character you actually want.
Not an official model by the artist
──────────────────────────────────────────────────────
About this version — Anima v2
Trigger: @lrlt at the very start, then regular Danbooru tags. No style tags needed.
Prompt template:
@lrlt safe, masterpiece, best quality, score_7, <your tags>
Negative:
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, nsfw, nude, explicit
LoRA strength: 1.0
Recommended:
DPM++ 2M SDE, simple, 25 steps, CFG 5, 832×1248 (2:3)
Hires fix 1.5×, lanczos, 20 steps, denoise 0.35 — sharpens lines, eyes and small details; the style stays the same.
Tips
sparkle, light rays in the prompt add twinkling stars, light beams and big grins.
Group scenes shine: characters eating, playing music, studying or cooking together.
If a secondary character comes out wrong, add their hair colour.
What changed since v1
No style tags needed: v1 relied on six style tags; v2 draws the style from the trigger alone.
No more signature: signatures were removed from the training images; in our test renders the signature never appeared.
New trigger @lrlt replaces @leorlontyan, which the model sometimes wrote into the picture.
Characters keep their features (e.g. Makima's ringed eyes, which v1 sometimes lost).
Stronger style: rank 32 with 10 epochs (v1: rank 16 with 4 epochs).
Epoch 10 of 12 (1030 steps), not the final epoch.
Training
Base: Anima Base v1.0. DiT with AdaLN trained, text encoder and LLM adapter frozen.
Rank/alpha 32/32 · 103 images, artist signatures removed · short tag captions · 1024 res, buckets 512–1536 · AdamW8bit · LR 1e-4, constant, 5% warmup
Trained on Anima Base v1.0 only. Behavior on Anima Turbo and third-party merges is unknown.
──────────────────────────────────────────────────────
About this version — Illustrious
Trigger: @lrlt right after the quality tags, then regular Danbooru tags. No style tags needed.
Prompt template:
masterpiece, best quality, @lrlt, <your tags>
Negative:
lowres, bad quality, worst quality, worst detail, bad anatomy, bad hands, extra fingers, watermark, signature, censor, nsfw, nude
LoRA strength: 1.2 (1.0–1.4; at 1.4 faces can sink into shadow)
Recommended:
Euler a, normal, 30 steps, CFG 5.5, 832×1216
Hires fix 1.5×, lanczos, 20 steps, denoise 0.42 — the style comes out noticeably stronger with it.
Tips
Brighter and more poppy than Anima, with less of the golden backlight and olive palette.
Everything turns golden hour — even "midday, clear sky" gets a pink sky and orange sand. That is part of the style.
Epoch 12 of 20 (732 steps), not the final epoch.
Training
Base: WAI-Illustrious-SDXL v16. UNet and text encoder trained.
Rank/alpha 32/32, conv 32/32 · the same 103 images · short tag captions · 1024 res, buckets 512–2048 · batch 2 · AdamW8bit · UNet LR 3e-4, TE LR 5e-5, cosine
Tested on WAI-Illustrious-SDXL v16 only.
──────────────────────────────────────────────────────
About this version — MOSS-Image
Trigger: @lrlt. at the very start, then describe the scene in plain sentences. No tags, no style words needed.
Prompt template:
@lrlt. <who or what>, <what they are doing>, <where>, <light and mood>.
Example: @lrlt. Three friends sit at a wooden table in a tavern: on the left a blond man in a green vest, in the middle a red-haired woman in a white blouse raising a mug, on the right a dark-skinned man with a black beard in a blue coat, all laughing.
Negative:
nsfw, worst quality, bad anatomy, watermark, jpeg artifact, signature
LoRA strength: 1.0
Recommended:
MOSS-Image SFT for generation. The LoRA was trained on Base; all tests and showcase images use SFT int8.
er_sde, simple, 25 steps, CFG 3, shift 5, skip layer 22
1024×1536, no hires
Tips
The style lands mostly on the characters; backgrounds keep more of MOSS's own look.
Ask for an emotion — a grin, a laugh or a blush suits this style best.
Step 1400 of 2400, not the final step.
Training
Base: MOSS-Image Base (int8). LoRA on DiT attention and MLP layers, text encoder frozen.
Rank/alpha 16/16 · 101 images, each with a long and a short natural-language caption · about 1 MP, native aspect · AdamW · LR 2e-4, constant, 20-step warmup · KohakuMossTrainer
Tested on MOSS-Image SFT int8 in ComfyUI only.
──────────────────────────────────────────────────────
About this version — Krea 2
Trigger: @lrlt. at the very start, then describe the scene in plain sentences. No tags, no style words needed.
Prompt template:
@lrlt. <who or what, with hair, eyes and clothes>, <what they are doing>, <where>, <light and mood>.
Example: @lrlt. A young woman with short black hair and blue eyes in a big red sweater, her fists near her cheeks, grinning at the viewer, plain white background.
LoRA strength: 1.0 — 0.8 keeps the pencil line and can clean up busy scenes.
Recommended:
Krea 2 Turbo, 8 steps, CFG 1, euler / simple, 832×1248
Hires pass: lanczos to 1216×1824, same prompt and seed, 8 steps, denoise 0.25 — sharper lines and more grain. All showcase images use it and keep their prompts and settings.
Tips
Describe your character. With no look in the prompt it falls back on the artist's usual design: yellow eyes, a mustard hoodie, a ponytail. Hair, eyes and clothes in the prompt are followed.
Plain backgrounds and close-ups show the pencil line, colored outline and translucent paint best.
Small warm moments suit it: a tough hero with a kitten, a huge parfait, a bowl of ramen. Grins, blushes and sparkles do the rest.
Unlike the MOSS version, it restyles the whole frame, backgrounds included.
Famous characters keep their canon look. Less common ones can drift to a similar famous character, so describe them in full.
The same seed and settings keep the shot, but small details change between runs, so copying a showcase image gives a close match, not an exact one.
Two files, pick either: fp32 is the original, bf16 is half the size. In our test they were no further apart than two runs of the same file.
Epoch 12 of 12 (2424 steps), the final epoch — earlier epochs keep less of the artist's line.
Training
Base: Krea 2 Raw. LoRA on the DiT.
Rank/alpha 32/32 · 101 images, each with a long and a short natural-language caption · 1024 res, buckets · batch 1 · AdamW8bit · LR 1e-4, constant · flow shift 2.5 · Musubi Tuner
Tested on Krea 2 Turbo (fp8) in ComfyUI only.