General description
Style LoRA based on illustrations by iwamushi (岩虫).
Rough uneven linework, deep shadows, real paper grain, muted print colors.
Available for four bases — pick the version that matches your model:
Anima v10 — trigger @iwamushi (new, recommended)
Anima v8 — trigger @iwamushi (previous version, flatter and warmer)
Anima v5 — trigger @iwamushi (previous version, darker and rougher)
Illustrious — trigger iwamushi
Pony — trigger iwamushi
Krea 2 — trigger @iwamushi
MOSS-Image — trigger @iwamushi (new)
This is a style, not a character — name the character you actually want.
Not an official model by the artist.
──────────────────────────────────────────────────────
About this version — Anima v10
Trigger: @iwamushi at the very start, then regular Danbooru tags. No style words needed.
Prompt template:
@iwamushi safe, masterpiece, best quality, score_7, <your tags>
Negative:
worst quality, low quality, lowres, artist name, signature, text, (bad anatomy:1.2), (bad hands:1.1), missing finger, (extra digits:1.4), fewer digits, disfigured, mutation, 4 fingers, (6 fingers:1.4)
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 — use it. The paper grain and the rough line come out at the higher resolution; without hires the picture looks noticeably simpler.
Color control
The default palette is cool and nocturnal — great for rain, snow, neon and night scenes. You can steer it:
Warm and brown scenes: describe a warm light source (morning light, sunset, lamp light) and add (warm colors:1.4), brown theme.
Dark red scenes: add warm colors, red theme.
Night, rain and winter scenes need nothing — leave the colors alone.
Lowering the LoRA strength barely changes the color, so keep it at 1.0.
The prompts of every showcase image are attached, so you can see where the color tags are used and where they are not.
Tips
rough uneven linework, halftone texture at the end of the prompt adds more dust and grain. Optional — some people like it, some don't.
The artist's signature can still show up now and then, most often in Re:Zero scenes (that's where the artist signs most). artist name, signature, text in the negative removes most of it — reroll the rest.
What changed since v8
Drawn, not flat: real light and shadow, dark masses and depth instead of flat fills and even gradients.
The artist's paper grain: uneven, chaotic texture on skin, cloth and sky — with hires it is close to the original scans.
More life in the scenes: emotions, poses and small props; scenes are composed instead of falling back to one template.
Prompts are followed as well as by base Anima, and color tags work without breaking the style.
Fewer signatures: in our side-by-side tests (same seeds and negative) v8 wrote the artist's name on several images, v10 on none.
The last images of the gallery are side-by-side comparisons: the same prompt, seed and settings with v8 (left) and v10 (right).
A weighted average of epochs 8 and 10 (80/20) of the same run — the averaged epochs keep the deep shadows of epoch 8 and add the finer detail of epoch 10.
Training
Base: Anima Base v1.0. DiT with AdaLN trained, text encoder and LLM adapter frozen.
Rank/alpha 64/64 · 112 images · 1024 res, buckets 512–1536 · AdamW8bit · LR 1e-4, constant, 5% warmup · 12 epochs, release = 0.8 × epoch 8 + 0.2 × epoch 10 (exact merge of the LoRA deltas, refactored to rank 64)
Trained on Anima Base v1.0 only. Behavior on Anima Turbo and third-party merges is unknown.
──────────────────────────────────────────────────────
Anima v8
Trigger: @iwamushi (put it at the start of the prompt)
Recommended style tags:
rough uneven linework, halftone texture
LoRA strength: 1.0
Tips
Don't add color tags like "olive muted palette" — they tint clothes olive-yellow. The LoRA already carries the palette.
Avoid the word "dirty" in texture tags — it paints stains on clothing.
If you ever see an "@iwamushi" signature in a corner, add signature to the negative prompt.
What changed since v5
Dataset grew from 79 to 112 images: new works added, low-res previews replaced with high-res originals, off-style pieces (neon, UI-heavy, 3D-like) removed.
Much less Re:Zero bias — characters stay on-model more often (hair color, outfits).
Noticeably more detail: backgrounds, clothing, lighting and shadows.
Handles complex scenes better: multiple characters, crowds, detailed interiors.
Warmer, sandier tone, closer to the artist's originals.
Epoch 7 of 8 (784 steps), not the final epoch.
Training
Base: Anima Base v1.0. DiT with AdaLN trained, text encoder and LLM adapter frozen.
Rank/alpha 16/16 · 112 images · 1024 res, buckets 512–1536 · AdamW8bit · LR 5e-5, constant, 5% warmup
Trained on Anima Base v1.0 only. Behavior on Anima Turbo and third-party merges is unknown.
──────────────────────────────────────────────────────
Anima v5
Previous version. Superseded by v8, which has more detail and handles complex scenes better. v5 is kept for its darker, rougher look.
Trigger: @iwamushi (put it at the start of the prompt)
Optional style tags: flat print color, olive muted palette, rough uneven linework, dirty halftone texture
Epoch 6 of the v5 run (474 steps), not the final epoch.
Training
Base: Anima Base v1.0. DiT with AdaLN trained, text encoder and LLM adapter frozen.
Rank/alpha 16/16 · 79 images · 1024 res, buckets 512–1536 · AdamW8bit · LR 5e-5, constant, 5% warmup
Trained on Anima Base v1.0 only. Behavior on Anima Turbo and third-party merges is unknown.
──────────────────────────────────────────────────────
Illustrious
Trigger: iwamushi (put it first, then regular Danbooru tags). No style tags needed.
Recommended:
LoRA 1.0 (up to 1.2)
euler_a, 30 steps, CFG 5–6
Faces and expressions
Expression tags work, but the style draws dim eyes and shadows on the face, partly because the artist draws eyes that way, so smiles can look tired.
Add empty eyes, shaded face to the negative prompt for brighter eyes and lighter faces. It helps most characters, not all, and can slightly change the image.
Lowering the LoRA strength doesn't help — keep it at 1.0.
Epoch 16 of 24 (640 steps), not the final epoch.
Training
Base: Illustrious XL v2.0. UNet and text encoder trained.
Rank/alpha 32/16, conv 16/8 · 79 images · 1024 res, buckets 512–2048 · batch 2 · AdamW8bit · UNet LR 1e-4, TE LR 1e-5, constant, 5% warmup · min SNR 5 · noise offset 0.03
Tested on WAI-Illustrious v16, Hyphoria v0.02 and Nova Anime XL il v1.9.
──────────────────────────────────────────────────────
Krea 2
Trigger: @iwamushi. at the very start, then describe the scene in plain sentences. No tags, no style words needed.
Prompt template:
@iwamushi. <who or what>, <what they are doing>, <where>, <light and mood>.
Example: @iwamushi. Guts from Berserk in battered black armor sits by a small campfire at night, his huge sword leaning against a rock, embers rising into the dark, his face lit only by the fire.
LoRA strength: 1.0 (0.8–1.2)
Below 0.8 Krea starts writing "@iwamushi" in a corner — it reads the trigger as a social-media handle. The LoRA at full strength suppresses it.
Above 1.2 outfits and small details drift toward the training images.
Recommended:
Krea 2 Turbo, 8 steps, CFG 1, euler / simple
Native resolution, about 1 MP: 832×1248, 1248×832 or 1024×1024. No hires pass needed.
Tips
Atmosphere is where it shines: dusk, rain, firelight, fog, night streets.
Turbo gives very similar images across seeds — change the wording of the prompt for variety instead of rerolling.
Epoch 6 of 12 (1344 steps), not the final epoch — later epochs start copying signatures and outfits from the dataset.
Training
Base: Krea 2 Raw. LoRA on the DiT.
Rank/alpha 32/32 · 112 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
Tested on Krea 2 Turbo (fp8).
──────────────────────────────────────────────────────
Pony
LoRA strength: 1.0 — a balanced default for both base Pony and merges.
Pony Diffusion V6 XL: ~0.8 if colors get heavy or red.
AutismMix and other Pony merges: up to 1.5 for a stronger style.
Trigger: iwamushi (right after the score tags). No style tags needed.
Prompt template:
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_anime, iwamushi, <your tags>
Recommended:
DPM++ 2M Karras, 30 steps, CFG 5–6, clip skip 2
A 1.5× hires pass at denoise ~0.45 makes a big difference on Pony.
Negative: score_6, score_5, score_4, source_pony, source_furry, source_cartoon, 3d, realistic, signature, artist name
Tips
Too much strength on base V6 (above ~1.2) turns everything muddy red-brown — lower it.
Closest to the artist's look is still Anima v8; the Pony version is a lighter take on the style.
Epoch 16 (944 steps), final epoch.
Training
Base: Pony Diffusion V6 XL. UNet and text encoder trained, clip skip 2.
Rank 32, conv 16, alpha scaled ×1.3 after training (32/20.8, conv 16/10.4) so 1.0 is the balanced strength · 112 images, Danbooru tags · 1024 res, buckets 512–2048 · batch 2 · AdamW8bit · UNet LR 1e-4, TE LR 1e-5, constant, 5% warmup · min SNR 5 · noise offset 0.03
Tested on Pony Diffusion V6 XL and AutismMix Pony.
──────────────────────────────────────────────────────
MOSS-Image v1
Trigger: @iwamushi. at the very start, then describe the scene in plain sentences. No tags, quality words or style words needed.
Prompt template:
@iwamushi. <who or what>, <what they are doing>, <where>, <light and mood>.
Example: @iwamushi. Guts from Berserk in battered black armor sits by a small campfire at night, his huge sword leaning against a rock, embers rising into the dark, his face lit only by the fire.
Negative:
nsfw, worst quality, bad anatomy, watermark, jpeg artifact
LoRA strength: 1.0 — 0.8 only weakens the style.
Recommended:
MOSS-Image SFT for generation. The LoRA was trained on Base, and all tests and showcase images use SFT int8.
er_sde, simple, 25 steps, CFG 3, shift 5, skip layer 22
1024×1536, no hires — this size gave the firmest line in our tests. 832×1248 is softer, 1248×1872 adds stray objects, and a hires pass added little.
All showcase images keep their settings, so you can copy them directly.
Faces and expressions
Expressions follow the prompt, but the style draws matte eyes with few highlights, so a calm face can look tired.
Ask for an emotion: a smile, a grin or a laugh brings the face to life.
Bright clear eyes with soft white highlights. at the end of the prompt helps on some images, but it can put a star in the pupil.
empty eyes, shaded face in the negative does not help on MOSS.
Tips
Name the colors and the light in a sentence or two at the end of the prompt, for example The colors are pale green, off-white and dusty blue. The lighting is even and diffused, casting soft shadows. Daytime scenes get a more muted, printed look. Night scenes change less, and cool blue light can tint the skin.
Speech bubbles and lettering can appear on their own, especially in shouting close-ups and scenes with bags or signs. speech bubble, text in the negative removes some of them — reroll the rest.
Characters and anatomy come from MOSS. The LoRA adds the paper grain, the ink line and the muted paint. For the closest match to the artist's hand, use Anima v10.
Step 1800 of 3000, not the final step.
Training
Base: MOSS-Image Base (int8). LoRA on DiT attention and MLP layers, text encoder frozen.
Rank/alpha 8/8 · 112 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.