MoeMoeMoss | 20 Artist Styles + deepseek-tan | Anime Illustration LoRA for MOSS Image

Trained on B200 · 192 GB and RTX PRO 6000 · 96 GB
18,500 steps · ~50 hours
Training Base: OpenMOSS-Team/MOSS-Image-SFT | Showcase Inference Base: moss_image_sft_int8_convrot.safetensors (int8)
Architecture: LoKr (f8-v1, 74.33M trainable params, 504 tensors)
Format & File Names: ComfyUI LoraLoaderModelOnly .safetensors (148,746,416 B / ~141.86 MiB) — style21-lokr-f8-step18500-comfy.safetensors (same file as style21-lokr-f8-v1-comfy.safetensors)

This adapter packs 20 distinct anime illustrator styles and 1 built-in character (deepseek-tan) into a single 74.33M-parameter LoKr module for MOSS-Image, trained across two top-tier machines (NVIDIA B200 · 192 GB and RTX PRO 6000 · 96 GB) over 18,500 steps (~50 hours).

Place one of the 20 <|artist|>...<|/artist|> trigger tags on the first line of your prompt—followed by either a natural-language scene description or a MOSS JSON structured layout prompt—to steer brushwork, lighting, and color harmony toward a specific illustrator, or write deepseek-tan to summon the built-in blue-haired whale-girl character.

1. How to Prompt & All 20 Artist Trigger Words

Standard Usage — Line-1 <|artist|>...<|/artist|> Tag + Scene / JSON Body

Place one of the 20 <|artist|>...<|/artist|> tags at the very start of your positive prompt (or concatenate it ahead of your main prompt using a StringConcatenate node in ComfyUI), followed by your natural-language description or MOSS JSON layout block. (You can also omit the tag if you want the adapter's blended default style.)

Keep all underscores (_), hyphens (-), and parentheses (()) exact:

<|artist|>ask_(askzy)<|/artist|>
<|artist|>binggong_asylum<|/artist|>
<|artist|>blade_(galaxist)<|/artist|>
<|artist|>ciloranko<|/artist|>
<|artist|>fkey<|/artist|>
<|artist|>fuzichoco<|/artist|>
<|artist|>happoubi_jin<|/artist|>
<|artist|>hiten_(hitenkei)<|/artist|>
<|artist|>hito_komoru<|/artist|>
<|artist|>jyt<|/artist|>
<|artist|>ke-ta<|/artist|>
<|artist|>mika_pikazo<|/artist|>
<|artist|>ogipote<|/artist|>
<|artist|>piromizu<|/artist|>
<|artist|>quan_(kurisu_tina)<|/artist|>
<|artist|>rin_yuu<|/artist|>
<|artist|>shou937<|/artist|>
<|artist|>torino_aqua<|/artist|>
<|artist|>weri<|/artist|>
<|artist|>wlop<|/artist|>

All 20 Artist Triggers & Corresponding Gallery Reference Images

# Exact Trigger Tag (Line 1) Shown in Gallery (Image #) # Exact Trigger Tag (Line 1) Shown in Gallery (Image #) 01 <|artist|>ask_(askzy)<|/artist|> 00259, 00351, 00371, 00393, 00486 11 <|artist|>ke-ta<|/artist|> — 02 <|artist|>binggong_asylum<|/artist|> (See Article 6-artist sweep) 12 <|artist|>mika_pikazo<|/artist|> 00264, 00286, 00442, 00466 03 <|artist|>blade_(galaxist)<|/artist|> 00181 13 <|artist|>ogipote<|/artist|> — 04 <|artist|>ciloranko<|/artist|> 00206 14 <|artist|>piromizu<|/artist|> — 05 <|artist|>fkey<|/artist|> 00410 15 <|artist|>quan_(kurisu_tina)<|/artist|> 00316 06 <|artist|>fuzichoco<|/artist|> 00153 16 <|artist|>rin_yuu<|/artist|> — 07 <|artist|>happoubi_jin<|/artist|> — 17 <|artist|>shou937<|/artist|> — 08 <|artist|>hiten_(hitenkei)<|/artist|> — 18 <|artist|>torino_aqua<|/artist|> (See Article 6-artist sweep) 09 <|artist|>hito_komoru<|/artist|> 00171, 00391, 00476 19 <|artist|>weri<|/artist|> — 10 <|artist|>jyt<|/artist|> 00496 20 <|artist|>wlop<|/artist|> —

2. Who is deepseek-tan?

deepseek-tan is the anthropomorphized character of DeepSeek (深度求索)—featuring gradient blue wavy hair, whale-fin ears, a curled ahoge, and a deep-blue whale tail. Simply include deepseek-tan in your prompt (paired with any <|artist|> tag) to summon her (see gallery showcase images Moss__00259_.png and Moss__00153_.png).

3. Recommended ComfyUI Settings

A. Exact Showcase Two-Stage Workflow (~3.2 MP, Used Across All 18 Gallery Images)

Every showcase image on this page was generated on the int8 base model (moss_image_sft_int8_convrot.safetensors) with step-18500 (style21-lokr-f8-step18500-comfy.safetensors / style21-lokr-f8-v1-comfy.safetensors) using the following two-stage ComfyUI pipeline:

Parameter Showcase Value Notes Inference Base UNet moss_image_sft_int8_convrot.safetensors (int8) Actual UNet used across all 18 gallery images Style Adapter (Stage 1 & Stage 2) LoraLoaderModelOnly → style21-lokr-f8-step18500-comfy.safetensors strength_model = 1.0 (0.6 – 0.8 for a lighter touch; same file as style21-lokr-f8-v1-comfy.safetensors) ModelSamplingSD3 shift = 5.0 Connected right after the style21 LoKr loader Stage 1 (Base Pass — Style LoKr Only) ~1 MP (0.75 MP on 00206), 25 steps, CFG 5.0, euler + sgm_uniform, denoise = 1.0 Runs without the DMD speed-up LoRA (16 steps on Moss__00410_.png) Upscale Between Stages VAE Decode → 4x-AnimeSharp.pth → lanczos ×0.4375 (net 1.75×) → VAE Encode (×0.5 on Moss__00206_.png) Stage 2 (Refiner Pass — + DMD LoRA) Stack lora_sft_8step_dmd.safetensors @ 1.0 → 10 steps, CFG 1.8, euler + sgm_uniform, denoise = 0.40 – 0.55 Needs an 8-step DMD speed-up LoRA (lora_sft_8step_dmd.safetensors) stacked only on Stage 2. Default denoise = 0.40 (14 images); 0.45 on 00410; 0.55 on 00442 / 00466 / 00476 Final Resolution (~3.2 MP) 1456×2176, 1568×2064 (Portrait) / 2176×1456, 2176×1472, 2064×1568 (Landscape) Standard single-stage 1024×1536 (~1.57 MP) is also supported Negative Prompt Recommended base negative string (plus optional <|texture_and_style_neg|> prefix on JSON layout prompts) See exact negative prompt templates in Section 4 below

B. Standard Single-Stage Workflow (Without Stage 2 Refiner / DMD LoRA)

If you run a single-stage pass on moss_image_sft_int8_convrot.safetensors:

• Resolution: 1024 × 1536 (portrait) or 1536 × 1024 (landscape)

• Steps / CFG: 50 steps (or 25 steps), CFG 5.0

• Sampler / Scheduler / Shift: euler, sgm_uniform, ModelSamplingSD3 (shift = 5.0)

• Adapter Strength: 1.0 (0.6 – 0.8 if you want a subtler blend)

4. Prompt & Negative Prompt Templates

Example 1 — Summoning deepseek-tan + Trilingual Signs (Moss__00259_.png, 1456×2176, Stage 2 denoise = 0.4)

Positive Prompt:

<|artist|>ask_(askzy)<|/artist|>
A warm, sunlit maid cafe scene where Chinese, English, and Japanese signs blend naturally into the environment. In the center foreground, deepseek-tan, a cute blue-haired girl with dark blue whale-fin ears, a curled ahoge, a navy whale tail, and a dark navy-and-white ruffled maid outfit leans warmly over a wooden counter, holding a rustic wooden welcome plaque in her hands that clearly reads "快来用MOSS" in large handwritten Chinese characters. On the brick wall behind her left shoulder, a warm golden vintage neon sign clearly reads "MOSS Please" in glowing English script. On the right side of the counter, a standing wooden menu chalkboard clearly reads "MOSSを使ってね" in neat white Japanese chalk lettering decorated with a little whale doodle.

Recommended Base Negative Prompt (Used Across Natural-Language Showcase Images):

unfinished, incomplete, work in progress, wip, draft, rough sketch, lineart only, uncolored, messy lines, low quality, worst quality, blurry, jpeg artifacts, gigantic breasts, hyper breasts, bad anatomy, extra limbs, extra hands, extra fingers, mutated hands, fused fingers, watermark, signature, artist name, username, twitter handle, patreon logo, youtube logo, social media icon, copyright, text, subtitle, bottom left text, corner text

Example 2 — Line-1 Artist Tag + MOSS JSON Structured Layout Mode (Moss__00466_.png, 1568×2064, Stage 2 denoise = 0.55)

Concatenate the <|artist|>...<|/artist|> tag directly before a MOSS JSON layout object (summary + items with bbox and text fields), and optionally prefix your negative prompt with <|texture_and_style_neg|>:

Negative Prompt for JSON Layout Prompts (00410 / 00442 / 00466 / 00486):

<|texture_and_style_neg|>cluttered background, complex scenery, photorealistic, real human, 3d render, cgi, photograph, monochrome, rough sketch, uncolored patches, camera flash, microscopic pores, film grain, over-sharpened hard edges, plastic skin, ugly, deformed, bad anatomy, extra fingers, extra limbs, distorted face, crooked eyes, blurry text, misspelled text, garbled letters, extra characters, watermark, signature, username, frame, border

5. Architecture & Training Specifications

Item Specification Training Hardware & Scale B200 · 192 GB and RTX PRO 6000 · 96 GB — 18,500 steps · ~50 hours Adapter Type LoKr (f8-v1) Trainable Parameters 74,333,952 (504 tensors) File Size & Names 148,746,416 bytes (style21-lokr-f8-step18500-comfy.safetensors / style21-lokr-f8-v1-comfy.safetensors) Trained Triggers 20 <|artist|>...<|/artist|> tags + 1 character (deepseek-tan)

6. Notes & License

• ComfyUI Loader: Always load with LoraLoaderModelOnly (not standard LoraLoader), as this adapter targets the MOSS-Image transformer backbone only.

• Exact Tag Spelling: Keep underscores and parentheses identical to the table in Section 1 (e.g., <|artist|>ask_(askzy)<|/artist|>, <|artist|>hiten_(hitenkei)<|/artist|>).

• License & Hosting: Free for personal local generation and Civitai on-site generation. Hosting on third-party commercial generation services or reselling this model / its merges is strictly prohibited.