CivArchive
    mbl9 - a synthetic character for MOSS-Image (LoRA + full training data) - v1 (SFT)
    NSFW
    Preview 149673757
    Preview 149673745
    Preview 149673760
    Preview 149673762

    Meet mbl9 ("Mei Bunny Lin"): a fictional, quirky streetwear girl with heavy winged eyeliner, bangs, freckles, purple hair streaks and a Jean Paul Gaultier-style wardrobe of spiral-striped corsets, patched denim and chains. She doesn't exist. I made her so my LoRA guides can show every image, every sheet and the whole dataset, which I can't do with real people.

    She's the worked example of my two articles:

    How I made a person who doesn't exist

    1. Design in ChatGPT Images. My starting prompt, with ChatGPT free to expand it: create a fictional character, a woman, Asian, eclectic, kind of quirky and weird but sexy at the same time, who loves funky streetwear like Jean Paul Gaultier; give me 4 examples to choose the base from.

    2. Pick a base. I had it make 44 different versions of the one I liked and chose one.

    3. Dataset. ChatGPT generated the training images of that one character: mostly studio headshots with different tops, expressions and angles (images 1–40), and full-body shots in her signature punk outfit (41–79), 60 images in all.

    4. Captions with my dataset tagger workflow (Qwen-VL): one natural-language sentence per image that starts with the trigger mbl9 and describes what changes from image to image (clothes, place, light, expression, framing), so the trigger word carries the person.

    Lesson learned: rename only the images, never the caption files. A bulk rename turned 20 of my captions into .png files.

    Training in ai-toolkit

    • Model: OpenMOSS-Team/MOSS-Image-SFT, transformer and text encoder quantized to convrot8, low VRAM on.

    • LoRA: rank 32, alpha 32. Layer 22 is left out of training (ignore_if_contains: layers.22.) because Skip Layer Guidance uses it at inference.

    • Training: AdamW 8-bit, LR 1e-4, batch 1, flowmatch with shift timesteps, 512/768/1024 buckets, caption dropout 0.05, cached text embeddings, gradient checkpointing.

    • 3000 steps, a checkpoint every 250.

    • Hardware: RTX 4090 capped at 320 W, about 9.7 GB VRAM. 41 minutes of training plus about an hour of sample images.

    The exact job is in the training data zip (mbl9_moss_sft_v1.yaml, paths made relative).

    Which checkpoint ships

    This file isn't a single checkpoint. Part 1 of the guide compared all 12 checkpoints against her training images, then ran a 3-seed head-to-head. The winner was an exact 50/50 merge of steps 1500 and 2250 (rank-stacked, so rank 64, 290 MB). It won 8.8 of 15 rows. The one-seed favourite, step 2500, won none.

    Using it

    • Trigger: mbl9, at strength 1.0. If you use LoRA Manager's trigger words, it puts mbl9 in front of your prompt for you.

    • Best base model I found: Jib Mix Moss Realistic by J1B. Part 2 of the guide swept her across 8 MOSS models, each on its own recommended settings, and it's my pick. My Jib Mix Moss workflow is set up for her.

    • Her look lives in the styling. Her identity is strongest in studio portraits and in her own outfit. In everyday scenes (cooking, cycling) the eyeliner and freckles fade, so name them in the prompt: heavy winged eyeliner, freckles, blush, bangs, purple streaks in her black hair.

    • Trained on MOSS SFT. It also loads on MOSS Base and the SFT-based finetunes.

    Training data

    mbl9-training-data.zip has all 60 images with their captions and the ai-toolkit config. Use it to retrain her, to try other settings or models, or to follow the guide step by step. She's synthetic, so the images are free to use for that.

    Happy training! Kiko

    Description

    Trained on MOSS-Image SFT in ai-toolkit (rank 32, 3000 steps). This file is an exact 50/50 merge of steps 1500 and 2250 (rank 64), picked against the training images. Use mbl9 at strength 1.0. Training data (60 images, captions, ai-toolkit config) is attached.

    LORA
    MOSS Image

    Details

    Downloads
    23
    Platform
    CivitAI
    Platform Status
    Available
    Created
    11/20/2026
    Updated
    11/20/2026
    Deleted
    -
    Trigger Words:
    mbl9

    Files

    mbl9-training-data.zip

    Size:
    129.64 MB
    SHA256:

    Mirrors

    mbl9-moss-sft-v1.safetensors

    Mirrors

    CivitAI (1 mirrors)