MOSS-Image 5B finetune hits ~90% artist-style accuracy with just 3,685 images on one MI300X

Turbulent-Bass-649 · reddit · 2026-10-23

A Reddit user released MossAD-Demo, a proof-of-concept full finetune of OpenMOSS's MOSS-Image Base 5B T2I model on a curated 3,685-image Danbooru/E621 dataset covering 57 artists, claiming MOSS-Image beats earlier mid-size architectures like MageFlow 4B.

Why it's hyped

Pretrained on 3.2B images: SDXL-like knowledge in a 5B model, paired with Flux2VAE and a 9B text encoder
Native bounding-box layout prompting reportedly better than Ideogram 4; the author says most users are missing 50% of its potential by ignoring the official userguide
Runs semi-comfortably on 16GB VRAM; fits 12GB with int8 convrot

Training results (39 hours on one AMD MI300X)

90% style accuracy, transferring to subjects the artists never drew (furry styles onto anime girls and vice versa)
4.5k views/artist gives 70–75% style accuracy; 7k views gives 85–90%; 4k views only works in-domain
Most training stays at cheap 512px with short 768/1024px stages for hands and detail; LR must be halved per resolution stage (4e-5→2e-5→1e-5) or training destabilizes within 5 epochs
No major architectural flaws found yet; occasional instability and tricky character LoRA training noted

Original post →

MOSS-Image 5B finetune hits ~90% artist-style accuracy with just 3,685 images on one MI300X

More from Multimodal