RicemanT/Nanosaur-1.2B-Preview-Latest

The original trainer of the model, well9472/Metal63 throughout the project's lifetime would keep deleting the model repo due to an unknown reason. As a silent observer of the project progress, I'm uploading my latest saved checkpoint of the project here to perserve a promising open source image model even if it's original creator have abandoned it. It's also still potentially useful as a toy model for training practice.

Nanosaur 1.2B is a custom pretrained from scratch arch with (will be updated once I scour back arch info- basically a custom arch, Gemma3 270m and a custom VAE called PS-VAE from what I remember), using a mixed 621/Danbooru dataset. So this is actually the first model of 1-3B param range that was trained on e621 since NoobAI XL. Money roughly spent before it was discontinued was somewhere around 1k-3k dollars, with H100 support from Lodestone Rock.

Original README.md:

WIP model for research purposes. Still in progress. 18 days of 1xH100 from scratch.

Install instructions

Copy nanosaursupport folder to customnodes in ComfyUI Copy diffusion model, text encoder and VAE to model folders in ComfyUI Start ComfyUI Drag nanosaur_workflow.json onto ComfyUI

Tags or natural language. Model responds well to prompt emphasis like (character:2) or (artist:2)

Lora training:

uv run cachelora.py --dataset-dir /path/to/imagesandtxtcaptionfiles --batch-size 4 uv run trainlora.py --cache-path cache/lora_latents.pt --batch-size 1 --rank 16 --alpha 16