Nanosaur2-670M

Nanosaur2-670M

Makes illustrations from text prompts with a small 670M-parameter diffusion model, run through ComfyUI. Built as a low-budget research model for people who want to try it or train their own LoRA.

Made by well9472 on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Nanosaur2-670M is a text-to-image model for illustrations, with 670M parameters and a diffusion transformer design. The package adds ComfyUI support, so you load it as custom nodes and a ready workflow. Its author trained it for a total cost of $600 to see what is possible on a small compute budget.

Who it suits: It suits ComfyUI users and researchers who want a small illustration model to try, study or adapt with their own images.

Industry | Who uses it | How they use it | Example
Research | Machine learning researchers | Study a small image model trained on a low budget | Read its training recipe and compare it with larger models
Illustration | Digital artists | Make illustrations in ComfyUI from tags or plain language | Describe a scene and add the quality tags the card suggests
Open models | LoRA trainers | Train a LoRA on a folder of captioned images | Run the included training script on a folder of pictures

People who need fine detail or deep character knowledge can skip it, because the author says not to expect that from this model.

Where to use it: The model runs inside ComfyUI, and the download steps work wherever Python does.

• ComfyUI, through the nanosaur2_support custom node folder
• Windows, macOS and Linux, with Python for the download

Use case | Who | What they get back
Text-to-image illustration | Artists and hobbyists | A picture made from a tag list or a natural language prompt
LoRA training | Model tinkerers | A safetensors LoRA file you load back into ComfyUI

How to set it up: You need ComfyUI installed, plus Python with pip to fetch the files.

1. Download the repository's files with the hf command.
2. Copy the nanosaur2_support folder into the custom_nodes folder of ComfyUI.
3. Put the three model files into ComfyUI's model folders, as the table below shows.
4. Start ComfyUI and drag the nanosaur2_workflow.json file onto it.

pip install -U huggingface_hub
hf download well9472/Nanosaur2-670M

Title: Where each model file goes
File | Folder
nanosaur2_diffusion_model.safetensors | models/diffusion_models
nanosaur2_text_encoder.safetensors | models/text_encoders
nanosaur2_vae.safetensors | models/vae

How to use it: Open the workflow in ComfyUI and write your prompt as tags or as plain sentences. The card's sampler setting is Euler simple, with 50 steps and CFG 4. Shift 3 is already built in.

Add the tags newest and masterpiece to the positive prompt, and oldest and low quality to the negative one. Some artist tags work too, and you may need to weight them up. To train a LoRA, run the script below on a folder of images that each have a text file with the same name.

uv run python custom_nodes/nanosaur2_support/train_lora.py /path/to/images

Tool | What it does
nanosaur2_support | Adds the model's custom nodes to ComfyUI
nanosaur2_workflow.json | Loads the ready-made workflow when dragged onto ComfyUI
train_lora.py | Trains a LoRA and saves it into the models/loras folder
LoraLoaderModelOnly | Loads the trained LoRA back into a workflow

Example: Describe a scene such as a fennec girl in a sunlit forest, the subject of the card's own sample picture, and run the workflow. You get a finished illustration. The sample PNG also carries its ComfyUI workflow, so dragging it onto ComfyUI loads those settings.

For newcomers and experts: If you are new to this, install ComfyUI first, then follow the four setup steps and try the workflow with a short prompt. Expect a research model, not a polished one.

If you know the field, the model is a 670M diffusion transformer with adaLN-single conditioning, 2D RoPE, SwiGLU, QK-norm and sparse middle blocks. It pairs a frozen gemma-3-270m text encoder with a DINOv2-based VAE of 129M parameters. Sampling alternates CFG with path-drop guidance.

You write tags or plain sentences
The frozen gemma-3-270m encoder reads the prompt
The diffusion transformer denoises the picture
The DINOv2-based VAE decodes the final image

Good to know: The author is looking for compute funding to train it further and to make a 2B version.

• License: MIT
• Cost to run: none named
• Data it reaches: model files in ComfyUI's folders, and for LoRA training the folder of images and captions you point it to
• Not for: matching fully trained models, since the card calls this a research model
• Other versions: the record counts 0 quantized and 5 fine-tuned copies made by other people

Title: Base training time by picture size, in days
256x256: 3 days
512x512: 3 days
1024x1024: 5 days

On Hugging Face, 13 November 2026

• Formats: Safetensors
• Downloads in the last 30 days: 0
• Likes: 112
• License: mit
• Fine-tunes: 5
• Tags: safetensors

Links

• Model: https://huggingface.co/well9472/Nanosaur2-670M
• Files: https://huggingface.co/well9472/Nanosaur2-670M/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download well9472/Nanosaur2-670M

Then follow the usage in its model card. Nothing in this kit runs by itself.