AuroraIMG2-100M

AuroraIMG2-100M

Generates 256x256 images from text with a tiny 100M-parameter model trained from scratch on one GPU. For people who want a small text-to-image model to run or study.

Made by AuroraAI-Research 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: AuroraIMG2-100M is a small text-to-image model that makes 256x256 pictures. Inside is a diffusion transformer of 100.6M parameters that uses flow matching, plus a text encoder of 44.9M parameters. Both were trained from scratch with one NVIDIA RTX PRO 6000. It reuses the FLUX.2 VAE latent space and a frozen FLUX.2 small decoder.

Who it suits: It suits developers and researchers who want a tiny image model they can run and study, and people who want to see what a model trained on one GPU can do.

Industry | Who uses it | How they use it | Example
Research | Image model researchers | Study a small flow-matching model | Compare it with the 50M Preview
Software | Developers testing pipelines | Load the custom pipeline through diffusers | Generate small test images in a CI job
Education | Students of diffusion models | See each part of a compact model | Read the settings table and run the quick start

People who need larger pictures can skip it, since the resolution is fixed at 256x256.

Where to use it:

• Python with the diffusers library
• An NVIDIA GPU, or a CPU, since the quick start picks whichever is available
• Windows, macOS or Linux, where you can download the files with Python

Use case | Who | What they get back
Small image generation | Developers | 256x256 pictures from a text prompt
Batch generation | Experimenters | Several images per prompt or a list of prompts at once
Model study | Researchers | A compact model with published settings and scores

How to set it up: You need these first:

• Python and the huggingface_hub package
• The diffusers library and PyTorch
• Permission to run the repository's custom pipeline, since the card says it needs trust_remote_code set to True

1. Install the Hugging Face hub package.
2. Download the model files to your computer. They take about 402 MB in F32.
3. Run the quick start script below.

pip install -U huggingface_hub
hf download AuroraAI-Research/AuroraIMG2-100M

How to use it: Run the quick start script with your own prompt. For batches, give it a list of prompts. To get several images from one prompt, set num_images_per_prompt.

import torch
from diffusers import DiffusionPipeline

device = "cuda" if torch.cuda.is_available() else "cpu"

pipe = DiffusionPipeline.from_pretrained(
    "AuroraAI-Research/AuroraIMG2-100M",
    custom_pipeline="AuroraAI-Research/AuroraIMG2-100M",
    trust_remote_code=True,
).to(device)

generator = torch.Generator(device=device).manual_seed(0)
image = pipe(
    prompt="Vintage Yellow Enamel Tea Kettle with Black Handle and Spout",
    num_inference_steps=50,
    guidance_scale=4.0,
    generator=generator,
).images[0]
image.save("auroraimg2.png")

Tool | What it does
num_inference_steps | Sets the sampler steps, 50 by default and 20 for fast generation
guidance_scale | Sets the CFG value, 4.0 in the card's tests
shift | Sets the time-grid shift, 1.5 in the card's tests
guidance_off_below | Switches guidance off for the last low-noise steps

Example: You pass a prompt such as a vintage yellow enamel tea kettle with a black handle and spout. The script saves one 256x256 image as a PNG file named auroraimg2.png.

For newcomers and experts: A text-to-image model draws a picture from a sentence. A first small thing to try is the quick start script with its sample prompt, then your own prompt.

The model uses latent flow matching with classifier-free guidance and the Euler sampler. Its latent is 128 by 16 by 16 in the FLUX.2 VAE space, and the text encoder is a 12-layer bidirectional transformer with a byte-level BPE vocabulary. It takes no negative prompt. Its unconditional branch is a learned null prompt, and arbitrary negative prompts never appeared in training.

Write a text prompt
The text encoder reads up to 190 tokens
The transformer denoises a latent in 50 steps
The FLUX.2 decoder turns it into pixels

Good to know:

• Size: 101M parameters per Hugging Face, about 402 MB in F32 and about 50 MB at 4-bit, as an estimate
• Training: 235,000 steps at batch size 512 on one RTX PRO 6000, around 80 GPU-hours
• Scores: the card reports Soft-TIFA GM of 10.80 and AM of 48.62 on GenEval 2, both higher than the Preview

Title: GenEval 2 Soft-TIFA GM
AuroraIMG2-100M: 10.80
AuroraIMG2-50M-Preview: 6.00

Title: GenEval 2 Soft-TIFA AM
AuroraIMG2-100M: 48.62
AuroraIMG2-50M-Preview: 36.03

• License: apache-2.0, and the bundled FLUX.2 small decoder is Apache-2.0 too
• Cost to run: none named, no API key or account
• Data it reaches: it downloads model files from Hugging Face, and the quick start runs the custom pipeline code that ships with the repository
• Not for: pictures larger than 256x256, or arbitrary negative prompts, since the card says neither was part of the design

On Hugging Face, 13 November 2026

• Task: text to image
• Library: diffusers
• Parameters: 101M
• Weights: about 402 MB in F32; about 50 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 179
• Likes: 24
• License: apache-2.0
• Languages: en
• Tags: diffusers, safetensors, text-to-image, rectified-flow, flow-matching, latent-diffusion, diffusion-transformer, tiny-model, from-scratch, muon, custom-pipeline, aurora

Links

• Model: https://huggingface.co/AuroraAI-Research/AuroraIMG2-100M
• Files: https://huggingface.co/AuroraAI-Research/AuroraIMG2-100M/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download AuroraAI-Research/AuroraIMG2-100M

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