AuroraIMG2-50M-Preview

AuroraIMG2-50M-Preview

Tiny text-to-image model that makes fixed 256 by 256 pictures from English prompts, built on a 50.4M flow-matching transformer trained from scratch. A preview release for researchers and hobbyists.

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: AuroraIMG2-50M-Preview makes 256 by 256 pictures from a text prompt. Its denoiser has 50.4M parameters and its text encoder 12.6M, and both were trained from scratch. A frozen FLUX.2 small decoder, shipped unchanged in the vae folder, turns the model's latents into the final pixels. The authors call it a Preview release.

Who it suits: It suits researchers, students and hobbyists who want to try a tiny text-to-image model that runs from a short Python script.

Industry | Who uses it | How they use it | Example
Research | Diffusion researchers | Study a small flow-matching model trained on a single GPU | Testing how few steps still give a usable image
Education | Students | Learn how a text encoder and denoiser work together | Reading the pipeline code beside the outputs
Hobby AI | Hobbyists | Generate small pictures from English prompts on their own machine | Making a storybook-style picture
Prototyping | Developers | Try a light text-to-image step in a test app | Batching several prompts in one call

If you need larger pictures or prompts in other languages, you can skip it, since size is fixed at 256 by 256 and the language listed is English.

Where to use it: You run it from a Python script that uses the diffusers library.

• Windows, macOS or Linux with Python
• A CUDA GPU if you have one, otherwise the CPU, as the card's code picks
• A script or notebook using diffusers with remote code allowed

Use case | Who | What they get back
One picture from a prompt | Hobbyists | A 256 by 256 image saved as a file
Several pictures at once | Developers | A batch from a list of prompts
Fast drafts | Researchers | Quicker results with 20 steps

How to set it up: The repository uses a custom pipeline, so loading it needs trust_remote_code turned on. The model files are about 202 MB in full precision.

• Python with torch and diffusers
• Permission to run the pipeline code that ships with the model
• Disk space for about 202 MB of weights

1. Download the model files, or let the code fetch them.
2. Open a Python script and import torch and DiffusionPipeline.
3. Load the pipeline with its custom pipeline name and trust_remote_code set to True.

pip install -U huggingface_hub
hf download AuroraAI-Research/AuroraIMG2-50M-Preview

How to use it: Load the pipeline, pick a seed, then call it with a prompt, a step count and a guidance scale. The card recommends 50 steps and a guidance scale of 3.0, with 20 steps for fast runs.

Tool | What it does
DiffusionPipeline.from_pretrained | Loads the model with its custom pipeline
pipe | Turns a prompt into one or more images
num_images_per_prompt | Makes several images for each prompt

import torch
from diffusers import DiffusionPipeline

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

pipe = DiffusionPipeline.from_pretrained(
    "AuroraAI-Research/AuroraIMG2-50M-Preview",
    custom_pipeline="AuroraAI-Research/AuroraIMG2-50M-Preview",
    trust_remote_code=True,
).to(device)

generator = torch.Generator(device=device).manual_seed(0)
image = pipe(
    prompt="A tiny wooden house on top of a giant red mushroom in an enchanted forest, warm glowing windows, soft morning sunlight, little flowers and moss, magical storybook illustration, vibrant colors, charming and detailed.",
    num_inference_steps=50,
    guidance_scale=3.0,
    generator=generator,
).images[0]
image.save("auroraimg2.png")

Pass a list of prompts for a batch. There is no negative prompt, because the model was trained only with its own learned empty prompt.

Example: You ask for a lighthouse on a cliff at dusk, painted in oils, with seed 0 and 50 steps. The script saves a 256 by 256 picture as a png file, and the same seed gives the same picture again.

For newcomers and experts: A text-to-image model draws a picture from words. Start small. Run the script once as written, open auroraimg2.png, then change the prompt and the seed.

For experts, it is latent flow matching with classifier-free guidance, working in the FLUX.2 VAE latent space of 128 by 16 by 16. The text encoder is a 6-layer transformer with an 8192-token byte-level vocabulary, trained jointly with the denoiser. Prompts run to 62 tokens plus start and end. Training lasted about 10 hours on one RTX PRO 6000, for 247,805 steps at batch size 256.

A prompt is split into byte-level tokens
The text encoder turns tokens into features
The denoiser refines a latent over 50 Euler steps
The frozen FLUX.2 decoder turns latents into pixels
A 256 by 256 image comes out

Good to know: On GenEval 2, the card lists a Soft-TIFA GM of 6.00 and an AM of 36.03. Its eight sample pictures were the first ones generated, not picked from many tries. Benchmarks used fp32 weights with bfloat16 autocast, and loading in bfloat16 saves memory but was not tested on its own.

Title: Parameters by part, in millions
Denoiser, trained from scratch: 50.4 M
Text encoder, trained from scratch: 12.6 M
Frozen FLUX.2 decoder: 27.9 M

• License: Apache 2.0, and the bundled decoder is also Apache 2.0
• Cost to run: none named
• Data it reaches: your prompts and the model files you download, plus the pipeline code that loads with it
• Not for: pictures at any size other than 256 by 256, or prompts that need a negative prompt

On Hugging Face, 13 November 2026

• Task: text to image
• Library: diffusers
• Parameters: 50.4M
• Weights: about 202 MB in F32; about 25 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 159
• Likes: 19
• License: apache-2.0
• Languages: en
• Tags: diffusers, safetensors, text-to-image, flow-matching, rectified-flow, latent-diffusion, diffusion-transformer, tiny-model, from-scratch, muon, auroraimg

Links

• Model: https://huggingface.co/AuroraAI-Research/AuroraIMG2-50M-Preview
• Files: https://huggingface.co/AuroraAI-Research/AuroraIMG2-50M-Preview/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download AuroraAI-Research/AuroraIMG2-50M-Preview

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