How to Run AuroraIMG-6M Locally: AuroraAI-Research's 6.82M-Parameter Text-to-Image Model

AuroraIMG-6M is a text-to-image model from AuroraAI-Research with just 6.82 million parameters, including its text encoder, denoiser and VAE. Released October 11, 2026 under Apache 2.0, it generates 64 by 64 pixel images from short English captions using latent flow matching, and runs comfortably on CPU with the diffusers library.

Written by Priya Raghunathan, local-AI hardware reviewer. I check parameter counts and hardware requirements against the primary source before recommending an install path.

A new top-20 entrant on Hugging Face's trending list

AuroraIMG-6M entered the top 20 of Hugging Face's text-to-image trending list. At time of writing it has 34 likes and 104 downloads on the AuroraAI-Research/AuroraIMG-6M model page, released October 11, 2026, under the Apache 2.0 license. The model card lists no GitHub repository; the top linked community Space is hugging-apps/auroraimg-6m (2 likes).

Specs at a glance

How it was built

AuroraIMG-6M is made of three small pieces, all trained together from scratch, no pretrained text encoder or VAE borrowed from elsewhere. A convolutional VAE with residual blocks squeezes a 64 by 64 image down into an 8-channel, 16 by 16 latent, reaching 31.9 dB PSNR on holdout images. A 4-layer transformer with rotary position embeddings and its own byte-level BPE tokenizer reads the caption, trained jointly with the image side so it learns whatever text representation the rest of the model actually needs. The biggest of the three pieces, a 4.42 million parameter U-Net denoiser, removes noise from the latent step by step, using AdaLN-Zero conditioning so each residual block starts out as an identity function before training nudges it.

Results

All numbers below come from the model card's own evaluation, measured at the model's native 64 by 64 resolution rather than upsampled for comparison against larger models.

GenEval checks whether a generated image actually matches what a prompt asked for, counts, colors, positions and the like, rather than just looking photorealistic. A GenEval overall score of 0.1387 is modest, which tracks with a 6.82 million parameter model: it is built to prove a tiny, from-scratch pipeline can generate recognizable images at all, not to compete with billion-parameter models on prompt accuracy.

Run it

AuroraIMG-6M loads through Hugging Face's diffusers library with a custom pipeline hosted on its own model repo.

from diffusers import DiffusionPipeline
import torch

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.bfloat16 if device == "cuda" else torch.float32

pipe = DiffusionPipeline.from_pretrained(
    "AuroraAI-Research/AuroraIMG-6M",
    custom_pipeline="AuroraAI-Research/AuroraIMG-6M",
    trust_remote_code=True,
    torch_dtype=dtype
).to(device)

prompt = "a sunset over the ocean"
print(f"Generating: '{prompt}'...")

generator = torch.Generator(device=device).manual_seed(42)
image = pipe(
    prompt=prompt,
    num_inference_steps=50,
    guidance_scale=6.0,
    generator=generator
).images[0]

filename = "Aurora-Image.png"
image.save(filename)
print(f"Saved to {filename}!")

The model card calls it small enough to run on CPU, so the cuda check above falls back cleanly on a laptop with no GPU. If you just want to see sample output before installing anything, the hugging-apps/auroraimg-6m Space runs the same model in the browser.

Known limitations

The model card is specific about where this falls short. Output is capped at 64 by 64 pixels, so fine detail simply is not there to render. It handles a single clear subject reasonably well but struggles with counting, relative position and prompts that stack several attributes at once. Captions are English only and lowercased; the text encoder supports up to 64 tokens, though training captions themselves were capped at 62. CC12M carries the usual stock-photo and web-scraped biases of any dataset pulled from the open internet, and the model card is explicit that this is a research artifact, not a tool for realistic portraiture or any use where fairness and accuracy matter.

Licensing

AuroraIMG-6M's weights are Apache 2.0. The training data is a different story: CC12M images carry whatever license their original source applied, so the model card asks that you verify licensing case by case rather than treat the Apache 2.0 grant on the weights as covering generated output outright.

FAQ

What is AuroraIMG-6M?

AuroraIMG-6M is a 6.82 million parameter text-to-image model from AuroraAI-Research, trained entirely from scratch on CC12M. It generates 64 by 64 pixel images from short English captions using latent flow matching.

Does it need a GPU to run?

No. The model card describes it as small enough to run on CPU, and at 6.82 million parameters total it is far smaller than mainstream text-to-image models.

How good are the images it generates?

On a CC12M holdout set it scores an FID of 33.01 and a CLIP score of 24.8, about 88 percent of what real photos score on the same CLIP metric. It handles single, clear subjects reasonably well but struggles with counting and multi-attribute prompts.

What resolution does it output?

64 by 64 pixels, well below most current text-to-image models. That is a direct trade-off for a model this small trained entirely from scratch.

Can I use it commercially?

The weights are Apache 2.0, but CC12M's training images carry their own original licenses, so check those case by case before relying on generated output commercially.

Where to go from here

• Need an image model that handles complex prompts and renders real text? Qwen-Image-2512 covers a 20-billion-parameter alternative built for exactly that.
• Want another tiny from-scratch text-to-image model to compare against? Run Supra2-IMG locally covers a 104-million-parameter model trained on synthetic images at 256 by 256.
• Browse other image and diffusion models: the local AI model directory lets you compare specs side by side.