Aurora

Aurora - developed by Malik Ayaan Ahmed.

Aurora is a text-to-image system built for photorealistic images, accurate on-image text and calligraphy in many typefaces, logos, and profile pictures. It pairs a fine-tuned image generator with a "Thinker" stage that writes a director-style prompt, renders several candidates, and keeps the one that best matches your idea.

What Aurora does

• Photorealistic scenes with cinematic lighting and camera-style detail
• Legible text and calligraphy in varied styles (script, serif, condensed sans, display)
• Logos, stickers, posters, signs, and circular profile-picture avatars
• Prompt-guided image editing (aurora.edit(image, "make it sunset")), best for light-to-moderate changes
• Knows its own name: ask for "an image of you" and it depicts Aurora

Development (developed by Malik Ayaan Ahmed)

• Two-stage QLoRA fine-tuning on high-resolution photographs plus synthetic typography, logo and avatar data rendered with open-licensed fonts, merged into standalone weights
• Thinker pipeline, typography safeguards, identity handling, and training tooling written by Malik Ayaan Ahmed
• Intermediate v1 weights are kept in transformer_v1/

Credits and license

Aurora's image backbone is a fine-tuned derivative of Freepik/flux.1-lite-8B by its original authors; its license terms continue to apply to this repository. The Thinker uses Qwen/Qwen2.5-3B-Instruct.

Status: v2 trained and merged

Use

!pip install -q -U diffusers transformers accelerate bitsandbytes sentencepiece peft
from huggingface_hub import hf_hub_download
exec(open(hf_hub_download("AyaanAhmed123/Aurora", "aurora_generate.py")).read())
aurora = Aurora(repo="AyaanAhmed123/Aurora")
image, info = aurora('a neon sign that reads "OPEN LATE" in flowing script, rainy street, 35mm photo')
image.save("out.png")