Boogu-Image-0.1 tops HuggingFace Daily on a $400K training bill

Boogu-Image-0.1 topped HuggingFace Daily on July 16 with 727 upvotes, per the arXiv listing from its 33 authors. The next paper on the list got 154, a 573-vote gap.

The number that matters is the training cost: around $400,000, theoretical, for a unified model that both understands and generates images. That's roughly what two senior engineers cost a year, a budget within reach for independent labs and small teams chasing frontier-adjacent results.

For that cost, the authors release weights, code, and training recipes under Apache 2.0, no usage restrictions, no gated access request. The model trained on 208.62 million unique images. Four variants ship: Base, Turbo, Edit, and Edit-Turbo. Together they cover text-to-image generation, fast inference, instruction-based editing, and bilingual Chinese-English text rendering, one family instead of four separate specialist models.

Why the margin matters

A 727-to-154 gap is a rout. HuggingFace Daily surfaces dozens of papers a day, and the community picked this one by nearly 5-to-1 over whatever ranked second. Open licensing paired with a documented low training cost is what pulled votes here.

If the $400K figure holds up under scrutiny, it becomes the reference point every future efficient-multimodal-model paper gets measured against. The arXiv paper doesn't detail what compute assumptions produced that estimate, so it's the authors' own number, not an independently audited one. Watch for follow-up work that confirms or challenges it.