Reve holds its own against the giants with ten times fewer GPUs.

Reve 2.1, a frugal image model by independent lab Reve, delivers native 4K visuals with ten times fewer GPUs than rivals on the Design Arena.

Just one month after Reve 2.0, independent lab Reve is following up with Reve 2.1, a version of its image model focused on better understanding intent. The announced improvements are most noticeable, according to the publisher, in marketing visuals, abstract patterns, and people, with enhanced rendering of text in foreign languages.

Reve's uniqueness stems from a gamble made two years ago: representing an image not as the simple product of a prompt, but as a set of structured and hierarchical regions, addressable one by one, like code. The model therefore plans the layout and reasons about the structure and spatial relationships between elements before rendering, which targets dense and complex scenes. With each area remaining independently editable, a portion can be modified and then regenerated without redoing the rest. The output is native 4K, or sixteen megapixels.

The other argument is frugality. Reve presents itself as an independent lab training its models on significantly fewer resources: it claims the spot of best 4K model and leading independent image lab on the Arena and Design Arena rankings, and says it takes second place in the overall ranking while training on less than one-tenth of the computing power of the labs ranked around it.

The model is accessible via Reve's application and API.