Stable Labs unveils CanvasGen: a finetuning toolkit for creative models focused on layout fidelity

AI · 5 min read

CanvasGen provides data pipelines, loss functions, and evaluation metrics specifically tuned for layout fidelity—things like grid adherence, component overlap penalties, and tokenized style loss. The toolkit simplifies the process of taking a base generative model and aligning it to a particular design language or component library.

Included are tools for synthetic data augmentation so teams can expand limited proprietary datasets with realistic variants. CanvasGen also comes with a visual diff utility to compare generated layouts against target specifications and quantify drift over time.

Agencies and studios that need to protect brand assets while leveraging generative models may find CanvasGen particularly useful. Stable Labs recommends a phased approach: start with conservative finetuning for token alignment, then expand to more creative generation once fidelity improves.