Atelier: A New Framework for Artist-Grounded Image Generation
A new framework named Atelier has been developed by researchers to enhance artist-based image generation through shortcut-aware control-state planning. This framework tackles the issue where image models depend on conventional shortcuts—like common motifs, typical color schemes, or frequently seen period styles—when interpreting artist names, instead of accurately reflecting the user's desired scene. Atelier converts vague artistic intentions into a clear control state that differentiates scene anchors, decisions on preservation or transformation, style hypotheses, artist-specific evidence, and constraints to avoid shortcuts. Utilizing artist-level insights and local patch references, it formulates backend-aware generation strategies, refining options through authenticity feedback. Additionally, the researchers present ArtIntentBench, a benchmark featuring artists Van Gogh and Qi Baishi for various tasks, including artwork re-rendering. The paper can be found on arXiv with the identifier 2608.06751.
Key facts
- Atelier is a shortcut-aware control-state planning framework for artist-grounded image generation.
- It addresses the issue of image models using canonical shortcuts like recurring motifs, generic palettes, or period signatures.
- The framework separates scene anchors, preserve/transform decisions, style-regime hypotheses, role-bound artist evidence, and shortcut-avoidance constraints.
- It uses artist-level knowledge and local patch references to ground the control state.
- Atelier compiles backend-aware generation plans and iteratively refines candidates with global and local authenticity feedback.
- ArtIntentBench is a new benchmark introduced, covering Van Gogh and Qi Baishi.
- The benchmark includes artwork re-rendering and other tasks.
- The paper is available on arXiv with identifier 2608.06751.
Entities
Artists
- Vincent van Gogh
- Qi Baishi
Institutions
- arXiv