A consistent simulation model for environmental art design generation driven by multimodal transformer

Aiming at semantic disconnection and visual distortion between generated results and actual scenes in environmental art design, this paper proposes a consistent generation and simulation model based on multimodal transformer.Traditional methods have limitations in coordinating complex elements and ensuring spatial logic, hindering design implementation.By integrating multi-source information including text, sketches, and scene images, an end-to-end generation-simulation framework achieves consistent mapping from concept to high-fidelity visual output.Using the public dataset MIT ade20k, results show the model achieves significant improvements in visual fidelity (area under the curve 0.92, an increase of 8.2%) and user preference (normalised discounted cumulative gain @10 an increase of 15.7%), with all key indicators being statistically significant (p < 0.01).This confirms the model's effectiveness in enhancing automation and usability of environmental art design.