| Summary: | Currently, Latent Diffusion Models (LDMs) are very adept at generating completely novel images. However, they tend to be lacking in generating images following specific conditions. This final year research project addresses the challenge of enhancing image generation using LDMs by incorporating conditional control.
The purpose of the project is to explore the potential of conditional LDMs in facilitating light editing on images. This allows users to create realistic modifications without deep technical knowledge of the underlying processes. There is an increasingly large artistic community growing around generative AI, primarily developed with text prompts. However, there is a gap in light editing capabilities, hence expanding in this area can provide additional creative options.
The project schedule breaks down the project into planning, researching, development, testing and evaluation stages. The method involved preparation of indoor dataset and light mask dataset, research of LDM functionality, and exploration of techniques for integrating conditioning mechanisms into LDMs. We developed a conditional LDM utilizing concatenation mechanism and binary light mask dataset which is able to produce high fidelity panoramic image outpainting. Thus, the results of the project demonstrate the feasibility and effectiveness of utilizing conditional LDMs for light editing tasks. Recommendations include the exploration of dynamic light masks dataset and the development of an intuitive user interface.
|
|---|