Here are some possible evaluation metrics to evaluate the quality of composite images from different aspects.

• Evaluate whether the foreground is harmonious with background.

  • Harmony score: use illumination encoder to extract the illumination codes from foreground and background, and measure their similarity.

  • Inharmony hit: use inharmonious region localization model to detect the inharmonious region, and calculate the overlap (e.g., IoU) between detected region and foreground region.

• Evaluate whether the foreground object placement is reasonable.

  • OPA score: use object placement assessment model to predict the accuracy of object placement.

• Evaluate whether the foreground is compatible with background in terms of geometry and semantics.

  • FOS score: use foreground object search model to calculate the compatibility score between foreground and background in terms of geometry and semantics.

• Evaluate the fidelity of foreground, i.e., whether the synthesized foreground is similar to the input foreground.

  • Clip score: use CLIP to extract the embeddings from the input foreground image and the generated foreground patch, and measure their similarity.

  • Dino score: use DINO to measure the average cosine similarity between the input and generated foreground.

• Evaluate the over quality of foreground or the whole composite image.

  • FID: use pretrained image encoder (e.g., InceptionNet, CLIP) to extract the embeddings from real images and generated images, and measure their Fréchet Inception Distance.
  • QS: use quality score to measure the quality of each single generated image, and compute average score.