Hi, thanks for sharing this impressive work! The idea of controllable bokeh shapes is fascinating.
I have a question regarding the generalization capability of the shape-conditioned model (described in Section 3.3 and Appendix C).
I noticed that the PointLight-1K dataset is specifically constructed to "reveal aperture shape" by mining keywords like "night" and "bokeh" (Appendix C). This makes me wonder:
1. Training Data Bias: Since the shape-conditioned LoRA is fine-tuned primarily on night-time or strong point-light scenarios, how does the model perform on daylight scenes where there are no obvious point light sources?
2. Generalization Mechanism: In a typical daytime photo, the bokeh is formed by the integration of many small textures (not single points). Does the model learn the "integration kernel" well enough to apply the shape to these complex textures, or does it tend to fail/ignore areas without strong luminance?
I understand that the paper's figures mainly showcase night scenes (e.g., Fig. 7). Have you observed any specific failure cases or limitations when applying custom aperture shapes to bright, texture-rich daylight environments?
Hi, thanks for sharing this impressive work! The idea of controllable bokeh shapes is fascinating.
I have a question regarding the generalization capability of the shape-conditioned model (described in Section 3.3 and Appendix C).
I noticed that the PointLight-1K dataset is specifically constructed to "reveal aperture shape" by mining keywords like "night" and "bokeh" (Appendix C). This makes me wonder:
1. Training Data Bias: Since the shape-conditioned LoRA is fine-tuned primarily on night-time or strong point-light scenarios, how does the model perform on daylight scenes where there are no obvious point light sources?
2. Generalization Mechanism: In a typical daytime photo, the bokeh is formed by the integration of many small textures (not single points). Does the model learn the "integration kernel" well enough to apply the shape to these complex textures, or does it tend to fail/ignore areas without strong luminance?
I understand that the paper's figures mainly showcase night scenes (e.g., Fig. 7). Have you observed any specific failure cases or limitations when applying custom aperture shapes to bright, texture-rich daylight environments?