Xiangjun Tang

1/5 Dropping a 3D Gaussian Splatting avatar into a captured 3DGS scene is easy. Making it cast a believable shadow is not. Happy to share RAGA, accepted at #ECCV2026.

5/5 Project Page at miraymen.github.io/raga. Great collaboration with @peter_wonka @_JianWang_ @GerardPonsMoll1

@Jaeyeon_Kim_0 @kiwhansong0 Sorry for misunderstanding. It seems that t equals to 0 for clean image. But there comes another question. For an almost clean image, the predicted score should be obvious so the conditional and unconditional prediction should be the same, but the inference diff is huge.

@Jaeyeon_Kim_0 @kiwhansong0 thank you for the inspiring work! I am stuck reading fig 2. When t equals 0, the input to the network should be a pure noise. At this setting, I think the behavior of inference and training should be the same. Why score diff in training is zero but in inference is huge?

We introduce a new ''rule'' for understanding diffusion models: Selective Underfitting. It explains: 🚨 How diffusion models generalize beyond training data 🚨 Why popular training recipes (e.g., DiT, REPA) are effective and scale well Co-led with @kiwhansong0! (1/n)