Open seaman1900 opened 5 months ago
https://github.com/gnobitab/RectifiedFlow/blob/9df93aa874f7d2c5097bf7c089698c071c71b5c8/ImageGeneration/losses.py#L101 You mentioned that xt = (1-t)x + ty, and in the code, t starts with eps (1e-3) in training, however, you used t=eps in inference and the input of model should be x_0 (1-eps)+ yeps accordingly. I feel quiet strange about this.
在采样的时候用的是xt = (1-t)x + ty, t的范围是[1e-3,1]，训练的时候t=1e-3对应的是x_0 (1-eps)+ yeps，但是推理的时候t的起点也是1e-3，这时候是没有y用于计算x_t的，是出于什么原因这样设计的呢？
https://github.com/gnobitab/RectifiedFlow/blob/9df93aa874f7d2c5097bf7c089698c071c71b5c8/ImageGeneration/losses.py#L101 You mentioned that xt = (1-t)x + ty, and in the code, t starts with eps (1e-3) in training, however, you used t=eps in inference and the input of model should be x_0 (1-eps)+ yeps accordingly. I feel quiet strange about this.
在采样的时候用的是xt = (1-t)x + ty, t的范围是[1e-3,1]，训练的时候t=1e-3对应的是x_0 (1-eps)+ yeps，但是推理的时候t的起点也是1e-3，这时候是没有y用于计算x_t的，是出于什么原因这样设计的呢？