Q1: 为什么你们的diffusion反向推理的时候的 ddim_timestep_seq 是[991, 981, 971, ... , 11, 1]，不应该是[1000, 990, 980, ... , 20, 10]吗？也就是说反向推理时高斯噪声对应的是 x_991 而不是 x_1000
Q2: 而且将时间作为求alpha的索引的时候不应该减 1 吗？
Q1: Why the ddim_timestep_seq in your diffusion reverse inference is [991, 981, 971, ... , 11, 1]? Shouldn't it be [1000, 990, 980, ... , 20, 10]? That is to say, when inferring backward, the Gaussian noise corresponds to x_991 instead of x_1000
Q2: And shouldn't 1 be subtracted when using time as the index for alpha?
你们的代码 / your code:
self.ddim_timestep_seq = np.asarray(list(range(0, self.noise_steps, self.noise_steps // self.ddim_timesteps))) + 1
Q1: 为什么你们的diffusion反向推理的时候的
ddim_timestep_seq是[991, 981, 971, ... , 11, 1]，不应该是[1000, 990, 980, ... , 20, 10]吗？也就是说反向推理时高斯噪声对应的是x_991而不是x_1000Q2: 而且将时间作为求alpha的索引的时候不应该减 1 吗？
Q1: Why the
ddim_timestep_seqin your diffusion reverse inference is [991, 981, 971, ... , 11, 1]? Shouldn't it be [1000, 990, 980, ... , 20, 10]? That is to say, when inferring backward, the Gaussian noise corresponds tox_991instead ofx_1000Q2: And shouldn't 1 be subtracted when using time as the index for alpha?
你们的代码 / your code:
self.ddim_timestep_seq = np.asarray(list(range(0, self.noise_steps, self.noise_steps // self.ddim_timesteps))) + 1