Inconsistency between timestep and noise level in Null-Text Inversion? about prompt-to-prompt

[KiwiXR]

To support my thoughts, I further modify the code in ddim_loop into:

class NullInversion:
    ...
    @torch.no_grad()
    def ddim_loop(self, latent):
        uncond_embeddings, cond_embeddings = self.context.chunk(2)
        all_latent = [latent]
        latent = latent.clone().detach()
        for i in range(NUM_DDIM_STEPS):
            t = self.model.scheduler.timesteps[len(self.model.scheduler.timesteps) - i - 1]
            next_t = min(t - self.scheduler.config.num_train_timesteps // self.scheduler.num_inference_steps, 999)  # copied from `self.next_step`
            noise_pred = self.get_noise_pred_single(latent, next_t, cond_embeddings)  # modified
            latent = self.next_step(noise_pred, t, latent)
            all_latent.append(latent)
        return all_latent

Then I calculate the null-text inverted image's PSNR with the cat example image:

from skimage.metrics import peak_signal_noise_ratio
psnr = peak_signal_noise_ratio(image_gt, image_inv[0])
print(psnr)

The original version gives 29.56082923568291, while the modified version gives 29.605481827030523 (greater is better).

Hope this can demonstrate my ideas.