About "reversed_latents_w"

[Qilong Zhang]

Hi Yuxin,

Sorry for bothering your again. In run_tree_ring_watermark.py, the reversed_latents_w is obtained by

  reversed_latents_w = pipe.forward_diffusion(
      latents=image_latents_w,
      text_embeddings=text_embeddings,    # why provides the information about prompt? 
      guidance_scale=1,
      num_inference_steps=args.test_num_inference_steps,
  )

However, I notice you write "While it may not be surprising that inversion is accurate for unconditional diffusion models, inversion also succeeds well-enough for conditional diffusion models, even when the conditioning c is not provided. This property of inversion will be exploited heavily by our watermark below."

Does the conditioning c mean text_embeddings?

[Yuxin Wen]

Hi Qilong, thanks for asking. text_embeddings is actually a dummy embedding of an empty string as we defined here. Therefore, we don't provide any prompt information during the inversion.

[Qilong Zhang]

Thanks for your reply! It's right