Question about reference image timestep conditioning in image editing inference

[Liu-Yangshuai]

Hi, thanks for releasing JoyAI-Image-Edit.

I have a question about the timestep conditioning for the reference image during image editing inference.

From reading the current inference code, my understanding is that the input image is first VAE-encoded into a clean latent, and during the denoising loop this reference latent is copied back before every transformer call, so its latent value stays clean.

However, the transformer seems to receive only one shared global timestep:

t_expand = t.repeat(latent_model_input.shape[0])

So the reference image latent is clean and fixed, but it appears to share the same current denoising timestep embedding/modulation as the noisy target latent. I do not see a per-item timestep mechanism where the reference image uses a fixed timestep=0 while the target image uses the current timestep t.

Could you clarify whether this is an intentional design in JoyAI-Image-Edit? Was training also done with clean reference latents and noisy target latents sharing the same current timestep?

I am asking because semantically the reference image seems closer to a clean x_0 condition, so one might expect it to use fixed timestep 0. I would appreciate any clarification on the design choice.

Thanks!