Question about model architecture of HiDream-E1.

[CyberPegasus]

Congratulations on your great work !
I have some questions about the architecture of HiDream-E1, I would be grateful if you could provide some insignts.

1. Why concat latent image and condition image at width dimension before patchify, instead of token dimension (height dimension) after patchify ?
2. The position embedding seems to regard latents and condition image as an Union Image. Why not using separating position embedding ( e.g., OminiControl ) for latents and condition image ?

[bghira]

or keep the flux ids in the prev impl a lot of the ideas were taken from. the 0 ids for the "denoising" image tokens allows simple differentiation to "reference" tokens, where the img_ids are 1.

[liuliu]

Probably just the same training code with updated dataset? I am happy to see the model is adopting well with this kind of RoPE encoding (even you can consider it sub-optimal to FLUX Kontext's). It suggests dataset is ever more important than subtle architecture differences.