Question about RoPE

[yinian-lw]

https://github.com/Saquib764/omini-kontext/blob/main/src/train/model.py#L123

x_0, x_ids = encode_images(self.flux_pipe, target_images)
          
x_init, init_img_ids = encode_images(self.flux_pipe, input_images)
# Prepare reference image with delta
x_ref, ref_img_ids = encode_images(self.flux_pipe, reference_images)

# Apply position delta to reference image IDs
delta = reference_deltas[0]
ref_img_ids[:, 0] += delta[0]
ref_img_ids[:, 1] += delta[1]
ref_img_ids[:, 2] += delta[2]

# Combine input and reference images
condition = torch.cat([x_init, x_ref], dim=1)
condition_ids = torch.cat([init_img_ids, ref_img_ids], dim=0)

https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/train_dreambooth_lora_flux_kontext.py#L1940

cond_latents_ids = FluxKontextPipeline._prepare_latent_image_ids(
          cond_model_input.shape[0],
          cond_model_input.shape[2] // 2,
          cond_model_input.shape[3] // 2,
          accelerator.device,
          weight_dtype,
      )
cond_latents_ids[..., 0] = 1
latent_image_ids = torch.cat([latent_image_ids, cond_latents_ids], dim=0)

Why do not add 1 on init_img_ids[:, 0] like flux kontext?

[Saquib764]

Please go through this document - #12