Image blurring in restoring predicted image

[Holmes2002]

Hi, I reimplemented this code to train. However, Fitdit lora just worked with Noise training when I add Frequency loss restoration in predicted image was a problem

noise_pred = net(
                    hidden_states=torch.cat([noisy_model_input, vton_model_input, batch["mask_input"].to(dtype=weight_dtype)], dim=1),
                    timesteps=timesteps,
                    pooled_projections=cloth_image_embeds,
                    pose_image=batch["pose_image"].to(dtype=weight_dtype),
                    ref_key=ref_key,
                    ref_value=ref_value,
                    encoder_hidden_states=None,
                    return_dict=False,
                )
                # mask torch.Size([1, 1, 128, 96])
                x0_pred = (noisy_model_input - sigmas * noise_pred) # / (1.0 - sigmas)
                x0_pred = (x0_pred / vae.config.scaling_factor) + vae.config.shift_factor

                pixel_x0_pred = vae.decode(x0_pred, return_dict=False)[0]

I believed SD3 was trained in flow matching so I modified a little. However, predicted images was blurred so they affected the whole training process. Any advice or help for me