Question about the method 'ddim_step' in DDIMSolver

[luochunfen]

Hi ! I have some problems about the method --ddim_step in DDIMSolver. When caculating the previous latent state, the input pred_noise is only treated as noise without considering other prediction types. Is this forgetting to consider other prediction types?

[huskydoge]

Seems like already been fixed in huggingface diffuser:

• https://github.com/huggingface/diffusers/blob/298ab6eb01f3ef475c15218ea87de1494e1250aa/examples/consistency_distillation/train_lcm_distill_lora_sd_wds.py#L1351

• latent-consistency-model/LCM_Training_Script/consistency_distillation/train_lcm_distill_lora_sd_wds.py

  Lines 1218 to 1231 in a9ad795

  	encoder_hidden_states=uncond_prompt_embeds.to(weight_dtype),
  	).sample
  	uncond_pred_x0 = predicted_origin(
  	uncond_teacher_output,
  	start_timesteps,
  	noisy_model_input,
  	noise_scheduler.config.prediction_type,
  	alpha_schedule,
  	sigma_schedule,
  	)
  	# 20.4.11. Perform "CFG" to get x_prev estimate (using the LCM paper's CFG formulation)
  	pred_x0 = cond_pred_x0 + w * (cond_pred_x0 - uncond_pred_x0)
  	pred_noise = cond_teacher_output + w * (cond_teacher_output - uncond_teacher_output)