Classifier free guidance(CFG) on different prediction types and karras style schedulers

There is banch of repos [1](https://github.com/duxiaodan/intrinsic-lora/blob/b69770c53035b936eae3132354afb1e14d2043ea/rescale_cfg_pipeline_forward.py#L259), [2](https://github.com/Amblyopius/Stable-Diffusion-ONNX-FP16/blob/55d050957f70c90f14aed13d9854f7aa59ff1f70/pipeline_onnx_stable_diffusion_instruct_pix2pix.py#L361), [3](https://github.com/scnuhealthy/video_try_on/blob/4b72bb32c59d37ed6b16a61f6906d014bc511e69/video_models/video_pipeline_mae_guided.py#L308) there autors take into account type of noise scheduler and for karras like change model output before cfg calculation:
```
# Hack:
# For karras style schedulers the model does classifer free guidance using the
# predicted_original_sample instead of the noise_pred. So we need to compute the
# predicted_original_sample here if we are using a karras style scheduler.
if scheduler_is_in_sigma_space:
step_index = (self.scheduler.timesteps == t).nonzero()[0].item()
sigma = self.scheduler.sigmas[step_index]
noise_pred = latent_model_input - sigma * noise_pred

# perform guidance
if self.do_classifier_free_guidance:
noise_pred_text_image, noise_pred_text = noise_pred.chunk(2)
noise_pred = (
noise_pred_text
+ self.image_guidance_scale * (noise_pred_text_image - noise_pred_text)
)

# Hack:
# For karras style schedulers the model does classifer free guidance using the
# predicted_original_sample instead of the noise_pred. But the scheduler.step function
# expects the noise_pred and computes the predicted_original_sample internally. So we
# need to overwrite the noise_pred here such that the value of the computed
# predicted_original_sample is correct.
if scheduler_is_in_sigma_space:
noise_pred = (noise_pred - latents) / (-sigma)
```

Firstly, I don’t know is this step so important and where find the original idea for cfg in karras like schedulers. Secondly, I wondered whether a similar effect should be considered for different types of network output (x0, eps, v). I will be very grateful for any investigaion on this topic.

Commencez par comparer les trois implémentations de pipeline liées, en vous concentrant sur les branches scheduler_is_in_sigma_space et sur le calcul de classifier-free guidance. Suivez la manière dont scheduler.step consomme noise_pred et dont les exemples gèrent les valeurs predicted-original-sample, puis examinez si la même question s’applique aux types de prédiction x0, eps et v. La tâche sera considérée comme terminée lorsque le comportement correct sera documenté et qu’un périmètre d’implémentation ou de test convenu sera identifié.

Rédigé par le modèle d'indexation à partir du texte de l'issue.