Not really tested WAN Phantom Support.

[comfyanonymous]

No description provided.

[George0726]

Hello author, I am currently using Phantom models. I found that this kind of model should use two negative conditions for the CFG guidance.
In your node WanPhantomSubjectToVideo, the two negative nodes refer to negative_text and negative_img_text. I am wondering if there are nodes that can be used for multiple CFG calculations, just like the functioning in the Phantom source code below?

self.model.to(self.device)
pos_it = self.model(
                    [torch.cat([latent[:,:-ref_latent.shape[1]], ref_latent], dim=1) for latent, ref_latent in zip(latents, ref_latents)], t=timestep, **arg_c
                    )[0]
pos_i = self.model(
                    [torch.cat([latent[:,:-ref_latent.shape[1]], ref_latent], dim=1) for latent, ref_latent in zip(latents, ref_latents)], t=timestep, **arg_null
                    )[0]
neg = self.model(
                    [torch.cat([latent[:,:-ref_latent_neg.shape[1]], ref_latent_neg], dim=1) for latent, ref_latent_neg in zip(latents, ref_latents_neg)], t=timestep, **arg_null
                    )[0]
noise_pred = neg + guide_scale_img * (pos_i - neg) + guide_scale_text * (pos_it - pos_i)