repair SDXL

        z = self.encode_with_transformers(tokens)

        pooled = getattr(z, 'pooled', None)

        # restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise

        batch_multipliers = torch.asarray(batch_multipliers).to(devices.device)

        original_mean = z.mean()

        new_mean = z.mean()

        z = z * (original_mean / new_mean)

        if pooled is not None:

            z.pooled = pooled

        return z