Upload 6 files (#1)

- Upload 6 files (a2c2e9a29781f1deb4c9f895dc89c3aa948fbbb3)


Co-authored-by: Shaswat Garg <ShaswatRobotics@users.noreply.huggingface.co>

ctrl_world/droid/checkpoint-10000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed17de48180d4e6f89fd33c53e9fb7a0196189c1a67d44c2c486a279a80ea8a8
+size 9281040326

ctrl_world/droid/config.json

@@ -0,0 +1,52 @@
+{
+  "name": "droid_ctrl_world",
+  "env": "DROID",
+  "model_type": "ctrl_world",
+  "metadata": {
+    "num_history": 6,
+    "num_frames": 5,
+    "action_dim": 7
+  },
+  "util_folders":{
+    "models": "../src/models"
+  },
+  "models": [
+    {
+      "name": "world_model",
+      "framework": null,
+      "format": "state_dict",
+      "source": {
+        "weights_path": "checkpoint-10000.pt",
+        "class_path": "../src/world_model.py",
+        "class_name": "CrtlWorld",
+        "class_args": [
+        {
+          "svd_model_path": "stabilityai/stable-video-diffusion-img2vid",
+          "clip_model_path": "openai/clip-vit-base-patch32",
+          "num_history": 6,
+          "num_frames": 5,
+          "action_dim": 7,
+          "text_cond": true,
+          "motion_bucket_id": 127,
+          "fps": 7,
+          "guidance_scale": 1.0,
+          "num_inference_steps": 50,
+          "decode_chunk_size": 7,
+          "width": 320,
+          "height": 192
+        }]
+      },
+      "methods":
+      [
+        {
+          "name": "blocks_left_in_kv_cache",
+          "method_name": "blocks_left_in_kv_cache"
+        },
+        {
+          "name": "reset_kv_cache",
+          "method_name": "reset_kv_cache"
+        }
+      ]
+    }
+  ]
+}

ctrl_world/src/models/pipeline_ctrl_world.py

@@ -0,0 +1,823 @@
+from typing import Callable, Dict, List, Optional, Union
+import torch
+from einops import rearrange, repeat
+import PIL
+import einops
+
+# from diffusers import TextToVideoSDPipeline, StableVideoDiffusionPipeline
+from diffusers import TextToVideoSDPipeline
+from models.pipeline_stable_video_diffusion import StableVideoDiffusionPipeline
+
+
+from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_synth import TextToVideoSDPipelineOutput
+from diffusers.pipelines.stable_video_diffusion.pipeline_stable_video_diffusion import StableVideoDiffusionPipelineOutput
+from diffusers.utils.torch_utils import randn_tensor
+
+def svd_tensor2vid(video: torch.Tensor, processor, output_type="np"):
+    # Based on:
+    # https://github.com/modelscope/modelscope/blob/1509fdb973e5871f37148a4b5e5964cafd43e64d/modelscope/pipelines/multi_modal/text_to_video_synthesis_pipeline.py#L78
+
+    batch_size, channels, num_frames, height, width = video.shape
+    outputs = []
+    for batch_idx in range(batch_size):
+        batch_vid = video[batch_idx].permute(1, 0, 2, 3)
+        batch_output = processor.postprocess(batch_vid, output_type)
+
+        outputs.append(batch_output)
+
+    return outputs
+
+class LatentToVideoPipeline(TextToVideoSDPipeline):
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt = None,
+        height= None,
+        width= None,
+        num_frames: int = 16,
+        num_inference_steps: int = 50,
+        guidance_scale= 9.0,
+        negative_prompt= None,
+        eta: float = 0.0,
+        generator= None,
+        latents= None,
+        prompt_embeds= None,
+        negative_prompt_embeds= None,
+        output_type= "np",
+        return_dict: bool = True,
+        callback= None,
+        callback_steps: int = 1,
+        cross_attention_kwargs= None,
+        condition_latent=None,
+        mask=None,
+        timesteps=None,
+        motion=None,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the video generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated video.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated video.
+            num_frames (`int`, *optional*, defaults to 16):
+                The number of video frames that are generated. Defaults to 16 frames which at 8 frames per seconds
+                amounts to 2 seconds of video.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality videos at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate videos that are closely linked to the text `prompt`,
+                usually at the expense of lower video quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the video generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for video
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`. Latents should be of shape
+                `(batch_size, num_channel, num_frames, height, width)`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            output_type (`str`, *optional*, defaults to `"np"`):
+                The output format of the generate video. Choose between `torch.FloatTensor` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.TextToVideoSDPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion.TextToVideoSDPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.TextToVideoSDPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated frames.
+        """
+        # 0. Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        num_images_per_prompt = 1
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
+        )
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        #device = self._execution_device
+        device = latents.device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        text_encoder_lora_scale = (
+            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+        )
+        prompt_embeds = self._encode_prompt(
+            prompt,
+            device,
+            num_images_per_prompt,
+            do_classifier_free_guidance,
+            negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            lora_scale=text_encoder_lora_scale,
+        )
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        if timesteps is None:
+            timesteps = self.scheduler.timesteps
+        else:
+            num_inference_steps = len(timesteps)
+        # 5. Prepare latent variables. do nothing
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        uncondition_latent = condition_latent
+        condition_latent = torch.cat([uncondition_latent, condition_latent]) if do_classifier_free_guidance else condition_latent 
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+                if motion is not None:
+                    motion = torch.tensor(motion, device=device)
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    condition_latent=condition_latent,
+                    mask=mask,
+                    motion=motion
+                ).sample
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # reshape latents
+                bsz, channel, frames, width, height = latents.shape
+                latents = latents.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height)
+                noise_pred = noise_pred.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+                # reshape latents back
+                latents = latents[None, :].reshape(bsz, frames, channel, width, height).permute(0, 2, 1, 3, 4)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        callback(i, t, latents)
+
+        video_tensor = self.decode_latents(latents)
+
+        if output_type == "pt":
+            video = video_tensor
+        else:
+            video = tensor2vid(video_tensor)
+
+        # Offload last model to CPU
+        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+            self.final_offload_hook.offload()
+
+        if not return_dict:
+            return (video, latents)
+
+        return TextToVideoSDPipelineOutput(frames=video)
+
+def _append_dims(x, target_dims):
+    """Appends dimensions to the end of a tensor until it has target_dims dimensions."""
+    dims_to_append = target_dims - x.ndim
+    if dims_to_append < 0:
+        raise ValueError(f"input has {x.ndim} dims but target_dims is {target_dims}, which is less")
+    return x[(...,) + (None,) * dims_to_append]
+
+class CtrlWorldDiffusionPipeline(StableVideoDiffusionPipeline):
+    @torch.no_grad()
+    def __call__(
+        self,
+        image,
+        text,
+        height: int = 576,
+        width: int = 1024,
+        num_frames: Optional[int] = None,
+        num_inference_steps: int = 25,
+        min_guidance_scale: float = 1.0,
+        max_guidance_scale: float = 3.0,
+        fps: int = 7,
+        motion_bucket_id: int = 127,
+        noise_aug_strength: int = 0.02,
+        decode_chunk_size: Optional[int] = None,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        return_dict: bool = True,
+        mask = None,
+        cond_wrist=None,
+        history=None,
+        frame_level_cond=False,
+        his_cond_zero=False,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.FloatTensor`):
+                Image or images to guide image generation. If you provide a tensor, it needs to be compatible with
+                [`CLIPImageProcessor`](https://huggingface.co/lambdalabs/sd-image-variations-diffusers/blob/main/feature_extractor/preprocessor_config.json).
+            height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+                The width in pixels of the generated image.
+            num_frames (`int`, *optional*):
+                The number of video frames to generate. Defaults to 14 for `stable-video-diffusion-img2vid` and to 25 for `stable-video-diffusion-img2vid-xt`
+            num_inference_steps (`int`, *optional*, defaults to 25):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference. This parameter is modulated by `strength`.
+            min_guidance_scale (`float`, *optional*, defaults to 1.0):
+                The minimum guidance scale. Used for the classifier free guidance with first frame.
+            max_guidance_scale (`float`, *optional*, defaults to 3.0):
+                The maximum guidance scale. Used for the classifier free guidance with last frame.
+            fps (`int`, *optional*, defaults to 7):
+                Frames per second. The rate at which the generated images shall be exported to a video after generation.
+                Note that Stable Diffusion Video's UNet was micro-conditioned on fps-1 during training.
+            motion_bucket_id (`int`, *optional*, defaults to 127):
+                The motion bucket ID. Used as conditioning for the generation. The higher the number the more motion will be in the video.
+            noise_aug_strength (`int`, *optional*, defaults to 0.02):
+                The amount of noise added to the init image, the higher it is the less the video will look like the init image. Increase it for more motion.
+            decode_chunk_size (`int`, *optional*):
+                The number of frames to decode at a time. The higher the chunk size, the higher the temporal consistency
+                between frames, but also the higher the memory consumption. By default, the decoder will decode all frames at once
+                for maximal quality. Reduce `decode_chunk_size` to reduce memory usage.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableVideoDiffusionPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableVideoDiffusionPipelineOutput`] is returned,
+                otherwise a `tuple` is returned where the first element is a list of list with the generated frames.
+
+        Examples:
+
+        ```py
+        from diffusers import StableVideoDiffusionPipeline
+        from diffusers.utils import load_image, export_to_video
+
+        pipe = StableVideoDiffusionPipeline.from_pretrained("stabilityai/stable-video-diffusion-img2vid-xt", torch_dtype=torch.float16, variant="fp16")
+        pipe.to("cuda")
+
+        image = load_image("https://lh3.googleusercontent.com/y-iFOHfLTwkuQSUegpwDdgKmOjRSTvPxat63dQLB25xkTs4lhIbRUFeNBWZzYf370g=s1200")
+        image = image.resize((1024, 576))
+
+        frames = pipe(image, num_frames=25, decode_chunk_size=8).frames[0]
+        export_to_video(frames, "generated.mp4", fps=7)
+        ```
+        """
+        # 0. Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        num_frames = num_frames if num_frames is not None else self.unet.config.num_frames
+        decode_chunk_size = decode_chunk_size if decode_chunk_size is not None else num_frames
+        # device = self._execution_device
+        device = self.unet.device   
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = max_guidance_scale > 1.0
+
+        # # 1. Check inputs. Raise error if not correct
+        # self.check_inputs(image, height, width)
+
+        # # 2. Define call parameters
+        # if isinstance(image, PIL.Image.Image):
+        #     batch_size = 1
+        # elif isinstance(image, list):
+        #     batch_size = len(image)
+        # else:
+        #     batch_size = image.shape[0]
+        # # 3. Encode input image
+        # # clip_imgae = self.video_processor.preprocess(image, height=224, width=224)
+        # clip_image = _resize_with_antialiasing(image, (224, 224))
+        # image_embeddings = self._encode_image(clip_image, device, num_videos_per_prompt, do_classifier_free_guidance)
+        image_embeddings = text
+        batch_size = image_embeddings.shape[0]
+        if do_classifier_free_guidance:
+            negative_image_embeddings = torch.zeros_like(image_embeddings)
+            image_embeddings = torch.cat([negative_image_embeddings, image_embeddings])
+        
+
+        # NOTE: Stable Diffusion Video was conditioned on fps - 1, which
+        # is why it is reduced here.
+        # See: https://github.com/Stability-AI/generative-models/blob/ed0997173f98eaf8f4edf7ba5fe8f15c6b877fd3/scripts/sampling/simple_video_sample.py#L188
+        # fps = fps - 1 # we only use fps = 7 in train, so just set to 7
+
+        # 4. Encode input image using VAE
+        needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+        if image.shape[-3] == 3: # (batch, 3, 256, 256)
+            image = self.video_processor.preprocess(image, height=height, width=width)
+            noise = randn_tensor(image.shape, generator=generator, device=image.device, dtype=image.dtype)
+            # image = image + noise_aug_strength * noise
+
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float32)
+
+            image_latents = self._encode_vae_image(image, device, num_videos_per_prompt, do_classifier_free_guidance)
+            image_latents = image_latents.to(image_embeddings.dtype)
+
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
+        else: # (batch, 4, 32, 32)
+            image_latents = image/self.vae.config.scaling_factor
+            if do_classifier_free_guidance:
+                # negative_image_latent = torch.zeros_like(image_latents)
+                # image_latents = torch.cat([negative_image_latent, image_latents])
+                image_latents = torch.cat([image_latents]*2)
+            image_latents = image_latents.to(image_embeddings.dtype)
+
+        # Repeat the image latents for each frame so we can concatenate them with the noise
+        # image_latents [batch, channels, height, width] ->[batch, num_frames, channels, height, width]
+        if history is not None:
+            B, num_his, C, H, W = history.shape
+            num_frames_all = num_frames + num_his
+            image_latents = image_latents.unsqueeze(1).repeat(1, num_frames_all, 1, 1, 1)
+            if his_cond_zero:
+                image_latents[:,:num_his] = 0.0 # set history to 0
+        else:
+            image_latents = image_latents.unsqueeze(1).repeat(1, num_frames, 1, 1, 1)
+        # mask = repeat(mask, '1 h w -> 2 f 1 h w', f=num_frames)
+        # 5. Get Added Time IDs
+        added_time_ids = self._get_add_time_ids(
+            fps,
+            motion_bucket_id,
+            noise_aug_strength,
+            image_embeddings.dtype,
+            batch_size,
+            num_videos_per_prompt,
+            do_classifier_free_guidance,
+        )
+        added_time_ids = added_time_ids.to(device)
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_frames,
+            num_channels_latents,
+            height,
+            width,
+            image_embeddings.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 7. Prepare guidance scale
+        guidance_scale = torch.linspace(min_guidance_scale, max_guidance_scale, num_frames).unsqueeze(0)
+        guidance_scale = guidance_scale.to(device, latents.dtype)
+        guidance_scale = guidance_scale.repeat(batch_size * num_videos_per_prompt, 1)
+        guidance_scale = _append_dims(guidance_scale, latents.ndim)
+
+        self._guidance_scale = guidance_scale
+
+        # 8. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+        # print("prediction type",self.scheduler.config.prediction_type)
+        if cond_wrist is not None:
+            B,F, C, H, W = latents.shape
+            cond_wrist = einops.repeat(cond_wrist, 'b l c h w -> b (f l) (n c) h w', n=3,f=num_frames) # (B, 8, 12 , 24, 40)
+            cond_wrist = torch.cat([cond_wrist]*2) if do_classifier_free_guidance else cond_wrist
+        
+        if history is not None:
+            history = torch.cat([history] * 2) if do_classifier_free_guidance else history
+        
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                if history is not None:
+                    latent_model_input = torch.cat([history, latent_model_input], dim=1) # (bsz*2,frame+F,4,32,32)
+
+                # Concatenate image_latents over channels dimention
+                latent_model_input = torch.cat([latent_model_input, image_latents], dim=2)
+
+                if cond_wrist is not None and i==0:
+                    # print('cond_wrist_shape:',cond_wrist.shape, 'latent_model_input_shape:',latent_model_input.shape)
+                    latent_model_input = torch.cat([latent_model_input, cond_wrist], dim=3) # (B, 8, 12, 96, 40)
+                
+
+                # predict the noise residual
+                latent_model_input = latent_model_input.to(self.unet.dtype)
+                image_embeddings = image_embeddings.to(self.unet.dtype)
+                # print('extract_layer_idx:',extract_layer_idx)
+                # print('latent_model_input_shape:',latent_model_input.shape)
+                # print('encoder_hidden_states:',image_embeddings.shape)
+                # print('added_time_ids:',added_time_ids.shape)
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=image_embeddings,
+                    added_time_ids=added_time_ids,
+                    return_dict=False,
+                    frame_level_cond=frame_level_cond,
+                )[0]
+
+                if cond_wrist is not None:
+                    noise_pred = noise_pred[:, :,:,:H, :W] # remove cond_wrist
+                if history is not None:
+                    # print('history_shape:',history.shape)
+                    # print('noise_pred_shape:',noise_pred.shape)
+                    noise_pred = noise_pred[:, num_his:, :, :, :] # remove history
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond)
+
+                # model_output = noise_pred
+                # # sigma = self.scheduler.get_sigma(t)
+                # # sigma = self.scheduler.sigmas[t]
+                # self.scheduler._init_step_index(t)
+                # sigma = self.scheduler.sigmas[self.scheduler.step_index]
+                # print("sigma", sigma)
+                # print(t)
+                # pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (latents / (sigma**2 + 1))
+                # print(pred_original_sample.shape)
+                # latents = pred_original_sample
+                # # return pred_original_sample
+                # break
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents).prev_sample
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+        if not output_type == "latent":
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
+            # latents = latents/self.vae.config.scaling_factor
+            latents = latents.to(self.vae.dtype)
+            frames = self.decode_latents(latents, num_frames, decode_chunk_size)
+            frames = svd_tensor2vid(frames, self.video_processor, output_type=output_type)
+        else:
+            frames = latents
+
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return frames,latents
+
+        return StableVideoDiffusionPipelineOutput(frames=frames)
+        
+class TextStableVideoDiffusionPipeline(StableVideoDiffusionPipeline):
+    @torch.no_grad()
+    def __call__(
+        self,
+        image,
+        prompt_embeds = None,
+        negative_prompt_embeds = None,
+        height: int = 576,
+        width: int = 1024,
+        num_frames: Optional[int] = None,
+        num_inference_steps: int = 25,
+        min_guidance_scale: float = 1.0,
+        max_guidance_scale: float = 3.0,
+        fps: int = 7,
+        motion_bucket_id: int = 127,
+        noise_aug_strength: int = 0.02,
+        decode_chunk_size: Optional[int] = None,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        return_dict: bool = True,
+        mask = None,
+        condition_type = "image",
+        condition_latent = None,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.FloatTensor`):
+                Image or images to guide image generation. If you provide a tensor, it needs to be compatible with
+                [`CLIPImageProcessor`](https://huggingface.co/lambdalabs/sd-image-variations-diffusers/blob/main/feature_extractor/preprocessor_config.json).
+            height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+                The width in pixels of the generated image.
+            num_frames (`int`, *optional*):
+                The number of video frames to generate. Defaults to 14 for `stable-video-diffusion-img2vid` and to 25 for `stable-video-diffusion-img2vid-xt`
+            num_inference_steps (`int`, *optional*, defaults to 25):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference. This parameter is modulated by `strength`.
+            min_guidance_scale (`float`, *optional*, defaults to 1.0):
+                The minimum guidance scale. Used for the classifier free guidance with first frame.
+            max_guidance_scale (`float`, *optional*, defaults to 3.0):
+                The maximum guidance scale. Used for the classifier free guidance with last frame.
+            fps (`int`, *optional*, defaults to 7):
+                Frames per second. The rate at which the generated images shall be exported to a video after generation.
+                Note that Stable Diffusion Video's UNet was micro-conditioned on fps-1 during training.
+            motion_bucket_id (`int`, *optional*, defaults to 127):
+                The motion bucket ID. Used as conditioning for the generation. The higher the number the more motion will be in the video.
+            noise_aug_strength (`int`, *optional*, defaults to 0.02):
+                The amount of noise added to the init image, the higher it is the less the video will look like the init image. Increase it for more motion.
+            decode_chunk_size (`int`, *optional*):
+                The number of frames to decode at a time. The higher the chunk size, the higher the temporal consistency
+                between frames, but also the higher the memory consumption. By default, the decoder will decode all frames at once
+                for maximal quality. Reduce `decode_chunk_size` to reduce memory usage.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableVideoDiffusionPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableVideoDiffusionPipelineOutput`] is returned,
+                otherwise a `tuple` is returned where the first element is a list of list with the generated frames.
+
+        Examples:
+
+        ```py
+        from diffusers import StableVideoDiffusionPipeline
+        from diffusers.utils import load_image, export_to_video
+
+        pipe = StableVideoDiffusionPipeline.from_pretrained("stabilityai/stable-video-diffusion-img2vid-xt", torch_dtype=torch.float16, variant="fp16")
+        pipe.to("cuda")
+
+        image = load_image("https://lh3.googleusercontent.com/y-iFOHfLTwkuQSUegpwDdgKmOjRSTvPxat63dQLB25xkTs4lhIbRUFeNBWZzYf370g=s1200")
+        image = image.resize((1024, 576))
+
+        frames = pipe(image, num_frames=25, decode_chunk_size=8).frames[0]
+        export_to_video(frames, "generated.mp4", fps=7)
+        ```
+        """
+        # 0. Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        num_frames = num_frames if num_frames is not None else self.unet.config.num_frames
+        decode_chunk_size = decode_chunk_size if decode_chunk_size is not None else num_frames
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(image, height, width)
+
+        # 2. Define call parameters
+        if isinstance(image, PIL.Image.Image):
+            batch_size = 1
+        elif isinstance(image, list):
+            batch_size = len(image)
+        else:
+            batch_size = image.shape[0]
+        device = self._execution_device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = max_guidance_scale > 1.0
+
+        # 3. Encode input image
+        if condition_type=="image":
+            image_embeddings = self._encode_image(image, device, num_videos_per_prompt, do_classifier_free_guidance)
+        elif condition_type=="text":
+            if do_classifier_free_guidance:
+                prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+            image_embeddings = prompt_embeds
+        else:
+            image_embeddings = self._encode_image(image, device, num_videos_per_prompt, do_classifier_free_guidance)
+            if do_classifier_free_guidance:
+                prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+            image_embeddings = torch.cat([image_embeddings, prompt_embeds], dim=1)
+        motion_mask = self.unet.config.in_channels == 9
+        if do_classifier_free_guidance:
+            mask = torch.cat([mask]*2) 
+        # NOTE: Stable Diffusion Video was conditioned on fps - 1, which
+        # is why it is reduced here.
+        # See: https://github.com/Stability-AI/generative-models/blob/ed0997173f98eaf8f4edf7ba5fe8f15c6b877fd3/scripts/sampling/simple_video_sample.py#L188
+        fps = fps - 1
+
+        # 4. Encode input image using VAE
+        image = self.video_processor.preprocess(image, height=height, width=width)
+        noise = randn_tensor(image.shape, generator=generator, device=image.device, dtype=image.dtype)
+        image = image + noise_aug_strength * noise
+
+        needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+        if needs_upcasting:
+            self.vae.to(dtype=torch.float32)
+
+        if condition_latent is None:
+            image_latents = self._encode_vae_image(image, device, num_videos_per_prompt, do_classifier_free_guidance)
+            image_latents = image_latents.to(image_embeddings.dtype)
+
+            # Repeat the image latents for each frame so we can concatenate them with the noise
+            # image_latents [batch, channels, height, width] ->[batch, num_frames, channels, height, width]
+            condition_latent = image_latents.unsqueeze(1).repeat(1, num_frames, 1, 1, 1)
+        else:
+            if do_classifier_free_guidance:
+                condition_latent = torch.cat([condition_latent] * 2) 
+        # 5. Get Added Time IDs
+
+        # cast back to fp16 if needed
+        if needs_upcasting:
+            self.vae.to(dtype=torch.float16)
+
+        added_time_ids = self._get_add_time_ids(
+            fps,
+            motion_bucket_id,
+            noise_aug_strength,
+            image_embeddings.dtype,
+            batch_size,
+            num_videos_per_prompt,
+            do_classifier_free_guidance,
+        )
+        added_time_ids = added_time_ids.to(device)
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_frames,
+            num_channels_latents,
+            height,
+            width,
+            image_embeddings.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 7. Prepare guidance scale
+        guidance_scale = torch.linspace(min_guidance_scale, max_guidance_scale, num_frames).unsqueeze(0)
+        guidance_scale = guidance_scale.to(device, latents.dtype)
+        guidance_scale = guidance_scale.repeat(batch_size * num_videos_per_prompt, 1)
+        guidance_scale = _append_dims(guidance_scale, latents.ndim)
+
+        self._guidance_scale = guidance_scale
+
+        # 8. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                if motion_mask:
+                    # Concatenate image_latents over channels dimention
+                    latent_model_input = torch.cat([mask, latent_model_input, condition_latent], dim=2)
+                else:
+                    latent_model_input = torch.cat([latent_model_input, condition_latent], dim=2)
+                # predict the noise residual
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=image_embeddings,
+                    added_time_ids=added_time_ids,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond)
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents).prev_sample
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+        if not output_type == "latent":
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
+            frames = self.decode_latents(latents, num_frames, decode_chunk_size)
+            frames = svd_tensor2vid(frames, self.video_processor, output_type=output_type)
+        else:
+            frames = latents
+
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return frames
+
+        return StableVideoDiffusionPipelineOutput(frames=frames)

ctrl_world/src/models/pipeline_stable_video_diffusion.py

@@ -0,0 +1,742 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from dataclasses import dataclass
+from typing import Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
+
+from diffusers.image_processor import PipelineImageInput
+
+# import from our own models instead  of diffusers
+# from diffusers.models import AutoencoderKLTemporalDecoder, UNetSpatioTemporalConditionModel
+from diffusers.models import AutoencoderKLTemporalDecoder
+from models.unet_spatio_temporal_condition import UNetSpatioTemporalConditionModel
+
+from diffusers.schedulers import EulerDiscreteScheduler
+from diffusers.utils import BaseOutput, is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils.torch_utils import is_compiled_module, randn_tensor
+from diffusers.video_processor import VideoProcessor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> from diffusers import StableVideoDiffusionPipeline
+        >>> from diffusers.utils import load_image, export_to_video
+
+        >>> pipe = StableVideoDiffusionPipeline.from_pretrained(
+        ...     "stabilityai/stable-video-diffusion-img2vid-xt", torch_dtype=torch.float16, variant="fp16"
+        ... )
+        >>> pipe.to("cuda")
+
+        >>> image = load_image(
+        ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd-docstring-example.jpeg"
+        ... )
+        >>> image = image.resize((1024, 576))
+
+        >>> frames = pipe(image, num_frames=25, decode_chunk_size=8).frames[0]
+        >>> export_to_video(frames, "generated.mp4", fps=7)
+        ```
+"""
+
+
+def _append_dims(x, target_dims):
+    """Appends dimensions to the end of a tensor until it has target_dims dimensions."""
+    dims_to_append = target_dims - x.ndim
+    if dims_to_append < 0:
+        raise ValueError(f"input has {x.ndim} dims but target_dims is {target_dims}, which is less")
+    return x[(...,) + (None,) * dims_to_append]
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+@dataclass
+class StableVideoDiffusionPipelineOutput(BaseOutput):
+    r"""
+    Output class for Stable Video Diffusion pipeline.
+
+    Args:
+        frames (`[List[List[PIL.Image.Image]]`, `np.ndarray`, `torch.Tensor`]):
+            List of denoised PIL images of length `batch_size` or numpy array or torch tensor of shape `(batch_size,
+            num_frames, height, width, num_channels)`.
+    """
+
+    frames: Union[List[List[PIL.Image.Image]], np.ndarray, torch.Tensor]
+
+
+class StableVideoDiffusionPipeline(DiffusionPipeline):
+    r"""
+    Pipeline to generate video from an input image using Stable Video Diffusion.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    Args:
+        vae ([`AutoencoderKLTemporalDecoder`]):
+            Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+        image_encoder ([`~transformers.CLIPVisionModelWithProjection`]):
+            Frozen CLIP image-encoder
+            ([laion/CLIP-ViT-H-14-laion2B-s32B-b79K](https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K)).
+        unet ([`UNetSpatioTemporalConditionModel`]):
+            A `UNetSpatioTemporalConditionModel` to denoise the encoded image latents.
+        scheduler ([`EulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents.
+        feature_extractor ([`~transformers.CLIPImageProcessor`]):
+            A `CLIPImageProcessor` to extract features from generated images.
+    """
+
+    model_cpu_offload_seq = "image_encoder->unet->vae"
+    _callback_tensor_inputs = ["latents"]
+
+    def __init__(
+        self,
+        vae: AutoencoderKLTemporalDecoder,
+        image_encoder: CLIPVisionModelWithProjection,
+        unet: UNetSpatioTemporalConditionModel,
+        scheduler: EulerDiscreteScheduler,
+        feature_extractor: CLIPImageProcessor,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            image_encoder=image_encoder,
+            unet=unet,
+            scheduler=scheduler,
+            feature_extractor=feature_extractor,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
+        self.video_processor = VideoProcessor(do_resize=True, vae_scale_factor=self.vae_scale_factor)
+
+    def _encode_image(
+        self,
+        image: PipelineImageInput,
+        device: Union[str, torch.device],
+        num_videos_per_prompt: int,
+        do_classifier_free_guidance: bool,
+    ) -> torch.Tensor:
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.video_processor.pil_to_numpy(image)
+            image = self.video_processor.numpy_to_pt(image)
+
+            # We normalize the image before resizing to match with the original implementation.
+            # Then we unnormalize it after resizing.
+            image = image * 2.0 - 1.0
+            image = _resize_with_antialiasing(image, (224, 224))
+            image = (image + 1.0) / 2.0
+
+        # Normalize the image with for CLIP input
+        image = self.feature_extractor(
+            images=image,
+            do_normalize=True,
+            do_center_crop=False,
+            do_resize=False,
+            do_rescale=False,
+            return_tensors="pt",
+        ).pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        image_embeddings = self.image_encoder(image).image_embeds
+        image_embeddings = image_embeddings.unsqueeze(1)
+
+        # duplicate image embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = image_embeddings.shape
+        image_embeddings = image_embeddings.repeat(1, num_videos_per_prompt, 1)
+        image_embeddings = image_embeddings.view(bs_embed * num_videos_per_prompt, seq_len, -1)
+
+        if do_classifier_free_guidance:
+            negative_image_embeddings = torch.zeros_like(image_embeddings)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            image_embeddings = torch.cat([negative_image_embeddings, image_embeddings])
+
+        return image_embeddings
+
+    def _encode_vae_image(
+        self,
+        image: torch.Tensor,
+        device: Union[str, torch.device],
+        num_videos_per_prompt: int,
+        do_classifier_free_guidance: bool,
+    ):
+        image = image.to(device=device)
+        image_latents = self.vae.encode(image).latent_dist.mode()
+
+        # duplicate image_latents for each generation per prompt, using mps friendly method
+        image_latents = image_latents.repeat(num_videos_per_prompt, 1, 1, 1)
+
+        if do_classifier_free_guidance:
+            negative_image_latents = torch.zeros_like(image_latents)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            image_latents = torch.cat([negative_image_latents, image_latents])
+
+        return image_latents
+
+    def _get_add_time_ids(
+        self,
+        fps: int,
+        motion_bucket_id: int,
+        noise_aug_strength: float,
+        dtype: torch.dtype,
+        batch_size: int,
+        num_videos_per_prompt: int,
+        do_classifier_free_guidance: bool,
+    ):
+        add_time_ids = [fps, motion_bucket_id, noise_aug_strength]
+
+        passed_add_embed_dim = self.unet.config.addition_time_embed_dim * len(add_time_ids)
+        expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+        if expected_add_embed_dim != passed_add_embed_dim:
+            raise ValueError(
+                f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+            )
+
+        add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+        add_time_ids = add_time_ids.repeat(batch_size * num_videos_per_prompt, 1)
+
+        if do_classifier_free_guidance:
+            add_time_ids = torch.cat([add_time_ids, add_time_ids])
+
+        return add_time_ids
+
+    def decode_latents(self, latents: torch.Tensor, num_frames: int, decode_chunk_size: int = 14):
+        # [batch, frames, channels, height, width] -> [batch*frames, channels, height, width]
+        latents = latents.flatten(0, 1)
+
+        latents = 1 / self.vae.config.scaling_factor * latents
+
+        forward_vae_fn = self.vae._orig_mod.forward if is_compiled_module(self.vae) else self.vae.forward
+        accepts_num_frames = "num_frames" in set(inspect.signature(forward_vae_fn).parameters.keys())
+
+        # decode decode_chunk_size frames at a time to avoid OOM
+        frames = []
+        for i in range(0, latents.shape[0], decode_chunk_size):
+            num_frames_in = latents[i : i + decode_chunk_size].shape[0]
+            decode_kwargs = {}
+            if accepts_num_frames:
+                # we only pass num_frames_in if it's expected
+                decode_kwargs["num_frames"] = num_frames_in
+
+            frame = self.vae.decode(latents[i : i + decode_chunk_size], **decode_kwargs).sample
+            frames.append(frame)
+        frames = torch.cat(frames, dim=0)
+
+        # [batch*frames, channels, height, width] -> [batch, channels, frames, height, width]
+        frames = frames.reshape(-1, num_frames, *frames.shape[1:]).permute(0, 2, 1, 3, 4)
+
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+        frames = frames.float()
+        return frames
+
+    def check_inputs(self, image, height, width):
+        if (
+            not isinstance(image, torch.Tensor)
+            and not isinstance(image, PIL.Image.Image)
+            and not isinstance(image, list)
+        ):
+            raise ValueError(
+                "`image` has to be of type `torch.Tensor` or `PIL.Image.Image` or `List[PIL.Image.Image]` but is"
+                f" {type(image)}"
+            )
+
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+    def prepare_latents(
+        self,
+        batch_size: int,
+        num_frames: int,
+        num_channels_latents: int,
+        height: int,
+        width: int,
+        dtype: torch.dtype,
+        device: Union[str, torch.device],
+        generator: torch.Generator,
+        latents: Optional[torch.Tensor] = None,
+    ):
+        shape = (
+            batch_size,
+            num_frames,
+            num_channels_latents // 2,
+            height // self.vae_scale_factor,
+            width // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+    # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1`
+    # corresponds to doing no classifier free guidance.
+    @property
+    def do_classifier_free_guidance(self):
+        if isinstance(self.guidance_scale, (int, float)):
+            return self.guidance_scale > 1
+        return self.guidance_scale.max() > 1
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        image: Union[PIL.Image.Image, List[PIL.Image.Image], torch.Tensor],
+        height: int = 576,
+        width: int = 1024,
+        num_frames: Optional[int] = None,
+        num_inference_steps: int = 25,
+        sigmas: Optional[List[float]] = None,
+        min_guidance_scale: float = 1.0,
+        max_guidance_scale: float = 3.0,
+        fps: int = 7,
+        motion_bucket_id: int = 127,
+        noise_aug_strength: float = 0.02,
+        decode_chunk_size: Optional[int] = None,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        return_dict: bool = True,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.Tensor`):
+                Image(s) to guide image generation. If you provide a tensor, the expected value range is between `[0,
+                1]`.
+            height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+                The width in pixels of the generated image.
+            num_frames (`int`, *optional*):
+                The number of video frames to generate. Defaults to `self.unet.config.num_frames` (14 for
+                `stable-video-diffusion-img2vid` and to 25 for `stable-video-diffusion-img2vid-xt`).
+            num_inference_steps (`int`, *optional*, defaults to 25):
+                The number of denoising steps. More denoising steps usually lead to a higher quality video at the
+                expense of slower inference. This parameter is modulated by `strength`.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            min_guidance_scale (`float`, *optional*, defaults to 1.0):
+                The minimum guidance scale. Used for the classifier free guidance with first frame.
+            max_guidance_scale (`float`, *optional*, defaults to 3.0):
+                The maximum guidance scale. Used for the classifier free guidance with last frame.
+            fps (`int`, *optional*, defaults to 7):
+                Frames per second. The rate at which the generated images shall be exported to a video after
+                generation. Note that Stable Diffusion Video's UNet was micro-conditioned on fps-1 during training.
+            motion_bucket_id (`int`, *optional*, defaults to 127):
+                Used for conditioning the amount of motion for the generation. The higher the number the more motion
+                will be in the video.
+            noise_aug_strength (`float`, *optional*, defaults to 0.02):
+                The amount of noise added to the init image, the higher it is the less the video will look like the
+                init image. Increase it for more motion.
+            decode_chunk_size (`int`, *optional*):
+                The number of frames to decode at a time. Higher chunk size leads to better temporal consistency at the
+                expense of more memory usage. By default, the decoder decodes all frames at once for maximal quality.
+                For lower memory usage, reduce `decode_chunk_size`.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of videos to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `pil`, `np` or `pt`.
+            callback_on_step_end (`Callable`, *optional*):
+                A function that is called at the end of each denoising step during inference. The function is called
+                with the following arguments:
+                    `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`.
+                `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableVideoDiffusionPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableVideoDiffusionPipelineOutput`] is
+                returned, otherwise a `tuple` of (`List[List[PIL.Image.Image]]` or `np.ndarray` or `torch.Tensor`) is
+                returned.
+        """
+        # 0. Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        num_frames = num_frames if num_frames is not None else self.unet.config.num_frames
+        decode_chunk_size = decode_chunk_size if decode_chunk_size is not None else num_frames
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(image, height, width)
+
+        # 2. Define call parameters
+        if isinstance(image, PIL.Image.Image):
+            batch_size = 1
+        elif isinstance(image, list):
+            batch_size = len(image)
+        else:
+            batch_size = image.shape[0]
+        device = self._execution_device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        self._guidance_scale = max_guidance_scale
+
+        # 3. Encode input image
+        image_embeddings = self._encode_image(image, device, num_videos_per_prompt, self.do_classifier_free_guidance)
+
+        # NOTE: Stable Video Diffusion was conditioned on fps - 1, which is why it is reduced here.
+        # See: https://github.com/Stability-AI/generative-models/blob/ed0997173f98eaf8f4edf7ba5fe8f15c6b877fd3/scripts/sampling/simple_video_sample.py#L188
+        fps = fps - 1
+
+        # 4. Encode input image using VAE
+        image = self.video_processor.preprocess(image, height=height, width=width).to(device)
+        noise = randn_tensor(image.shape, generator=generator, device=device, dtype=image.dtype)
+        image = image + noise_aug_strength * noise
+
+        needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+        if needs_upcasting:
+            self.vae.to(dtype=torch.float32)
+
+        image_latents = self._encode_vae_image(
+            image,
+            device=device,
+            num_videos_per_prompt=num_videos_per_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+        )
+        image_latents = image_latents.to(image_embeddings.dtype)
+
+        # cast back to fp16 if needed
+        if needs_upcasting:
+            self.vae.to(dtype=torch.float16)
+
+        # Repeat the image latents for each frame so we can concatenate them with the noise
+        # image_latents [batch, channels, height, width] ->[batch, num_frames, channels, height, width]
+        image_latents = image_latents.unsqueeze(1).repeat(1, num_frames, 1, 1, 1)
+
+        # 5. Get Added Time IDs
+        added_time_ids = self._get_add_time_ids(
+            fps,
+            motion_bucket_id,
+            noise_aug_strength,
+            image_embeddings.dtype,
+            batch_size,
+            num_videos_per_prompt,
+            self.do_classifier_free_guidance,
+        )
+        added_time_ids = added_time_ids.to(device)
+
+        # 6. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, None, sigmas)
+
+        # 7. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_frames,
+            num_channels_latents,
+            height,
+            width,
+            image_embeddings.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 8. Prepare guidance scale
+        guidance_scale = torch.linspace(min_guidance_scale, max_guidance_scale, num_frames).unsqueeze(0)
+        guidance_scale = guidance_scale.to(device, latents.dtype)
+        guidance_scale = guidance_scale.repeat(batch_size * num_videos_per_prompt, 1)
+        guidance_scale = _append_dims(guidance_scale, latents.ndim)
+
+        self._guidance_scale = guidance_scale
+
+        # 9. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # Concatenate image_latents over channels dimension
+                latent_model_input = torch.cat([latent_model_input, image_latents], dim=2)
+
+                # predict the noise residual
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=image_embeddings,
+                    added_time_ids=added_time_ids,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents).prev_sample
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if not output_type == "latent":
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
+            frames = self.decode_latents(latents, num_frames, decode_chunk_size)
+            frames = self.video_processor.postprocess_video(video=frames, output_type=output_type)
+        else:
+            frames = latents
+
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return frames
+
+        return StableVideoDiffusionPipelineOutput(frames=frames)
+
+
+# resizing utils
+# TODO: clean up later
+def _resize_with_antialiasing(input, size, interpolation="bicubic", align_corners=True):
+    h, w = input.shape[-2:]
+    factors = (h / size[0], w / size[1])
+
+    # First, we have to determine sigma
+    # Taken from skimage: https://github.com/scikit-image/scikit-image/blob/v0.19.2/skimage/transform/_warps.py#L171
+    sigmas = (
+        max((factors[0] - 1.0) / 2.0, 0.001),
+        max((factors[1] - 1.0) / 2.0, 0.001),
+    )
+
+    # Now kernel size. Good results are for 3 sigma, but that is kind of slow. Pillow uses 1 sigma
+    # https://github.com/python-pillow/Pillow/blob/master/src/libImaging/Resample.c#L206
+    # But they do it in the 2 passes, which gives better results. Let's try 2 sigmas for now
+    ks = int(max(2.0 * 2 * sigmas[0], 3)), int(max(2.0 * 2 * sigmas[1], 3))
+
+    # Make sure it is odd
+    if (ks[0] % 2) == 0:
+        ks = ks[0] + 1, ks[1]
+
+    if (ks[1] % 2) == 0:
+        ks = ks[0], ks[1] + 1
+
+    input = _gaussian_blur2d(input, ks, sigmas)
+
+    output = torch.nn.functional.interpolate(input, size=size, mode=interpolation, align_corners=align_corners)
+    return output
+
+
+def _compute_padding(kernel_size):
+    """Compute padding tuple."""
+    # 4 or 6 ints:  (padding_left, padding_right,padding_top,padding_bottom)
+    # https://pytorch.org/docs/stable/nn.html#torch.nn.functional.pad
+    if len(kernel_size) < 2:
+        raise AssertionError(kernel_size)
+    computed = [k - 1 for k in kernel_size]
+
+    # for even kernels we need to do asymmetric padding :(
+    out_padding = 2 * len(kernel_size) * [0]
+
+    for i in range(len(kernel_size)):
+        computed_tmp = computed[-(i + 1)]
+
+        pad_front = computed_tmp // 2
+        pad_rear = computed_tmp - pad_front
+
+        out_padding[2 * i + 0] = pad_front
+        out_padding[2 * i + 1] = pad_rear
+
+    return out_padding
+
+
+def _filter2d(input, kernel):
+    # prepare kernel
+    b, c, h, w = input.shape
+    tmp_kernel = kernel[:, None, ...].to(device=input.device, dtype=input.dtype)
+
+    tmp_kernel = tmp_kernel.expand(-1, c, -1, -1)
+
+    height, width = tmp_kernel.shape[-2:]
+
+    padding_shape: List[int] = _compute_padding([height, width])
+    input = torch.nn.functional.pad(input, padding_shape, mode="reflect")
+
+    # kernel and input tensor reshape to align element-wise or batch-wise params
+    tmp_kernel = tmp_kernel.reshape(-1, 1, height, width)
+    input = input.view(-1, tmp_kernel.size(0), input.size(-2), input.size(-1))
+
+    # convolve the tensor with the kernel.
+    output = torch.nn.functional.conv2d(input, tmp_kernel, groups=tmp_kernel.size(0), padding=0, stride=1)
+
+    out = output.view(b, c, h, w)
+    return out
+
+
+def _gaussian(window_size: int, sigma):
+    if isinstance(sigma, float):
+        sigma = torch.tensor([[sigma]])
+
+    batch_size = sigma.shape[0]
+
+    x = (torch.arange(window_size, device=sigma.device, dtype=sigma.dtype) - window_size // 2).expand(batch_size, -1)
+
+    if window_size % 2 == 0:
+        x = x + 0.5
+
+    gauss = torch.exp(-x.pow(2.0) / (2 * sigma.pow(2.0)))
+
+    return gauss / gauss.sum(-1, keepdim=True)
+
+
+def _gaussian_blur2d(input, kernel_size, sigma):
+    if isinstance(sigma, tuple):
+        sigma = torch.tensor([sigma], dtype=input.dtype)
+    else:
+        sigma = sigma.to(dtype=input.dtype)
+
+    ky, kx = int(kernel_size[0]), int(kernel_size[1])
+    bs = sigma.shape[0]
+    kernel_x = _gaussian(kx, sigma[:, 1].view(bs, 1))
+    kernel_y = _gaussian(ky, sigma[:, 0].view(bs, 1))
+    out_x = _filter2d(input, kernel_x[..., None, :])
+    out = _filter2d(out_x, kernel_y[..., None])
+
+    return out

ctrl_world/src/models/unet_spatio_temporal_condition.py

@@ -0,0 +1,522 @@
+from dataclasses import dataclass
+from typing import Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import UNet2DConditionLoadersMixin
+from diffusers.utils import BaseOutput, logging
+from diffusers.models.attention_processor import CROSS_ATTENTION_PROCESSORS, AttentionProcessor, AttnProcessor
+from diffusers.models.embeddings import TimestepEmbedding, Timesteps
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.unets.unet_3d_blocks import UNetMidBlockSpatioTemporal, get_down_block, get_up_block
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNetSpatioTemporalConditionOutput(BaseOutput):
+    """
+    The output of [`UNetSpatioTemporalConditionModel`].
+
+    Args:
+        sample (`torch.Tensor` of shape `(batch_size, num_frames, num_channels, height, width)`):
+            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: torch.Tensor = None
+
+
+class UNetSpatioTemporalConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+    r"""
+    A conditional Spatio-Temporal UNet model that takes a noisy video frames, conditional state, and a timestep and
+    returns a sample shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample.
+        in_channels (`int`, *optional*, defaults to 8): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlockSpatioTemporal", "CrossAttnDownBlockSpatioTemporal", "CrossAttnDownBlockSpatioTemporal", "DownBlockSpatioTemporal")`):
+            The tuple of downsample blocks to use.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlockSpatioTemporal", "CrossAttnUpBlockSpatioTemporal", "CrossAttnUpBlockSpatioTemporal", "CrossAttnUpBlockSpatioTemporal")`):
+            The tuple of upsample blocks to use.
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        addition_time_embed_dim: (`int`, defaults to 256):
+            Dimension to to encode the additional time ids.
+        projection_class_embeddings_input_dim (`int`, defaults to 768):
+            The dimension of the projection of encoded `added_time_ids`.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unets.unet_3d_blocks.CrossAttnDownBlockSpatioTemporal`],
+            [`~models.unets.unet_3d_blocks.CrossAttnUpBlockSpatioTemporal`],
+            [`~models.unets.unet_3d_blocks.UNetMidBlockSpatioTemporal`].
+        num_attention_heads (`int`, `Tuple[int]`, defaults to `(5, 10, 10, 20)`):
+            The number of attention heads.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 8,
+        out_channels: int = 4,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlockSpatioTemporal",
+            "CrossAttnDownBlockSpatioTemporal",
+            "CrossAttnDownBlockSpatioTemporal",
+            "DownBlockSpatioTemporal",
+        ),
+        up_block_types: Tuple[str] = (
+            "UpBlockSpatioTemporal",
+            "CrossAttnUpBlockSpatioTemporal",
+            "CrossAttnUpBlockSpatioTemporal",
+            "CrossAttnUpBlockSpatioTemporal",
+        ),
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        addition_time_embed_dim: int = 256,
+        projection_class_embeddings_input_dim: int = 768,
+        layers_per_block: Union[int, Tuple[int]] = 2,
+        cross_attention_dim: Union[int, Tuple[int]] = 1024,
+        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+        num_attention_heads: Union[int, Tuple[int]] = (5, 10, 20, 20),
+        num_frames: int = 25,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+
+        # Check inputs
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+            )
+
+        # input
+        self.conv_in = nn.Conv2d(
+            in_channels,
+            block_out_channels[0],
+            kernel_size=3,
+            padding=1,
+        )
+
+        # time
+        time_embed_dim = block_out_channels[0] * 4
+
+        self.time_proj = Timesteps(block_out_channels[0], True, downscale_freq_shift=0)
+        timestep_input_dim = block_out_channels[0]
+
+        self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+
+        self.add_time_proj = Timesteps(addition_time_embed_dim, True, downscale_freq_shift=0)
+        self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+
+        self.down_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if isinstance(layers_per_block, int):
+            layers_per_block = [layers_per_block] * len(down_block_types)
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+        blocks_time_embed_dim = time_embed_dim
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block[i],
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=1e-5,
+                cross_attention_dim=cross_attention_dim[i],
+                num_attention_heads=num_attention_heads[i],
+                resnet_act_fn="silu",
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        self.mid_block = UNetMidBlockSpatioTemporal(
+            block_out_channels[-1],
+            temb_channels=blocks_time_embed_dim,
+            transformer_layers_per_block=transformer_layers_per_block[-1],
+            cross_attention_dim=cross_attention_dim[-1],
+            num_attention_heads=num_attention_heads[-1],
+        )
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_num_attention_heads = list(reversed(num_attention_heads))
+        reversed_layers_per_block = list(reversed(layers_per_block))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+        reversed_transformer_layers_per_block = list(reversed(transformer_layers_per_block))
+
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=reversed_layers_per_block[i] + 1,
+                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=1e-5,
+                resolution_idx=i,
+                cross_attention_dim=reversed_cross_attention_dim[i],
+                num_attention_heads=reversed_num_attention_heads[i],
+                resnet_act_fn="silu",
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=32, eps=1e-5)
+        self.conv_act = nn.SiLU()
+
+        self.conv_out = nn.Conv2d(
+            block_out_channels[0],
+            out_channels,
+            kernel_size=3,
+            padding=1,
+        )
+
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(
+            name: str,
+            module: torch.nn.Module,
+            processors: Dict[str, AttentionProcessor],
+        ):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor()
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor)
+
+    # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.enable_forward_chunking
+    def enable_forward_chunking(self, chunk_size: Optional[int] = None, dim: int = 0) -> None:
+        """
+        Sets the attention processor to use [feed forward
+        chunking](https://huggingface.co/blog/reformer#2-chunked-feed-forward-layers).
+
+        Parameters:
+            chunk_size (`int`, *optional*):
+                The chunk size of the feed-forward layers. If not specified, will run feed-forward layer individually
+                over each tensor of dim=`dim`.
+            dim (`int`, *optional*, defaults to `0`):
+                The dimension over which the feed-forward computation should be chunked. Choose between dim=0 (batch)
+                or dim=1 (sequence length).
+        """
+        if dim not in [0, 1]:
+            raise ValueError(f"Make sure to set `dim` to either 0 or 1, not {dim}")
+
+        # By default chunk size is 1
+        chunk_size = chunk_size or 1
+
+        def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
+            if hasattr(module, "set_chunk_feed_forward"):
+                module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
+
+            for child in module.children():
+                fn_recursive_feed_forward(child, chunk_size, dim)
+
+        for module in self.children():
+            fn_recursive_feed_forward(module, chunk_size, dim)
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        added_time_ids: torch.Tensor,
+        return_dict: bool = True,
+        frame_level_cond=False,
+    ) -> Union[UNetSpatioTemporalConditionOutput, Tuple]:
+        r"""
+        The [`UNetSpatioTemporalConditionModel`] forward method.
+
+        Args:
+            sample (`torch.Tensor`):
+                The noisy input tensor with the following shape `(batch, num_frames, channel, height, width)`.
+            timestep (`torch.Tensor` or `float` or `int`): The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.Tensor`):
+                The encoder hidden states with shape `(batch, sequence_length, cross_attention_dim)`.
+            added_time_ids: (`torch.Tensor`):
+                The additional time ids with shape `(batch, num_additional_ids)`. These are encoded with sinusoidal
+                embeddings and added to the time embeddings.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_slatio_temporal.UNetSpatioTemporalConditionOutput`] instead
+                of a plain tuple.
+        Returns:
+            [`~models.unet_slatio_temporal.UNetSpatioTemporalConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unet_slatio_temporal.UNetSpatioTemporalConditionOutput`] is
+                returned, otherwise a `tuple` is returned where the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+            logger.info("Forward upsample size to force interpolation output size.")
+            forward_upsample_size = True
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            is_npu = sample.device.type == "npu"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if (is_mps or is_npu) else torch.float64
+            else:
+                dtype = torch.int32 if (is_mps or is_npu) else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        batch_size, num_frames = sample.shape[:2]
+        timesteps = timesteps.expand(batch_size)
+
+        t_emb = self.time_proj(timesteps)
+
+        # `Timesteps` does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+
+        emb = self.time_embedding(t_emb)
+
+        time_embeds = self.add_time_proj(added_time_ids.flatten())
+        time_embeds = time_embeds.reshape((batch_size, -1))
+        time_embeds = time_embeds.to(emb.dtype)
+        aug_emb = self.add_embedding(time_embeds)
+        emb = emb + aug_emb
+
+        # Flatten the batch and frames dimensions
+        # sample: [batch, frames, channels, height, width] -> [batch * frames, channels, height, width]
+        sample = sample.flatten(0, 1)
+        # Repeat the embeddings num_video_frames times
+        # emb: [batch, channels] -> [batch * frames, channels]
+        emb = emb.repeat_interleave(num_frames, dim=0, output_size=emb.shape[0] * num_frames)
+
+        # encoder_hidden_states: [batch, 1, channels] -> [batch * frames, 1, channels]
+        # encoder_hidden_states = encoder_hidden_states.repeat_interleave(
+        #     num_frames, dim=0, output_size=encoder_hidden_states.shape[0] * num_frames
+        # )
+
+        ############################# newly added to support frame_level pose conditioning ########################################
+        # print('new one!!!!!!!!!')
+        if not frame_level_cond:
+            encoder_hidden_states = encoder_hidden_states.repeat_interleave(num_frames, dim=0)
+        else:
+            encoder_hidden_states = encoder_hidden_states.reshape(batch_size * num_frames, -1, encoder_hidden_states.shape[-1])
+        ############################################################################################################################
+
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        image_only_indicator = torch.zeros(batch_size, num_frames, dtype=sample.dtype, device=sample.device)
+
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    image_only_indicator=image_only_indicator,
+                )
+            else:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    image_only_indicator=image_only_indicator,
+                )
+
+            down_block_res_samples += res_samples
+
+        # 4. mid
+        sample = self.mid_block(
+            hidden_states=sample,
+            temb=emb,
+            encoder_hidden_states=encoder_hidden_states,
+            image_only_indicator=image_only_indicator,
+        )
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    upsample_size=upsample_size,
+                    image_only_indicator=image_only_indicator,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                    image_only_indicator=image_only_indicator,
+                )
+
+        # 6. post-process
+        sample = self.conv_norm_out(sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        # 7. Reshape back to original shape
+        sample = sample.reshape(batch_size, num_frames, *sample.shape[1:])
+
+        if not return_dict:
+            return (sample,)
+
+        return UNetSpatioTemporalConditionOutput(sample=sample)

ctrl_world/src/world_model.py

@@ -0,0 +1,185 @@
+from models.pipeline_stable_video_diffusion import StableVideoDiffusionPipeline
+from models.unet_spatio_temporal_condition import UNetSpatioTemporalConditionModel
+from models.pipeline_ctrl_world import CtrlWorldDiffusionPipeline
+import torch
+import torch.nn as nn
+
+import json
+import einops
+import numpy as np
+from huggingface_hub import snapshot_download
+from transformers import AutoTokenizer, CLIPTextModelWithProjection
+
+class Action_encoder2(nn.Module):
+    def __init__(self, action_dim, action_num, hidden_size, text_cond=True):
+        super().__init__()
+        self.action_dim = action_dim
+        self.action_num = action_num
+        self.hidden_size = hidden_size
+        self.text_cond = text_cond
+
+        input_dim = int(action_dim)
+        self.action_encode = nn.Sequential(
+            nn.Linear(input_dim, 1024),
+            nn.SiLU(),
+            nn.Linear(1024, 1024),
+            nn.SiLU(),
+            nn.Linear(1024, 1024)
+        )
+        # kaiming initialization
+        nn.init.kaiming_normal_(self.action_encode[0].weight, mode='fan_in', nonlinearity='relu')
+        nn.init.kaiming_normal_(self.action_encode[2].weight, mode='fan_in', nonlinearity='relu')
+
+    def forward(self, action,  texts=None, text_tokinizer=None, text_encoder=None, frame_level_cond=True,):
+        # action: (B, action_num, action_dim)
+        B,T,D = action.shape
+        if not frame_level_cond:
+            action = einops.rearrange(action, 'b t d -> b 1 (t d)')
+        action = self.action_encode(action)
+
+        if texts is not None and self.text_cond:
+            # with 50% probability, add text condition
+            with torch.no_grad():
+                inputs = text_tokinizer(texts, padding='max_length', return_tensors="pt", truncation=True).to(text_encoder.device)
+                outputs = text_encoder(**inputs)
+                hidden_text = outputs.text_embeds # (B, 512)
+                hidden_text = einops.repeat(hidden_text, 'b c -> b 1 (n c)', n=2) # (B, 1, 1024)
+            
+            action = action + hidden_text # (B, T, hidden_size)
+        return action # (B, 1, hidden_size) or (B, T, hidden_size) if frame_level_cond
+
+
+class CrtlWorld(nn.Module):
+    def __init__(self, config: dict):
+        super(CrtlWorld, self).__init__()
+
+        self.config = config
+        # load from pretrained stable video diffusion
+        model_local_path = snapshot_download(
+            repo_id=config["svd_model_path"],   # e.g. "stabilityai/stable-video-diffusion-img2vid"
+            repo_type="model"
+        )
+
+        # Load pipeline from downloaded path
+        self.pipeline = StableVideoDiffusionPipeline.from_pretrained(
+            model_local_path,
+            torch_dtype="auto"
+        )
+
+
+        unet = UNetSpatioTemporalConditionModel()
+        unet.load_state_dict(self.pipeline.unet.state_dict(), strict=False)
+        self.pipeline.unet = unet
+        
+        self.unet = self.pipeline.unet
+        self.vae = self.pipeline.vae
+        self.image_encoder = self.pipeline.image_encoder
+        self.scheduler = self.pipeline.scheduler
+
+        # freeze vae, image_encoder, enable unet gradient ckpt
+        self.vae.requires_grad_(False)
+        self.image_encoder.requires_grad_(False)
+        self.unet.requires_grad_(True)
+        self.unet.enable_gradient_checkpointing()
+
+        # SVD is a img2video model, load a clip text encoder
+        
+        model_local_path = snapshot_download(
+            repo_id=config["clip_model_path"],   # e.g. "stabilityai/stable-video-diffusion-img2vid"
+            repo_type="model"
+        )
+
+        self.text_encoder = CLIPTextModelWithProjection.from_pretrained(
+            model_local_path,
+            torch_dtype="auto"
+        )
+        self.tokenizer = AutoTokenizer.from_pretrained(model_local_path, use_fast=False)
+        self.text_encoder.requires_grad_(False)
+
+        # initialize an action projector
+        self.action_encoder = Action_encoder2(action_dim=config["action_dim"], action_num=int(config["num_history"]+config["num_frames"]), hidden_size=1024, text_cond=config["text_cond"])
+
+        with open(f"{config["data_stat_path"]}", 'r') as f:
+            data_stat = json.load(f)
+            self.state_p01 = np.array(data_stat['state_01'])[None,:]
+            self.state_p99 = np.array(data_stat['state_99'])[None,:]
+    
+    def normalize_bound(
+        self,
+        data: np.ndarray,
+        clip_min: float = -1,
+        clip_max: float = 1,
+        eps: float = 1e-8,
+    ) -> np.ndarray:
+        ndata = 2 * (data - self.state_p01) / (self.state_p99 - self.state_p01 + eps) - 1
+        return np.clip(ndata, clip_min, clip_max)
+    
+    def decode(self, latents: torch.Tensor):
+
+        bsz, frame_num = latents.shape[:2]
+        x = latents.flatten(0, 1)
+
+        decoded = []
+        chunk_size = self.config["decode_chunk_size"]
+        for i in range(0, x.shape[0], chunk_size):
+            chunk = x[i:i + chunk_size] / self.pipeline.vae.config.scaling_factor
+            decode_kwargs = {"num_frames": chunk.shape[0]}
+            out = self.pipeline.vae.decode(chunk, **decode_kwargs).sample
+            decoded.append(out)
+
+        videos = torch.cat(decoded, dim=0)
+        videos = videos.reshape(bsz, frame_num, *videos.shape[1:])
+        videos = ((videos / 2.0 + 0.5).clamp(0, 1))
+        videos = videos.detach().float().cpu()
+
+    def encode(self, img: torch.Tensor):
+
+        x = img.unsqueeze(0)
+        x = x * 2 - 1  # [0,1] → [-1,1]
+
+        vae = self.pipeline.vae
+        with torch.no_grad():
+            latent = vae.encode(x).latent_dist.sample()
+            latent = latent * vae.config.scaling_factor
+
+        return latent.detach()
+    
+    def action_text_encode(self, action: torch.Tensor, text):
+
+        action_tensor = action.unsqueeze(0)
+
+        # ── Encode action (+ optional text) ───────────────────
+        with torch.no_grad():
+            if text is not None and self.config["text_cond"]:
+                text_token = self.action_encoder(action_tensor, [text], self.tokenizer, self.text_encoder)
+            else:
+                text_token = self.action_encoder(action_tensor)
+
+        return text_token.detach()
+
+    def get_latent_views(self, frames, current_latent, text_token):
+
+        his_cond = torch.cat(frames, dim=0).unsqueeze(0)  # (1, num_history, 4, stacked_H, W)
+
+        # ── Run CtrlWorldDiffusionPipeline ────────────────────
+        with torch.no_grad():
+            _, latents = CtrlWorldDiffusionPipeline.__call__(
+                self.pipeline,
+                image=current_latent,
+                text=text_token,
+                width=self.config["width"],
+                height=int(self.config["height"] * 3),  # 3 views stacked
+                num_frames=self.config["num_frames"],
+                history=his_cond,
+                num_inference_steps=self.config["num_inference_steps"],
+                decode_chunk_size=self.config["decode_chunk_size"],
+                max_guidance_scale=self.config["guidance_scale"],
+                fps=self.config["fps"],
+                motion_bucket_id=self.config["motion_bucket_id"],
+                mask=None,
+                output_type="latent",
+                return_dict=False,
+                frame_level_cond=True,
+            )
+
+        return latents