| 1 | import os |
| 2 | import logging |
| 3 | import cv2 |
| 4 | from typing import List, Tuple, Union, Optional |
| 5 | from pydantic import BaseModel, Field |
| 6 | from tenacity import retry, stop_after_attempt |
| 7 | from langchain_core.messages import HumanMessage, SystemMessage |
| 8 | from langchain_core.output_parsers import PydanticOutputParser |
| 9 | from utils.robust_json_parser import TrailingCommaTolerantPydanticOutputParser as PydanticOutputParser |
| 10 | from scenedetect import open_video, SceneManager, split_video_ffmpeg |
| 11 | from scenedetect.detectors import ContentDetector |
| 12 | |
| 13 | from interfaces import ShotDescription, ShotBriefDescription, Camera, ImageOutput, VideoOutput |
| 14 | |
| 15 | |
| 16 | from moviepy import VideoFileClip |
| 17 | from PIL import Image |
| 18 | |
| 19 | |
| 20 | system_prompt_template_select_reference_camera = \ |
| 21 | """ |
| 22 | [Role] |
| 23 | You are a professional video editing expert specializing in multi-camera shot analysis and scene structure modeling. You have deep knowledge of cinematic language, enabling you to understand shot sizes (e.g., wide shot, medium shot, close-up) and content inclusion relationships. You can infer hierarchical structures between camera positions based on corresponding shot descriptions. |
| 24 | |
| 25 | [Task] |
| 26 | Your task is to analyze the input camera position data to construct a "camera position tree". This tree structure represents a relationship where a parent camera's content encompasses that of a child camera. Specifically, you need to identify the parent camera for each camera position (if one exists) and determine the dependent shot indices (i.e., the specific shots within the parent camera's footage that contain the child camera's content). If a camera position has no parent, output None. |
| 27 | |
| 28 | [Input] |
| 29 | The input is a sequence of cameras. The sequence will be enclosed within <CAMERA_SEQ> and </CAMERA_SEQ>. |
| 30 | Each camera contains a sequence of shots filmed by the camera, which will be enclosed within <CAMERA_N> and </CAMERA_N>, where N is the index of the camera. |
| 31 | |
| 32 | Below is an example of the input format: |
| 33 | |
| 34 | <CAMERA_SEQ> |
| 35 | <CAMERA_0> |
| 36 | Shot 0: Medium shot of the street. Alice and Bob are walking towards each other. |
| 37 | Shot 2: Medium shot of the street. Alice and Bob hug each other. |
| 38 | </CAMERA_0> |
| 39 | <CAMERA_1> |
| 40 | Shot 1: Close-up of the Alice's face. Her expression shifts from surprise to delight as she recognizes Bob. |
| 41 | </CAMERA_1> |
| 42 | </CAMERA_SEQ> |
| 43 | |
| 44 | |
| 45 | [Output] |
| 46 | {format_instructions} |
| 47 | |
| 48 | [Guidelines] |
| 49 | - The language of all output values (not include keys) should be consistent with the language of the input. |
| 50 | - Content Inclusion Check: The parent camera should as fully as possible contain the child camera's content in certain shots (e.g., a parent medium two-shot encompasses a child over-the-shoulder reverse shot). Analyze shot descriptions by comparing keywords (e.g., characters, actions, setting) to ensure the parent shot's field of view covers the child shot's. |
| 51 | - Transition Smoothness Priority: Larger shot size as parent camera is preferred, such as Wide Shot -> Medium Shot or Medium Shot -> Close-up. The shot sizes of adjacent parent and child nodes should be as similar as possible. A direct transition from a long shot to a close-up is not allowed unless absolutely necessary. |
| 52 | - Temporal Proximity: Each camera is described by its corresponding first shot, and the parent camera is located based on the description of the first shot. The shot index of the parent camera should be as close as possible to the first shot index of the child camera. |
| 53 | - Logical Consistency: The camera tree should be acyclic, avoid circular dependencies. If a camera is contained by multiple potential parents, select the best match (based on shot size and content). If there is no suitable parent camera, output None. |
| 54 | - When a broader perspective is not available, choose the shot with the largest overlapping field of view as the parent (the one with the most information overlap), or a shot can also serve as the parent of a reverse shot. When two cameras can be the parent of each other, choose the one with the smaller index as the parent of the camera with the larger index. |
| 55 | - Only one camera can exist without a parent. |
| 56 | - When describing the elements lost in a shot, carefully compare the details between the parent shot and the child shot. For example, the parent shot is a medium shot of Character A and Character B facing each other (both in profile to the camera), while the child shot is a close-up of Character A (with Character A facing the camera directly). In this case, the child shot lacks the frontal view information of Character A. |
| 57 | - The first camera must be the root of the camera tree. |
| 58 | """ |
| 59 | |
| 60 | |
| 61 | human_prompt_template_select_reference_camera = \ |
| 62 | """ |
| 63 | <CAMERA_SEQ> |
| 64 | {camera_seq_str} |
| 65 | </CAMERA_SEQ> |
| 66 | """ |
| 67 | |
| 68 | |
| 69 | class CameraParentItem(BaseModel): |
| 70 | parent_cam_idx: Optional[int] = Field( |
| 71 | default=None, |
| 72 | description="The index of the parent camera. Set to None if the camera has no parent (e.g., for a root camera).", |
| 73 | examples=[0, 1, None], |
| 74 | ) |
| 75 | parent_shot_idx: Optional[int] = Field( |
| 76 | default=None, |
| 77 | description="The index of the dependent shot. Set to None if the camera has no parent (e.g., for a root camera).", |
| 78 | examples=[0, 3, None], |
| 79 | ) |
| 80 | reason: str = Field( |
| 81 | description="The reason for the selection of the parent camera. If the camera has no parent, it should explain why it's a root camera.", |
| 82 | examples=[ |
| 83 | "The parent shot's field of view covers the child shot's field of view (from medium shot to close-up)", |
| 84 | "The parent shot and the child shot have a shot/reverse shot relationship.", |
| 85 | "CAMERA_0 (Shot 0) establishes the entire scene and contains all characters and the setting. It is the root camera." # 补充 LLM 实际输出的例子 |
| 86 | ], |
| 87 | ) |
| 88 | is_parent_fully_covers_child: Optional[bool] = Field( |
| 89 | default=None, |
| 90 | description="Whether the parent camera fully covers the child camera's content. Set to None if the camera has no parent.", |
| 91 | examples=[True, False, None], |
| 92 | ) |
| 93 | missing_info: Optional[str] = Field( |
| 94 | default=None, |
| 95 | description="The missing elements in the child shot that are not covered by the parent shot. If the parent shot fully covers the child shot, set this to None.", |
| 96 | examples=[ |
| 97 | "The frontal view of Alice.", |
| 98 | None, |
| 99 | ], |
| 100 | ) |
| 101 | |
| 102 | class CameraTreeResponse(BaseModel): |
| 103 | camera_parent_items: List[Optional[CameraParentItem]] = Field( |
| 104 | description="The parent camera items for each camera. If a camera has no parent, set this to None. The length of the list should be the same as the number of cameras.", |
| 105 | ) |
| 106 | |
| 107 | |
| 108 | |
| 109 | class CameraImageGenerator: |
| 110 | |
| 111 | def __init__( |
| 112 | self, |
| 113 | chat_model, |
| 114 | image_generator, |
| 115 | video_generator, |
| 116 | ): |
| 117 | self.chat_model = chat_model |
| 118 | self.image_generator = image_generator |
| 119 | self.video_generator = video_generator |
| 120 | |
| 121 | |
| 122 | async def construct_camera_tree( |
| 123 | self, |
| 124 | cameras: List[Camera], |
| 125 | shot_descs: List[Union[ShotDescription, ShotBriefDescription]], |
| 126 | ) -> List[Camera]: |
| 127 | parser = PydanticOutputParser(pydantic_object=CameraTreeResponse) |
| 128 | shot_desc_by_idx = {shot.idx: shot for shot in shot_descs} |
| 129 | |
| 130 | camera_seq_str = "<CAMERA_SEQ>\n" |
| 131 | for cam in cameras: |
| 132 | camera_seq_str += f"<CAMERA_{cam.idx}>\n" |
| 133 | for shot_idx in cam.active_shot_idxs: |
| 134 | shot_desc = shot_desc_by_idx.get(shot_idx) |
| 135 | if shot_desc is None: |
| 136 | raise ValueError(f"Camera {cam.idx} references missing shot {shot_idx}") |
| 137 | camera_seq_str += f"Shot {shot_idx}: {shot_desc.visual_desc}\n" |
| 138 | camera_seq_str += f"</CAMERA_{cam.idx}>\n" |
| 139 | camera_seq_str += "</CAMERA_SEQ>" |
| 140 | |
| 141 | messages = [ |
| 142 | SystemMessage(content=system_prompt_template_select_reference_camera.format(format_instructions=parser.get_format_instructions())), |
| 143 | HumanMessage(content=human_prompt_template_select_reference_camera.format(camera_seq_str=camera_seq_str)), |
| 144 | ] |
| 145 | |
| 146 | chain = self.chat_model | parser |
| 147 | response: CameraTreeResponse = await chain.ainvoke(messages) |
| 148 | parent_items = response.camera_parent_items |
| 149 | if len(parent_items) != len(cameras): |
| 150 | raise ValueError(f"Camera tree response length mismatch: expected {len(cameras)}, got {len(parent_items)}") |
| 151 | |
| 152 | valid_camera_idxs = {cam.idx for cam in cameras} |
| 153 | valid_shot_idxs = set(shot_desc_by_idx) |
| 154 | parent_by_camera = {} |
| 155 | for cam, parent_cam_item in zip(cameras, parent_items): |
| 156 | parent_cam_idx = parent_cam_item.parent_cam_idx if parent_cam_item is not None else None |
| 157 | parent_shot_idx = parent_cam_item.parent_shot_idx if parent_cam_item is not None else None |
| 158 | if parent_cam_idx is not None and parent_cam_idx not in valid_camera_idxs: |
| 159 | raise ValueError(f"Camera {cam.idx} has invalid parent camera {parent_cam_idx}") |
| 160 | if parent_cam_idx == cam.idx: |
| 161 | raise ValueError(f"Camera {cam.idx} cannot be its own parent") |
| 162 | if parent_shot_idx is not None and parent_shot_idx not in valid_shot_idxs: |
| 163 | raise ValueError(f"Camera {cam.idx} has invalid parent shot {parent_shot_idx}") |
| 164 | parent_by_camera[cam.idx] = parent_cam_idx |
| 165 | |
| 166 | for cam in cameras: |
| 167 | seen = set() |
| 168 | current = cam.idx |
| 169 | while parent_by_camera.get(current) is not None: |
| 170 | current = parent_by_camera[current] |
| 171 | if current in seen: |
| 172 | raise ValueError(f"Camera tree contains a cycle involving camera {cam.idx}") |
| 173 | seen.add(current) |
| 174 | |
| 175 | for cam, parent_cam_item in zip(cameras, parent_items): |
| 176 | cam.parent_cam_idx = parent_cam_item.parent_cam_idx if parent_cam_item is not None else None |
| 177 | cam.parent_shot_idx = parent_cam_item.parent_shot_idx if parent_cam_item is not None else None |
| 178 | cam.reason = parent_cam_item.reason if parent_cam_item is not None else None |
| 179 | cam.is_parent_fully_covers_child = parent_cam_item.is_parent_fully_covers_child if parent_cam_item is not None else None |
| 180 | cam.missing_info = parent_cam_item.missing_info if parent_cam_item is not None else None |
| 181 | return cameras |
| 182 | |
| 183 | |
| 184 | async def generate_transition_video( |
| 185 | self, |
| 186 | first_shot_visual_desc: str, |
| 187 | second_shot_visual_desc: str, |
| 188 | first_shot_ff_path: str, |
| 189 | progress=None, |
| 190 | ) -> VideoOutput: |
| 191 | |
| 192 | prompt = f"Two shots. The transition between the shots is a cut to. The style of the two shots should be consistent." |
| 193 | prompt += f"\nThe first shot description: {first_shot_visual_desc}." |
| 194 | prompt += f"\nThe second shot description: {second_shot_visual_desc}." |
| 195 | reference_image_paths = [first_shot_ff_path] |
| 196 | video_output = await self.video_generator.generate_single_video( |
| 197 | prompt=prompt, |
| 198 | reference_image_paths=reference_image_paths, |
| 199 | progress=progress, |
| 200 | ) |
| 201 | return video_output |
| 202 | |
| 203 | |
| 204 | def get_new_camera_image( |
| 205 | self, |
| 206 | transition_video_path: str, |
| 207 | ) -> ImageOutput: |
| 208 | video = open_video(transition_video_path) |
| 209 | scene_manager = SceneManager() |
| 210 | scene_manager.add_detector(ContentDetector()) |
| 211 | scene_manager.detect_scenes(video, show_progress=False) |
| 212 | scene_list = scene_manager.get_scene_list() |
| 213 | output_dir = os.path.join(os.path.dirname(transition_video_path), "cache") |
| 214 | os.makedirs(output_dir, exist_ok=True) |
| 215 | split_video_ffmpeg(transition_video_path, scene_list, output_dir, show_progress=True) |
| 216 | |
| 217 | |
| 218 | video_name = os.path.basename(transition_video_path).split('.')[0] |
| 219 | second_video_path = os.path.join(output_dir, f"{video_name}-Scene-002.mp4") |
| 220 | if os.path.exists(second_video_path): |
| 221 | # use first frame of second shot as new camera image |
| 222 | clip = VideoFileClip(second_video_path) |
| 223 | ff = clip.get_frame(0) |
| 224 | ff = Image.fromarray(ff.astype('uint8'), 'RGB') |
| 225 | return ImageOutput(fmt="pil", ext="png", data=ff) |
| 226 | else: |
| 227 | # use last frame of transition video to instead |
| 228 | clip = VideoFileClip(transition_video_path) |
| 229 | lf_time = clip.duration - (1 / clip.fps) |
| 230 | lf_time = max(0, lf_time) |
| 231 | lf = clip.get_frame(lf_time) |
| 232 | lf = Image.fromarray(lf.astype('uint8'), 'RGB') |
| 233 | return ImageOutput(fmt="pil", ext="png", data=lf) |
| 234 | |
| 235 | |
| 236 | async def generate_first_frame( |
| 237 | self, |
| 238 | shot_desc: ShotDescription, |
| 239 | character_portrait_path_and_text_pairs: List[Tuple[str, str]], |
| 240 | ) -> ImageOutput: |
| 241 | prompt = "" |
| 242 | reference_image_paths = [] |
| 243 | for i,(path, text )in enumerate(character_portrait_path_and_text_pairs): |
| 244 | prompt += f"Image {i}: {text}\n" |
| 245 | reference_image_paths.append(path) |
| 246 | prompt += f"Generate an image based on the following description: {shot_desc.ff_desc}." |
| 247 | image_output = await self.image_generator.generate_single_image( |
| 248 | prompt=prompt, |
| 249 | reference_image_paths=reference_image_paths, |
| 250 | size="1600x900", |
| 251 | ) |
| 252 | return image_output |
| 253 | |
| 254 | |
| 255 | |
| 256 | def _validate_camera_tree(cameras: List[Camera]) -> None: |
| 257 | """Reject parent assignments that would deadlock frame generation.""" |
| 258 | by_idx = {cam.idx: cam for cam in cameras} |
| 259 | for cam in cameras: |
| 260 | if cam.parent_cam_idx is None: |
| 261 | continue |
| 262 | if cam.parent_cam_idx == cam.idx: |
| 263 | raise ValueError(f"Camera {cam.idx} lists itself as its parent.") |
| 264 | if cam.parent_cam_idx not in by_idx: |
| 265 | raise ValueError(f"Camera {cam.idx} references unknown parent camera {cam.parent_cam_idx}.") |
| 266 | for cam in cameras: |
| 267 | seen = set() |
| 268 | current = cam |
| 269 | while current.parent_cam_idx is not None: |
| 270 | if current.idx in seen: |
| 271 | raise ValueError(f"Cycle detected in camera parent graph involving camera {current.idx}.") |
| 272 | seen.add(current.idx) |
| 273 | current = by_idx[current.parent_cam_idx] |
| 274 |