3DTrajMaster: Mastering 3D Trajectory for Multi-Entity Motion in Video Generation

         

Xiao Fu^{1 †}, Xian Liu¹, Xintao Wang^{2 ✉}, Sida Peng³, Menghan Xia², Xiaoyu Shi², Ziyang Yuan²,
Pengfei Wan² Di Zhang², Dahua Lin^1✉
1The Chinese University of Hong Kong ²Kuaishou Technology ³Zhejiang University
†: Intern at KwaiVGI, Kuaishou Technology, ✉: Corresponding Authors

ICLR 2025

🌟 Introduction

🔥 3DTrajMaster controls one or multiple entity motions in 3D space with entity-specific 3D trajectories for text-to-video (T2V) generation. It has the following features:

• 6 Domain of Freedom (DoF): control 3D entity location and orientation.
• Diverse Entities: human, animal, robot, car, even abstract fire, breeze, etc.
• Diverse Background: city, forest, desert, gym, sunset beach, glacier, hall, night city, etc.
• Complex 3D trajectories: 3D occlusion, rotating in place, 180°/continuous 90° turnings, etc.
• Fine-grained Entity Prompt: change human hair, clothing, gender, figure size, accessory, etc.

teaser.mp4

🔥 Release News

• [2025/01/23] 3DTrajMaster is accepted to ICLR 2025.
• [2025/01/22] Release inference and training codes based on CogVideoX-5B.
• [2024/12/10] Release paper, project page, dataset, and eval code.

⚙️ Quick Start

(1) Access to Our Internal Video Model

As per company policy, we may not release the proprietary trained model at this time. However, if you wish to access our internal model, please submit your request via (1) a shared document or (2) directly via email ([email protected]); we will respond to requests with the generated video as quickly as possible. Please ensure your request includes the following:

1. Entity prompts (1–3, with a maximum of 42 tokens, approximately 20 words per entity)
2. Location prompt
3. Trajectory template (you can refer to the trajectory template in our released 360°-Motion Dataset, or simply describe new ones via text)

(2) Access to Publicly Available Codebase

We open-source a model based on CogVideoX-5B. Below is a comparison between CogVideoX and our internal video model as of 2025.01.15.

CogVideoX-2025.01.15.mp4

Inference

1. [Environment Set Up] Our environment setup is identical to CogVideoX. You can refer to their configuration to complete the environment setup.

   conda create -n 3dtrajmaster python=3.10
   conda activate 3dtrajmaster
   pip install -r requirements.txt

2. [Download Weights and Dataset] Download the pretrained checkpoints (CogVideo-5B, LoRA, and injector) from here and place them in the CogVideo/weights directory. Then, download the dataset from here. Please note that in both training stages, we use only 11 camera poses and exclude the last camera pose as the novel pose setting.

3. [Inference on Generalizable Prompts] Change root path to CogVideo/inference. Note a higher LoRA scale and more annealed steps can improve accuracy in prompt generation but may result in lower visual quality. For entity input, you can use GPT to enhance the description to an appropriate length, such as "Generate a detailed description of approximately 20 words".

   python 3dtrajmaster_inference.py \
       --model_path ../weights/cogvideox-5b \
       --ckpt_path ../weights/injector \
       --lora_path ../weights/lora \
       --lora_scale 0.6 \
       --annealed_sample_step 20 \
       --seed 24 \
       --output_path output_example

   Argument	Description
   --lora_scale	LoRA alpha weight. Options: 0-1, float. Default: 0.6.
   --annealed_sample_step	annealed sampling steps during inference. Options: 0-50, int. Default: 20.
   Generalizable Robustness	prompt entity number: 1>2>3
   Entity Length	15-24 words, ~24-40 tokens after T5 embeddings

   The following code snapshot showcases the core components of 3DTrajMaster, namely the plug-and-play 3D-motion grounded object injector.

   # 1. norm & modulate
   norm_hidden_states, norm_empty_encoder_hidden_states, gate_msa, enc_gate_msa = self.norm1(hidden_states, empty_encoder_hidden_states, temb)
   bz, N_visual, dim = norm_hidden_states.shape
   max_entity_num = 3
   _, entity_num, num_frames, _ = pose_embeds.shape
   
   # 2. pair-wise fusion of trajectory and entity
   attn_input = self.attn_null_feature.repeat(bz, max_entity_num, 50, num_frames, 1)
   pose_embeds = self.pose_fuse_layer(pose_embeds)
   attn_input[:,:entity_num,:,:,:] = pose_embeds.unsqueeze(-3) + prompt_entities_embeds.unsqueeze(-2)
   attn_input = torch.cat((
       rearrange(norm_hidden_states, "b (n t) d -> b n t d",n=num_frames), 
       rearrange(attn_input, "b n t f d -> b f (n t) d")),
       dim=2
   ).flatten(1,2)
   
   # 3. gated self-attention
   attn_hidden_states, attn_encoder_hidden_states = self.attn1_injector(
       hidden_states=attn_input,
       encoder_hidden_states=norm_empty_encoder_hidden_states,
       image_rotary_emb=image_rotary_emb,
   )
   attn_hidden_states = attn_hidden_states[:,:N_visual,:]
   
   hidden_states = hidden_states + gate_msa * attn_hidden_states

Training

1. Change root path to CogVideo/finetune. First, train lora module to fit the synthetic data domain.

   bash finetune_single_rank_lora.sh

2. Then, train injector module to learn the entity motion controller. Here we set --block_interval to 2 to insert the injector every 2 transformer blocks. You can increase this value for a lighter model, but note that it will require a longer training time. For the initial fine-tuning stage, use --finetune_init. If resuming from a pre-trained checkpoint, omit --finetune_init and specify --resume_from_checkpoint $TRANSFORMER_PATH instead. Note that in both training stages, we use only 11 camera poses and exclude the last camera pose as the novel pose setting.

   bash finetune_single_rank_injector.sh

📦 360°-Motion Dataset (Download 🤗)

 ├── 360Motion-Dataset                      Video Number        Cam-Obj Distance (m)
   ├── 480_720/384_672
       ├── Desert (desert)                    18,000               [3.06, 13.39]
           ├── location_data.json
       ├── HDRI                                                      
           ├── loc1 (snowy street)             3,600               [3.43, 13.02]
           ├── loc2 (park)                     3,600               [4.16, 12.22]
           ├── loc3 (indoor open space)        3,600               [3.62, 12.79]
           ├── loc11 (gymnastics room)         3,600               [4.06, 12.32]
           ├── loc13 (autumn forest)           3,600               [4.49, 11.92]
           ├── location_data.json
       ├── RefPic
       ├── CharacterInfo.json
       ├── Hemi12_transforms.json

(1) Released Dataset Information (V1.0.0)

Argument	Description	Argument	Description
Video Resolution	(1) 480×720 (2) 384×672	Frames/Duration/FPS	99/3.3s/30
UE Scenes	6 (1 desert+5 HDRIs)	Video Samples	(1) 36,000 (2) 36,000
Camera Intrinsics (fx,fy)	(1) 1060.606 (2) 989.899	Sensor Width/Height (mm)	(1) 23.76/15.84 (2) 23.76/13.365
Hemi12_transforms.json	12 surrounding cameras	CharacterInfo.json	entity prompts
RefPic	50 animals	1/2/3 Trajectory Templates	36/60/35 (121 in total)
{D/N}_{locX}	{Day/Night}_{LocationX}	{C}_ {XX}_{35mm}	{Close-Up Shot}_{Cam. Index(1-12)} _{Focal Length}

Note that the resolution of 384×672 refers to our internal video diffusion resolution. In fact, we render the video at a resolution of 378×672 (aspect ratio 9:16), with a 3-pixel black border added to both the top and bottom.

(2) Difference with the Dataset to Train on Our Internal Video Diffusion Model

The release of the full dataset regarding more entities and UE scenes is still under our internal license check.

Argument	Released Dataset	Our Internal Dataset
Video Resolution	(1) 480×720 (2) 384×672	384×672
Entities	50 (all animals)	70 (20 humans+50 animals)
Video Samples	(1) 36,000 (2) 36,000	54,000
Scenes	6	9 (+city, forest, asian town)
Trajectory Templates	121	96

(3) Load Dataset Sample

1. Change root path to dataset. We provide a script to load our dataset (video & entity & pose sequence) as follows. It will generate the sampled video for visualization in the same folder path.

   python load_dataset.py

2. Visualize the 6DoF pose sequence via Open3D as follows:

   python vis_trajecotry.py

   After running the visualization script, you will get an interactive window like this. Note that we have converted the right-handed coordinate system (Open3D) to the left-handed coordinate system in order to better align with the motion trajectory of the video:

   

🚀 Benchmark Evaluation (Reproduce Paper Results)

├── eval
  ├── GVHMR
  ├── common_metrics_on_video_quality

(1) Evaluation on 3D Trajectory

1. Change root path to eval/GVHMR. Then follow GVHMR installation to prepare the setups and (recommend using a different Conda environment to avoid package conflicts). Our evaluation input is available at here. Please note that the 3D trajectories have been downsampled from 77 frames to 20 frames to match the RGB latent space of the 3D VAE.

2. Download the inference videos generated by our internal video diffusion model and corresponding evalution GT poses by using this command (you can check the 3D evaluated trajectory via our provided visualization script):

   bash download_eval_pose.sh

3. Estimation of human poses on evaluation sets:

   python tools/demo/demo_folder.py -f eval_sets -d outputs/eval_sets_gvhmr -s

4. Evaluation of all human samples (note to convert the left and right hand coordinate systems) :

   python tools/eval_pose.py -f outputs/eval_sets_gvhmr

(2) Evaluation on Visual Quality

1. Change root path to eval/common_metrics_on_video_quality. Then download fvd, inference videos and base T2V inference videos using the download script:

   bash download_eval_visual.sh

2. Evaluation of FVD, FID, and CLIP-SIM metrics.

   pip install pytorch-fid  clip
   bash eval_visual.sh

📚 Related Work

• MotionCtrl: the first to control 3D camera motion and 2D object motion in video generation
• TC4D: compositional text-to-4D scene generation with 3D trajectory conditions
• Tora: control 2D motions in trajectory-oriented diffusion transformer for video generation
• SynCamMaster: multi-camera synchronized video generation from diverse viewpoints
• StyleMaster: enable artistic video generation and translation with reference style image

🔗 Citation

If you find this work helpful, please consider citing:

@inproceedings{fu20243dtrajmaster,
  title={3DTrajMaster: Mastering 3D Trajectory for Multi-Entity Motion in Video Generation},
  author={Fu, Xiao and Liu, Xian and Wang, Xintao and Peng, Sida and Xia, Menghan and Shi, Xiaoyu and Yuan, Ziyang and Wan, Pengfei and Zhang, Di and Lin, Dahua},
  booktitle={ICLR},
  year={2025}
}