From Diffusion To Flow: Efficient Motion Generation In MotionGPT3

This repository contains official code for the paper From Diffusion To Flow: Efficient Motion Generation In MotionGPT3 (ReALM-GEN Workshop ICLR 2026) by Jaymin Bhan, JiHong Jeon and SangYeop Jeong.

This work was completed as our two-semester capstone project in the Applied AI department at Seoul National University of Science and Technology, where it received the department's grand prize (1st place). The accompanying Motion-Chat demo app is built on the improved motion-generation model from this project.

Motion-Chat App

Motion-Chat: an interactive interface that responds to a user prompt with both a text reply and a generated 3D motion animation, built on our improved rectified-flow motion model. It explores motion as a complementary response modality for chat-based systems, where the model can show a movement rather than just describe it.

Paper Summary

Recently, flow-based approaches have shown competitive results in generative modeling for image and audio generation compared to diffusion methods. In this work, we conduct an isolated comparison to see if the known advantages of flow methodology extend to the motion generation domain.

We found that by replacing MotionGPT3's diffusion head with a rectified flow approach, we were able to further improve evaluation metrics while reducing both the required training epochs and the number of inference steps for motion generation.

Poster (ICLR 2026 ReALM-GEN Workshop)

Result Table

Methods	R@1↑	R@2↑	R@3↑	FID↓	MMDist↓	Diversity↑	MModality↑
Real	0.511	0.703	0.797	0.002	2.974	9.503	–
T2M-GPT (Zhang et al., 2023)	0.491	0.680	0.775	0.116	3.118	9.761	1.856
DiverseMotion (Lou et al., 2023)	0.515	0.706	0.802	0.072	2.941	9.683	1.869
MoMask (Guo et al., 2023)	0.521	0.713	0.807	0.045	2.958	9.620	1.241
MotionGPT (Jiang et al., 2023)	0.492	0.681	0.733	0.232	3.096	9.528	2.00
TM2T (Guo et al., 2022b)	0.424	0.618	0.729	1.501	3.467	8.589	2.424
MotionGPT3 (Reported)	0.533	0.731	0.826	0.239	2.797	9.688	1.560
MotionGPT3 (Diffusion, ep. 142)	0.520	0.716	0.807	0.240	2.918	9.479	2.277
MotionGPT3 (Flow, ep. 54)	0.544	0.740	0.828	0.192	2.837	9.340	2.366

Comparison of text-to-motion generation methods on HumanML3D. Arrows indicate whether higher (↑) or lower (↓) values are better. Real denotes statistics computed from groundtruth motion data. Bold values indicate the best performance for each metric.

Training Curves

Inference Step Comparison

Setup and Download

Setup and download steps follow the MotionGPT3 repository with minor dependency updates to match a single GPU (RTX 5090) environment.

Clone this repository.

Install uv, then:

uv venv .venv --python 3.11 && source .venv/bin/activate
uv pip install -r requirements.txt
python -m spacy download en_core_web_sm

Run the script to download dependencies materials:

bash prepare/download_smpl_model.sh
bash prepare/prepare_gpt2.sh

For Text to Motion Evaluation

bash prepare/download_t2m_evaluators.sh

For pre-trained MotionVAE:

bash prepare/download_mld_pretrained_models.sh

Then run following script to process checkpoints:

python -m scripts.gen_mot_gpt

3. Pre-trained model

Run the script to download the pre-trained model

bash prepare/download_pretrained_motiongpt3_model.sh

4. (Optional) Download manually

Visit the Hugging Face to download the pretrained models.

Training

1. Set up the text-to-motion dataset following the HumanML3D instructions.

2. Copy the instruction data from prepare/instructions into the HumanML3D dataset folder.

3. Update the dataset and output paths in the following config files:

   • configs/MoT_vae_stage3.yaml
   • configs/MoT_vae_stage2_instruct.yaml
   • configs/MoT_vae_stage2_all-flow.yaml
   • configs/MoT_vae_stage1_t2m-flow.yaml
   • configs/MoT_vae_stage2_all-diffusion.yaml
   • configs/MoT_vae_stage1_t2m-diffusion.yaml
   • configs/assets.yaml

   And the training scripts:

   • run_stage1_training-diffusion.sh
   • run_stage1_training-flow.sh

4. Run training:

# Diffusion baseline
./run_stage1_training-diffusion.sh

# Rectified flow (ours)
./run_stage1_training-flow.sh