1 of 15

Diffuse-CLoC:

Guided Diffusion for Physics-based

Character Look-ahead Control

2025.  07.  18.

SEOUL NATIONAL UNIVERSITY

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING

(Bldg.302 #312-1)

WONJEONG SEO

Review | SIGGRAPH ’25 (TOG)

2 of 15

2

0. LINK

https://diffusecloc.github.io/website/

3 of 15

3

1. GOAL

Physically Steerable Character Control for Zero-Shot Tasks

Train Diffusion Model

Condition

Guidance

Unseen Tasks

w/o retraining, high-level controller

4 of 15

4

2. RELATED WORKS

Physics-based Control using Diffusion Model

steerability↓

fine-tuning (per task)

depends on kinematic trajectory (tracking)

5 of 15

5

2. RELATED WORKS

Physics-based Control using Diffusion Model

steerability↓

fine-tuning (per task)

depends on kinematic trajectory (tracking)

6 of 15

6

3. KEY CONCEPTS

  1. Joint Distribution Transformer Diffusion Model
  2. Non-causal/causal Attention for States and Actions
  3. Rolling Inference Scheme

Perform unseen task w/o high-level controller or retraining NN

7 of 15

7

4. FRAMEWORK

Joint Distribution Transformer Diffusion Model

Training Step]

8 of 15

8

4. FRAMEWORK

Joint Distribution Transformer Diffusion Model

Inference Step]

Backward process

9 of 15

9

4. FRAMEWORK

Cost Functions

Static obstacle avoidance

Dynamic obstacle avoidance

10 of 15

10

4. FRAMEWORK

Attention for States and Actions

state -> state (prevent compromising kinematic information)

action -> causal state/action pair (preserve consistency)

11 of 15

11

4. FRAMEWORK

Rolling Inference Scheme

12 of 15

12

5. RESULTS

13 of 15

13

5. RESULTS

14 of 15

14

5. RESULTS

15 of 15

15

5. RESULTS