Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fast-DiM: Towards Fast Diffusion Morphs

Zander W. Blasingame, Chen Liu

TL;DR

This work tackles the high computational cost of diffusion-based face morphs by redesigning the DiM pipeline around faster numerical solvers. By replacing the PF-ODE solver with the high-order DPM++ 2M method and introducing efficient forward-time encoding through principled forward ODE formulations, Fast-DiM achieves substantial reductions in network evaluations while preserving morph quality. The encoding NFE shrinks by up to 85% and forward-sampling NFE by about 50%, with only minimal drops in MMPMR (as low as 1.6%) and 0.23% in sampling. Evaluations on the SYN-MAD 2022 dataset across ArcFace, AdaFace, and ElasticFace demonstrate that Fast-DiM maintains competitive vulnerability while improving efficiency, and detectability analyses indicate that diverse training data is essential for robust S-MAD performance. Overall, the paper provides practical guidance for deploying diffusion-based morphing with substantially reduced computational demands and offers insights into the role of forward-time solving in diffusion models.

Abstract

Diffusion Morphs (DiM) are a recent state-of-the-art method for creating high quality face morphs; however, they require a high number of network function evaluations (NFE) to create the morphs. We propose a new DiM pipeline, Fast-DiM, which can create morphs of a similar quality but with fewer NFE. We investigate the ODE solvers used to solve the Probability Flow ODE and the impact they have on the the creation of face morphs. Additionally, we employ an alternative method for encoding images into the latent space of the Diffusion model by solving the Probability Flow ODE as time runs forwards. Our experiments show that we can reduce the NFE by upwards of 85% in the encoding process while experiencing only 1.6\% reduction in Mated Morph Presentation Match Rate (MMPMR). Likewise, we showed we could cut NFE, in the sampling process, in half with only a maximal reduction of 0.23% in MMPMR.

Fast-DiM: Towards Fast Diffusion Morphs

TL;DR

This work tackles the high computational cost of diffusion-based face morphs by redesigning the DiM pipeline around faster numerical solvers. By replacing the PF-ODE solver with the high-order DPM++ 2M method and introducing efficient forward-time encoding through principled forward ODE formulations, Fast-DiM achieves substantial reductions in network evaluations while preserving morph quality. The encoding NFE shrinks by up to 85% and forward-sampling NFE by about 50%, with only minimal drops in MMPMR (as low as 1.6%) and 0.23% in sampling. Evaluations on the SYN-MAD 2022 dataset across ArcFace, AdaFace, and ElasticFace demonstrate that Fast-DiM maintains competitive vulnerability while improving efficiency, and detectability analyses indicate that diverse training data is essential for robust S-MAD performance. Overall, the paper provides practical guidance for deploying diffusion-based morphing with substantially reduced computational demands and offers insights into the role of forward-time solving in diffusion models.

Abstract

Diffusion Morphs (DiM) are a recent state-of-the-art method for creating high quality face morphs; however, they require a high number of network function evaluations (NFE) to create the morphs. We propose a new DiM pipeline, Fast-DiM, which can create morphs of a similar quality but with fewer NFE. We investigate the ODE solvers used to solve the Probability Flow ODE and the impact they have on the the creation of face morphs. Additionally, we employ an alternative method for encoding images into the latent space of the Diffusion model by solving the Probability Flow ODE as time runs forwards. Our experiments show that we can reduce the NFE by upwards of 85% in the encoding process while experiencing only 1.6\% reduction in Mated Morph Presentation Match Rate (MMPMR). Likewise, we showed we could cut NFE, in the sampling process, in half with only a maximal reduction of 0.23% in MMPMR.
Paper Structure (14 sections, 13 equations, 5 figures, 8 tables, 1 algorithm)

This paper contains 14 sections, 13 equations, 5 figures, 8 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Prior Work
  3. Experimental Setup
  4. Dataset
  5. Face Recognition Systems
  6. Metrics
  7. Fast-DiM
  8. The ODE Solver
  9. Noise Injection
  10. Solving the Forward ODE
  11. Results
  12. Vulnerability
  13. Detectability Study
  14. Conclusion

Figures (5)

  • Figure 1: Example of a morph created using DiM blasingame2023leveraging. Samples are from the FRLL dataset frll.
  • Figure 2: From left to right: identity $a$, morph generated with DDIM $(N = 100)$, morph generated with DPM++ 2M $(N = 20)$, identity $b$.
  • Figure 3: From top left to lower right: original image, output from the DiffAE forward solver, white noise, original image, DDIM sampled image from DiffAE approach, DDIM sampled image from pure white noise.
  • Figure 4: From left to right: identity $a$, identity $b$, pixel-wise averaged image, noisy image, final morphed image.
  • Figure 5: From left to right: identity $a$, DiffAE forward solver $N_F = 250$, DDIM forward ODE solver $N_F=100$, DPM++ 2M forward ODE solver $N_F=100$, DPM++ 2M forward ODE solver $N_F = 50$, and identity $b$.