ALADDIN: Adversarial Latent Diffusion Distillation for One-Step Identity Swapping
Authors & Institutions
Ivan Shchekotov
Constructor University
Timofei Gritsaev
Constructor University
Viacheslav Pirogov
Sumsub
Dmitry Vetrov
Constructor University
What Problem It Solves
ALADDIN seeks one-step identity swapping with diffusion-level realism and a controllable balance between source identity transfer and preservation of target pose, expression, hair, lighting, accessories, and background. It also introduces an identity evaluation protocol based on an encoder independent of every compared method.
Key Result
On 1,000 FFHQ pairs, the balanced student reports FID 4.49, independent-encoder identity similarity 0.50, and 97.4% top-1 identity retrieval. A preservation-focused variant reaches FID 3.34 and matches APPLE's 0.40 identity score while improving pose error from 1.86 to 1.56. It uses one network evaluation instead of APPLE's 28 and about 0.115 million training pairs rather than 1.28 million.
Abstract
Face swapping transfers a source identity onto a target image while preserving the target's pose, expression, hair, and lighting. The most realistic current methods are diffusion-based, but they need many network evaluations per image. GAN-based methods are fast, yet trained from scratch they must learn realism from the adversarial signal alone and produce visibly weaker results. We present ALADDIN (Adversarial Latent Diffusion Distillation for One-Step Identity Swapping), a one-step face swapper obtained by initialising the generator from a pre-trained multi-step flow-matching model and training it adversarially under identity and target-preservation constraints. We expose a controllable trade-off between identity transfer and target-attribute preservation by training a student model with triplet pseudo-label supervision. We additionally identify a bias in the standard evaluation protocol, which favours methods trained against the same identity encoder used to score them, and re-evaluate all baselines with an independent one. ALADDIN is state-of-the-art among one-step methods in realism and target preservation at comparable identity transfer, and matches the best multi-step model with 28 times fewer network evaluations and an order of magnitude fewer training pairs.
Research Starting Point
Diffusion face swappers can be realistic but require many network evaluations, while one-step GAN systems are fast and often less faithful. Evaluation can also be biased when a method is trained against the same face encoder later used to score its identity similarity.
Method
The generator is initialized from a pretrained multi-step flow-matching model and trained adversarially. Its discriminator reads intermediate features from a frozen diffusion model rather than raw latents, while identity and target-preservation losses constrain the output. A second student uses triplet pseudo-label supervision to expose an adjustable reconstruction weight and therefore a useful identity-versus-attribute trade-off.
Paper Summary
ALADDIN materially lowers the inference and training cost of high-quality still-image swapping. Its independent identity evaluation is equally important for procurement: a favorable score can be partly caused by sharing the evaluator used during training.