← Back to Blog
Research RadarFace SwappingTalking FacearXivAugust 2026

Monthly arXiv Radar

August 2026 Face Swapping Papers: Real-Time Portraits, Minute-Long Avatars, and Emotional Control

August's face swapping and digital-human research is less about a single identity-transfer trick and more about sustained, controllable interaction. The selected systems separate reusable motion from appearance, distill an offline audio-video model into a minute-scale streamer, and combine latent emotion semantics with explicit blendshape geometry.

What This Month Signals

Proxy Avatar Meets Low-Rank Caching reaches 32 FPS on an RTX 4090 by separating a reusable emotional performer from one-shot identity retargeting and caching static appearance. Omni-LiveAvatar tackles the data-center end of the spectrum, using autoregressive distillation and bounded long-short memory to keep joint audio and video coherent for a full minute. GemTalk focuses on the control interface, letting explicit blendshape geometry scale the strength of implicit emotion features. Together they map three different product choices: personalized motion reuse, continuous streaming infrastructure, and fine-grained expressive direction.

Paper 012026-08-03cs.CV

Proxy Avatar Meets Low-Rank Caching: Real-Time One-Shot Emotion-Controllable Portrait Animation

Authors & Institutions

Haijie Yang

Nanjing University of Science and Technology, China

Jindi Bao

Nanjing University of Science and Technology, China

Yixuan Dong

Tsientang Institute for Advanced Study, China

Hongliang Zhang

Nanjing University of Science and Technology, China

Jian Bi

Nanjing University of Science and Technology, China

Hao Tang

Peking University, China

Zhenyu Zhang

Nanjing University, China

Jianjun Qian

Nanjing University of Science and Technology, China

Jian Yang

Nanjing University of Science and Technology, China

What Problem It Solves

The paper separates reusable emotional motion, cross-identity retargeting, and appearance rendering so that a single reference portrait can be animated in real time from audio and an emotion label. It specifically removes repeated reference-feature computation without retraining a cache for each target identity.

Key Result

On an RTX 4090 the full system runs at 32 FPS. In self-reconstruction it reaches 31.24 PSNR, 0.018 LPIPS, 0.883 SSIM, and a 6.314 lip-sync score, leading most quality and motion metrics despite EmoTag's higher raw frame rate. In cross-identity driving it records 29.64 FID, 0.745 identity similarity, 75.4% emotion accuracy, and 6.096 lip sync. Participants preferred it for emotional expressiveness in 70% of comparisons, lip synchronization in 60%, and visual realism in 65%; extended clips showed no obvious identity or sync drift.

Abstract

Audio-driven portrait animation has advanced rapidly with diffusion-based generative models, yet real-time one-shot generation with expressive emotion control remains challenging. Existing methods often suffer from insufficient emotion-aware motion priors and expensive appearance computation during multi-step denoising. To address these issues, we propose Proxy Avatar Meets Low-Rank Caching, a cascaded framework for real-time one-shot emotion-controllable portrait animation. Instead of directly generating the target portrait from audio, our method uses a Gaussian-based emotion proxy avatar as a reusable motion generator, which is trained once on a single identity to produce expressive driving videos from audio and emotion labels. Since the proxy avatar only provides motion rather than target appearance or geometry, a large-scale one-shot retargeting model further extracts identity-independent motion from the proxy performance and adapts it to arbitrary target portraits. To improve inference efficiency, we introduce zero-shot appearance reuse with low-rank caching, which caches reference appearance features at the initial denoising step and models subsequent feature variations using lightweight low-rank adapters. Extensive experiments demonstrate that our method achieves stronger emotional expressiveness, better identity-preserving animation, and substantially reduced inference cost, enabling real-time one-shot portrait animation.

Research Starting Point

One-shot talking portraits need to preserve an unseen person's appearance, follow speech, express a requested emotion, and run interactively on available hardware. Diffusion systems repeatedly recompute nearly static appearance features during denoising, while emotion controls often entangle identity and motion or require a reference video for every style. The result is either expressive but slow generation or fast animation with weak emotional range.

Method

A Gaussian-based proxy avatar is trained once on one performer and serves only as an audio-and-emotion-conditioned motion generator. A large-scale one-shot retargeting network extracts identity-independent motion from that proxy video and applies it to arbitrary target portraits. During few-step diffusion, zero-shot appearance reuse caches reference features at the first step, while lightweight low-rank adapters model only their later variation. This cascaded division of labor is evaluated in self-reconstruction, cross-identity one-shot driving, long-video stability, ablations, and a blinded study with twenty participants.

Paper Summary

The reusable proxy-plus-retargeting design makes real-time emotional avatars practical without per-user diffusion training. The reported 32 FPS on a consumer RTX 4090 is promising, but deployment sizing should include the one-time proxy training and the quality/cost trade-off of the large retargeting model, not just steady-state frame rate.

Paper 022026-08-07cs.MM

Omni-LiveAvatar: Minute-Level Real-Time Streaming Joint Audio-Video Avatar Generation

Authors & Institutions

Lunjie Zhu

iComAI Lab, Hong Kong University of Science and Technology, Hong Kong

Vivix Group Limited, Hong Kong

Xingtong Ge

iComAI Lab, Hong Kong University of Science and Technology, Hong Kong

Vivix Group Limited, Hong Kong

Fangyu Lin

iComAI Lab, Hong Kong University of Science and Technology, Hong Kong

Vivix Group Limited, Hong Kong

Yi Zhang

Vivix Group Limited, Hong Kong

Zhening Liu

iComAI Lab, Hong Kong University of Science and Technology, Hong Kong

Mengfei Li

iComAI Lab, Hong Kong University of Science and Technology, Hong Kong

Vivix Group Limited, Hong Kong

Yumeng Zhang

iComAI Lab, Hong Kong University of Science and Technology, Hong Kong

Guanglu Song

Vivix Group Limited, Hong Kong

Yu Liu

Vivix Group Limited, Hong Kong

Jun Zhang

iComAI Lab, Hong Kong University of Science and Technology, Hong Kong

What Problem It Solves

Omni-LiveAvatar converts a large offline joint audio-video teacher into a causal streaming generator and addresses the separate stability problems of long context and changing prompts. Its target is continuous minute-scale generation rather than a sequence of independently generated short clips.

Key Result

For five-second clips the model generates at 19.57 FPS, roughly 33 times faster than its 0.60-FPS teacher, while posting the highest overall quality score among tested systems (81.72) and perfect human identity and clothing fidelity scores. On sixty-second generation it reaches 21.99 FPS versus 16.18 for the next-fastest baseline. Its minute-level human identity score is 98.61 versus 67.60 and 50.19, Sync-C is 6.76 versus 0.72 and 0.28, and text-video alignment is 9.82 versus 6.68 and 5.46, with visibly reduced avatar and background drift.

Abstract

Joint audio-video generative models serve as foundation for immersive and interactive digital-human generation. Nevertheless, most existing models rely on bidirectional attention and multi-step denoising and can generate only short clips, making them unsuitable for real-time interaction over extended durations. We present Omni-LiveAvatar, the first framework for minute-level, real-time streaming joint audio-video avatar generation. Specifically, we propose (1) a progressive autoregressive distillation pipeline that transfers a large bidirectional joint audio-video diffusion model into a few-step autoregressive generator without auxiliary stabilization mechanisms; (2) a synchronized audio-video long-short-term memory that preserves global consistency under a bounded memory budget; and (3) a hierarchical rolling prompt planning strategy that enables coherent semantic evolution and seamless prompt transitions. Extensive experiments show that Omni-LiveAvatar generates high-quality, synchronized minute-level avatars in real time. In terms of speed, it achieves a 33$\times$ generation speedup over its teacher, LTX-2, on a single NVIDIA H200 GPU; in terms of generation quality, it outperforms accelerated baselines across visual quality, audio quality, cross-modal synchronization, and human fidelity. Our code is available at https://github.com/Aoko955/Omni-LiveAvatar.

Research Starting Point

Joint audio-video generators can synthesize convincing short clips, but bidirectional attention and multi-step diffusion prevent live interaction, while accelerated autoregressive variants drift in identity, background, audio quality, or lip synchronization over long sessions. A useful digital human must preserve both modalities for minutes under bounded memory and accept evolving prompts without abrupt semantic or acoustic transitions.

Method

A three-stage progressive autoregressive distillation pipeline transfers LTX-2 into a four-denoising-step causal generator without external reward models or modality-specific stabilization. Synchronized audio-video long-short-term memory combines a periodically re-anchored first macro-block with rolling key-value caches, retaining global identity and recent context inside a fixed budget. Hierarchical rolling prompt planning splits persistent appearance/background instructions from block-level local actions so prompts advance with the rolling window. Evaluation covers both five-second and sixty-second output at 512 by 768 on one NVIDIA H200.

Paper Summary

The paper makes long-horizon audio-video stability a first-class systems problem rather than an afterthought. It is a strong reference architecture for live digital humans, but the real-time claim is measured on an H200 at 512 by 768, so buyers should reproduce latency, memory, and concurrency on their own serving hardware.

Paper 032026-08-01cs.CV

Geometry-guided Emotion Modulation for Controllable and Photorealistic Emotional Talking Face Generation

Authors & Institutions

Chenggong Hu

School of Software Technology, Zhejiang University, Ningbo, China

Shaoyin Ma

School of Software Technology, Zhejiang University, Ningbo, China

Yi Wang

College of Computer Science and Technology, Zhejiang University, Hangzhou, China

Li Sun

Ningbo Global Innovation Center, Zhejiang University, Ningbo, China

Mingli Song

College of Computer Science and Technology, Zhejiang University, Hangzhou, China

Jie Song

School of Software Technology, Zhejiang University, Ningbo, China

What Problem It Solves

GemTalk connects audio-derived emotional semantics to identity-aware facial geometry and uses the geometry to control the strength of latent expression features. This makes emotion category and intensity continuously editable while retaining photorealism and lip synchronization.

Key Result

On the combined MEAD-front and RAVDESS-speech emotional test, GemTalk reaches 334.937 FVD, 31.625 FID, 7.027 Sync-C, 8.283 Sync-D, 2.392 expression FID, and 59.258% emotion accuracy. It improves emotion accuracy by 3.33 percentage points over the next-best method and lowers FVD by 20.84. On neutral HDTF it retains strong synchronization and the best reported expression FID of 1.386, while coefficient edits produce smooth weak-to-strong emotion transitions without obvious identity loss.

Abstract

Audio-driven emotional talking face generation aims to synthesize realistic videos with expressive facial dynamics. However, existing methods struggle to balance controllability and visual fidelity. Although implicit representations capture rich semantics, they lack structural guidance, often resulting in averaged emotional expressions. In contrast, explicit geometric methods offer better control over facial expressions but tend to sacrifice high-frequency texture details. To address it, we propose GemTalk, a diffusion-based framework that combines the semantic richness of implicit representations with the structural precision of explicit geometric priors. We introduce a Vision-guided Audio Emotion Projection (V-AEP) module to extract implicit emotional lip and expression features. At the same time, a Diffusion-based Geometric Priors Generator (D-GPG) generates identity-aware blendshape coefficients as explicit structural priors. Crucially, our Geometry-guided Emotion Modulation (GEM) module leverages these geometric priors to recalibrate the magnitude of implicit features, enabling precise, continuous control over emotional expressions, especially emotion intensity, without sacrificing visual quality. Extensive experiments show GemTalk achieves superior performance in photo-realism, and facial emotional dynamics.

Research Starting Point

Implicit diffusion features provide rich facial texture but tend to average emotional expressions and offer poor intensity control. Explicit landmarks or blendshapes provide interpretable structure, yet systems driven only by geometry often lose high-frequency appearance and can produce a mouth or upper face that conflicts with the audio. Emotional animation needs both semantic richness and controllable physical motion without leaking the expression already present in the reference portrait.

Method

Training begins with an emotion-agnostic lip-sync backbone on broad non-emotional video. Vision-guided Audio Emotion Projection learns separate but coordinated lip and upper-face emotion spaces by using visual emotion representations to supervise audio features. A diffusion-based geometric-prior generator predicts identity-aware Apple ARKit blendshape coefficients. Geometry-guided Emotion Modulation then preserves the latent feature direction that represents emotion category while recalibrating its magnitude from the blendshapes; conflict-aware sampling pairs reference images and audio with different emotions so the model cannot copy the source expression.

Paper Summary

Combining latent emotion semantics with explicit blendshape magnitude offers a useful control surface for creative tools and digital-human APIs. GemTalk is strongest when continuous emotional intensity matters, though teams should test whether its emotion labels and ARKit coefficient assumptions cover their languages, speaking styles, and target rigs.