Bibliography
A curated bibliography of peer-reviewed research, preprints and industry sources on digital humans, avatars, embodied agents and AI-led training — each with a plain-English summary and a link to the original.
Industry & market entries are analyst reports, vendor publications and marketing material. They are included for market context and are not peer-reviewed. Trace any specific figure to its original source before citing it.
arXiv preprint, 2026 · Oct 15, 2026
Shukesh Reddy, Abhijit Das
In plain English
Introduces a dataset of over 2 million face-swap images covering manipulations of individual facial regions and tests 12 detectors on it. The authors report that detectors lose roughly 10% in accuracy when the swap affects only part of the face.
arXiv preprint, 2026 · Oct 8, 2026
Fazhong Liu, Yan Meng, Tian Dong, Guoxing Chen, Haojin Zhu
In plain English
Builds a benchmark of real and synthetic full-body videos animated by pose-guided diffusion models and proposes MoDA, a detector that also identifies which generator made a clip. The authors report detection accuracy of 94.8% in-distribution and 89.1% across datasets.
arXiv preprint, 2026 · Aug 12, 2026
Ruibin Li, Zhiyuan Ma, Fangzhou Ai, Lei Zhang
In plain English
Proposes a framework for streaming audio-driven avatar video of unlimited length by training a fast generator and a stability adapter in parallel, plus a caching method to cut computation. The authors report about 27 frames per second at 768x512 on a single H100 GPU.
Global Market Insights, 2026 · Aug 1, 2026
Preeti Wadhwani, Aishwarya Ambekar
In plain English
Global Market Insights reports the AI avatars market at USD 6.3 billion in 2025 and USD 8.4 billion in 2026. It forecasts growth to USD 93.4 billion by 2035, a 30.6% CAGR.
JMIR AI, 2026 · Jul 9, 2026
Dafna Benadof, Josefa Cabezas, Daniel Schwartz
In plain English
Two online randomized experiments found humanized AI avatars were as effective as a real person at raising intention to follow dental aftercare instructions, while animated avatars lowered it. Disclosing the video was AI-generated did not change intention to comply but reduced willingness to return to the same dentist.
Behavior Research Methods (2026)
Matus Krajcovic, Peter Demcak, Eduard Kuric
In plain English
Eighty UK participants completed psychometric surveys either by speaking to a photorealistic video avatar or by typing with a chatbot, producing 2,265 conversational responses. The avatar condition drew significantly more informative and detailed answers (mean informativeness 285.41 versus 142.23, p < .001) and higher, more time-efficient engagement; satisfaction did not differ between the two. Participants attributed the lack of satisfaction difference to personal preference, turn-taking delay and uncanny valley reactions, which is a useful reminder that richer data and a better felt experience are not the same win.
arXiv
Kyra Wang, Boon-Kiat Quek, Jessica Goh, Dorien Herremans
In plain English
A within-subjects study compared embodied and non-embodied LLM conversational agents on cooperative tasks, and the embodied version came out worse: the text-only agent was rated significantly more competent, and many participants described the embodied agent as more sycophantic. The authors argue the usual credibility benefit of embodiment can reverse when the underlying model is sycophantic. Worth reading before assuming a face will make an agent more trusted; it may instead make its agreeableness more obvious.
arXiv · Oct 22, 2026
Zhuoyuan Jiang, Weijie Lyu, Xiaoyuan Wang, Tiancheng Shen, Ming-Hsuan Yang
In plain English
Most 3D talking avatar systems are built on parametric face models such as 3DMM or FLAME, which struggle with extreme expressions and anything worn on the head. This method skips the face model entirely: it first generates synchronised multi-view video from a single portrait and audio, then reconstructs a full-head Gaussian avatar from those views, tolerating the inconsistencies between them. For practitioners, the interesting claim is accessories and wide expression range from one photo with no per-person optimisation step.
arXiv · Oct 15, 2026
Junseok Ahn, Jaehyun Park, Jeongsoo Choi, Joon Son Chung, Ji-Hoon Kim
In plain English
Conversational agents usually generate speech and face separately, or skip the face altogether. This system runs both from one shared LLM through a joint audio-visual representation that then splits into streaming speech and face branches, removing the intermediate hand-off. The authors report initial latency roughly 200 times lower than a strong baseline, though no absolute millisecond figures, plus better lip sync and more appropriate expressions. They also release JAVIS-Syn, a 334-hour synthetic dataset of expressive multi-turn conversations, which addresses a real shortage of training data for audio-visual dialogue.
arXiv · Sep 30, 2026
Junyao Gao, Sibo Liu, Weidong Zhang, Cairong Zhao, Jun Zhang
In plain English
Audio-plus-reference-image avatar generation gives you a talking face but almost no control over the body. This framework adds SMPL-X 3D guidance on top of a video model so body pose and head motion can be directed explicitly, with separate audio and face attention modules for expression and identity. An audio-to-SMPL-X model predicts the motion sequence when the user supplies no pose input, so it still works audio-only. Note the abstract reports no quantitative metrics or benchmark comparisons, so treat the quality claims as unevaluated for now.