Organizations: Nanyang Technological University, Singapore. · Institute for Infocomm Research, A*STAR, Singapore. · Ant Group, China. · Nanjing University of Information Science and Technology, Nanjing, China.
We present DiVA, a deeply interactive digital life simulator pioneering a new paradigm for long-term, open-ended interactive experiences within digital character worlds. DiVA's architecture pairs a Multimodal Large Language Model (MLLM) as a router with a meticulously designed stacked video pipeline for seamless, multi-turn interactions with action and audio response. To maintain continuity and avoid degradation, we model generation as a three-part coupled system: waiting video, action video, and the transitions between them. These transitions are critically handled by our Anchored Video Continuation (AVC) module, which returns the character to stable states to prevent degradation. By encoding information from the preceding action video segment, AVC ensures smooth transitions, significantly reducing camera jitter and inconsistencies common in current video transition methods. This design also enables complex pose changes (e.g., sitting to standing) typically difficult for audio-driven models. These system designs together ensure high-fidelity identity, coherence, and dynamics for extended experiences. To validate our pipeline design, we comprehensively compare our system against alternatives by replacing our core generation module with mainstream long-video, continuation, and interpolation methods. We further analyze the necessity of the three-stage design, anchor-state selection, transition naturalness, spatial grounding, and the quality-latency trade-off, and we expand the comparison to additional long-form audio-driven avatar models. Results confirm DiVA is markedly superior in maintaining long-term visual quality and realism, validating its effectiveness as a sustainable, interactive simulation.
Figures & tables
Figure 1: Our DiVA enables infinite, continuous interaction with undiminished quality (five rounds shown here). Users can define character personality and interact via simultaneous verbal and non-verbal (spatial click) inputs, receiving corresponding realistic actions and voice replies. DiVA demonstrates advanced state transition capabilities (3rd interaction), while also inheriting the MLLM’s strong ability for factual responses (e.g., describing Hello Kitty), enabling limitless simulation within interactive digital worlds.
Paradigm
Speech Alignment
Macro Transition
State Routing
Quality Reset
E2E General long video generation
N/A
Unstable
None
None
E2E Audio-avatars
Strong
Restricted
Implicit
None
Decoupled Stateful Pipeline
Strong
Robust
Explicit
Anchored
Table 1: Positioning of representative video generation paradigms for DLS, including E2E general long video generation (FramePack, LongLive, LongCat-Video), E2E Audio-Avatars (InfiniteTalk, StableAvatar, WanS2V, Live Avatar, X-Streamer)
Figure 2: Given a text prompt defining the character’s role, an MLLM acts as a router, processing user inputs (e.g., clicks and dialogue) and translating them into character responses and behavioral prompts for the audio-text conditional action model. Our video generation component starts with a waiting video, followed by iterative action videos based on the MLLM output. Finally, the anchored video continuation (AVC) module, conditioned on the prior action, targets preset, high-quality anchor frames that represent distinct character states (e.g., sitting or leaning back). Guided by the MLLM state assessment, AVC transitions to the appropriate anchor state. This mechanism smoothly restores the sequence to a stable, high-quality condition and enables expressive transitions between states, significantly expanding the character’s motion range.
Figure 3: Qualitative comparison on long-term digital life simulation. Our method successfully generates high-quality, expressive, and drift-free results, accurately performing state transitions (e.g., the 3rd to 4th transition in both examples). In contrast, all baselines suffer from severe drift and fail to execute precise state transitions; for instance, InfiniteTalk’s subject only partially stands up in the second example. Notably, all baselines exhibit severe drift in fewer than 10 interaction rounds (marked in the bottom-right of each image), whereas our method remains stable indefinitely. Best viewed zoomed in.
Figure 4: Demonstration of our pipeline’s high-quality continuation capability. In this example, the model initially generates a coffee cup not present in the user-provided scene. Our AVC module seamlessly resolves this by generating a high-quality continuation of the character, plausibly setting the cup down. Subsequently, AVC executes a complex state transition: zooming out as the character stands and simultaneously closes the laptop, enabling more dynamic possibilities for character simulation.
Method
Subject Consis. ↑
Dynamic Degree ↑
Image Quality ↑
Quality Drift ↓
Sync-C ↑
Sync-D ↓
Click Acc. ↑
FramePack
0.8979
0.4167
0.5961
0.1293
–
–
–
InfiniteTalk
0.9202
0.3667
0.6143
0.0947
7.21
7.32
–
LongCat-Video
0.8948
0.2833
0.6238
0.1195
–
–
–
DiVA
0.9451
0.4833
0.6481
0.0341
7.53
7.16
0.8200
Table 2: Quantitative results. The best result is shown in bold . Higher is better except for Quality Drift and Sync-D.
Method
Quality Drift ↓
StableAvatar ( Tu et al., 2025 )
0.1393
WanS2V ( Gao et al., 2025a )
0.1151
InfiniteTalk ( Yang et al., 2025a )
0.0947
DiVA
0.0341
Table 3: Quality Drift for additional audio-driven avatar baselines. Lower is better.
Figure 5: Comparison of video frame interpolation methods. We compare our method against Wan FLF2V. Our approach generates more natural camera movements and character pose transitions, while the base Wan method (relying solely on the last frame) exhibits less harmonious camera work and character motion. Our method effectively avoids these inconsistencies. We recommend viewing the video samples for a clear demonstration of camera and character motion.
Method
Transition Fidelity ↑
Physical Plausibility ↑
Wan FLF2V
0.233
0.283
DiVA
0.767
0.717
Table 4: User study results on transition naturalness. Preference scores for Transition Fidelity (TF) and Physical Plausibility (PhyP). The best results are in bold .
Figure 6: Qualitative gallery of transition naturalness across diverse character styles. Under substantial pose and camera shifts, Wan FLF2V often introduces an abrupt shot change, whereas DiVA maintains continuous motion and camera flow.
Configuration
Subject Consis. ↑
Dynamic Degree ↑
Quality Drift ↓
1 Anchor (static)
0.8842
0.4145
0.0911
Historical anchors
0.8633
0.4512
0.1425
5 Curated
0.9451
0.4833
0.0341
Table 5: Anchor configuration ablation. Historical anchors are generated frames from the preceding interaction history rather than curated visual states.
Figure 7: Ablation of spatial click input. Direct coordinate input struggles with precise localization, leading to semantic confusion between adjacent features (e.g., ear versus eye). Visual markers provide explicit spatial grounding for the MLLM.
Figure 8: Comparison between high fidelity configuration and causal distillation.
Figure 9: Persona-driven interaction. Our method dynamically defines distinct character personas, enabling varied roles with appropriate verbal and non-verbal responses. The examples include a companion persona offering empathetic interaction with fitting gestures and an academic assistant explaining complex PDFs (e.g., “Attention Is All You Need”). The assistant provides factually correct explanations and uses gestures that enhance clarity.
Figure 10: Discussion of visual hallucination. The video model’s powerful generative imagination allows it to follow complex prompts, such as visualizing “Transformer” as a square matrix or hallucinating a key. These cases demonstrate advanced prompt-following fidelity, even when the results defy physical common sense.
State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University, China · China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd., China · Shenzhen Key Laboratory of Computer Vision and Pattern Recognition, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China