EgoControl: Controllable Egocentric Video Generation via 3D Full-Body Poses
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "EgoControl: Controllable Egocentric Video Generation via 3D Full-Body Poses".
Paper summary: Tom: So we're talking about EgoControl today. This paper is about making video generation controllable by the body of the person wearing the camera in an egocentric setting. It claims that you can use explicit three-dimensional full-body poses to generate future frames that match those poses and look realistic.
Jane: That sounds really practical for applications, especially since egocentric vision captures how we actually experience things—like driving or walking—and you need to predict what the view will look like next based on your intended movement. But the main thing is that existing state-of-the-art models don't let you control the actual articulated body parts of the camera wearer directly.
Lu: Exactly, that's the core gap they are addressing in "EgoControl: Controllable Egocentric Video Generation via three dee Full-Body Poses" <ref:2511.18173#pg2> <ref:2511.18173#pg1,EgoControl: Controllable Egocentric Video Generation via 3D Full-Body Poses>. They argue that in egocentric vision, your head movement handles the big camera motion, but the arms and hands—the local body parts—create all those occlusions and interactions you see.
Meng: So if a model can't predict how your hands will move based on a planned action, it can't simulate what happens when you reach for something or hide behind an object. That sounds like it would be really useful for things like augmented reality or even remote operation systems where the visual feedback has to match the intended physical action.
Lalam: Lalam's looking at the paper, and I see they are proposing a latent diffusion model trained on egocentric data that conditions future frames on both past observations and a sequence of target body poses. It's essentially learning how to map those explicit three dee poses into the video generation process <ref:2511.18173#pg1> <ref:2511.18173#pg1>.
Tom: Right, so they aren't just looking at your current pose; they are conditioning the generation on a whole sequence of future intended poses, which is what makes it more than just a simple image prompt. They need that full sequence of control to generate temporally coherent video.
Paper summary: Jane: And to get that full control, they introduce a new way to represent the pose—a compact representation that captures both the global camera motion and the detailed articulated body movements simultaneously. That’s where the method gets interesting <ref:2511.18173#pg6>.
Lu: They break it down into two parts: one part captures how fast the camera is moving, which they call delta h and delta r, and another part tracks all twenty-one joints relative to a pelvis at each frame, which they call J <ref:2511.18173#pg6>.
Meng: So they are creating this combined pose representation P as delta h, delta r, and J. It sounds like a lot of data to handle at once when you're training the diffusion model. What makes it better than just feeding the raw joint data?
Tom: The paper says this specific way of formulating the pose representation, which is differential rather than cumulative, gives them a more stable signal that is temporally accurate <ref:2511.18173#pg9>. It’s a technical detail they are emphasizing because they found it worked better for learning the dynamics.
Jane: That stability comes from how they integrate it into the generation process through two different control mechanisms, which I think is key to why their results are strong <ref:2511.18173#pg7>. They use both modulation and cross-attention to condition the video generation on that full body pose P.
Lu: They use a modulation approach with some two layer machine learning planes, ge and gm, to map the flattened pose vector Pflat into an embedding space where they get parameters for scaling, shifting, and gating in the diffusion process <ref:2511.18173#pg8>. This is like subtly nudging the generation based on your overall posture.
Tom: And then they add cross-attention where each frame's pose vector gets projected to some feature space to make a sequence of pose tokens, c m = PE(P′m), and those tokens are used as context for the diffusion model <ref:2511.18173#pg9>. It’s combining global nudges with localized, frame-by-frame control signals.
Meng: From an engineering standpoint, that sounds like a lot of computation during inference because you're running two pathways—modulation and cross-attention—for every frame generation step. What about the cost of running this model compared to just using a simpler head pose control method?
Paper summary: Jane: The paper shows that combining both methods gives them the best results, and it’s similar in performance to using only cross-attention for body control in terms of alignment metrics like mean Intersection-over-Union. It seems the dual approach is necessary for good visual quality and motion accuracy <ref:2511.18173#pg9>.
Lu: They found that AdaLN, which is their adaptive layer normalization method, captures the global styling and head motion well, but it struggles to enforce fine-grained articulated body control #Ablation Study Findings. The cross-attention pathway seems better for that specific body alignment detail.
Tom: That's interesting because it means they can pinpoint exactly where the model is failing—global style versus local articulation—and the cross-attention pathway targets that weakness effectively <ref:2511.18173#pg9>. It sounds like a very fine tuning process for the control signals.
Jane: So, what does this mean for us who are building embodied AI agents? It suggests we can finally move past just controlling the camera's view and start controlling the whole agent's physical interaction with its environment.
Meng: I wonder about generalization. The paper mentions they didn't explicitly model fine-grained hand articulation because their training data just didn't have those specific hand pose annotations #Limitations. So, if you want to control every finger movement perfectly, this might not be the answer yet without more specialized data.
Lu: They also noted a limitation regarding the training data itself; it’s biased toward fisheye cameras and subjects in motion-capture suits, so generalizing that to everyday viewpoints and clothing is still a challenge #Limitations.
Tom: So, to summarize the main point of "EgoControl: Controllable Egocentric Video Generation via three dee Full-Body Poses," they successfully translate global camera dynamics and articulated body motion into temporally coherent videos by using a compact representation for the pose and combining modulation with cross-attention control #Conclusion <ref:2511.18173#pg1,EgoControl: Controllable Egocentric Video Generation via 3D Full-Body Poses>.
Jane: It really shows that full-body information is critical if you want full control over egocentric videos, significantly improving both visual fidelity and motion alignment compared to previous methods #Conclusion.
Conclusion: Tom: So, to wrap up this paper on "EgoControl: Controllable Egocentric Video Generation via three dee Full-Body Poses," they show that you can take a short history of what your camera has seen and tell the AI exactly how you want your body to move in the next few frames.
Jane: It's about using explicit three-dimensional poses, or full body poses, to give direct control over the video generation process. The authors are essentially showing that you don't just need a text prompt for a video; you can use these three dee poses as an extra instruction set for the model.
Lu: They introduce a new pose representation that combines global camera motion with the relative positions of all your joints to your pelvis at each frame. This lets them capture both how fast the camera is moving and exactly how your arms and legs are positioned relative to each other.
Meng: From my side, it's interesting because they used two different ways to control the generation—a modulation approach that nudges the whole video, and a cross-attention pathway that gives frame-by-frame fine control over your body alignment. It sounds like a lot of computation for every single frame you generate.
Tom: Exactly, and they found that putting both methods together gives them the best results for both visual quality and motion accuracy compared to using just one or the other. That's a solid number behind it.
Jane: What does this mean in practice? It means we move past just controlling where the camera points, into controlling how you physically interact with things in a simulated world.
Lu: It shows that for any embodied AI agent, having this level of control over the visual output is really important for simulating actions and understanding consequences.
Tom: The implication is that future systems won't just be reactive; they'll be predictive and controllable based on a full body plan. So, we need to keep looking at how these pose representations scale up to handle more complex human movements in real-world scenarios.
Enrico Pallotta, Sina Mokhtarzadeh Azar, Lars Doorenbos, Serdar Ozsoy, Umar Iqbal, Juergen Gall
University of Bonn · Lamarr Institute for Machine Learning and Artificial Intelligence · NVIDIA
cs.CV
Submitted: 2025-11-22
Updated: 2026-11-17
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 80/100
The gist: EgoControl is a pose-controllable video diffusion model trained on egocentric data that generates temporally coherent and visually realistic future frames conditioned on explicit 3D full-body poses…
Key concepts
- Egocentric Vision
- This is first-person vision, like what a person sees through a camera they are holding. It captures the sensorimotor experience of an agent, meaning it tightly links the camera's movement with the body's movements, making it crucial for tasks like navigation and manipulation.
- Pose Representation P
- This is a compact mathematical description of the human body's state. It combines two parts: global camera motion (head transformation) and articulated body dynamics (joint transformations relative to the pelvis). This unified representation allows the model to understand both where the camera is looking and how the body is moving.
- Modulation Control
- This control mechanism uses two small neural networks to map a flattened pose vector into an embedding space. These embeddings generate parameters (scaling, shifting, gating) that are used to adjust features within the diffusion model's layers, providing a global influence on the generated video.
- Cross-Attention Control
- This method projects each frame's pose vector into a sequence of 'pose tokens.' These tokens are then used in a cross-attention pathway within the diffusion model. This provides temporally localized control signals that help align the generated video frames precisely with the desired body movements.
Terminology
Summary
EgoControl is a pose-controllable video diffusion model trained on egocentric data that generates temporally coherent and visually realistic future frames conditioned on explicit 3D full-body poses of the camera wearer, which is key for controllable embodied video simulation
Introduction to the Problem
Egocentric vision captures the sensorimotor experience of an agent, characterized by rapid camera motion and tight coupling between body pose and visual observations This perspective is critical for problems where perception and action must co-design one another, such as embodied navigation, manipulation, augmented reality, and human–robot collaboration For an embodied agent, the ability to forecast future egocentric frames is essential to internally simulate the visual consequences of planned actions Existing state-of-the-art generative video models do not provide direct, explicit control over the camera wearer’s articulated body in the egocentric setting
EgoControl Architecture
EgoControl is a latent diffusion model trained to synthesize future egocentric frames conditioned on a short sequence of past observations and a sequence of target body poses The model conditions the generation on two factors: the past visual context x=(x1, x2,..., xN), i.e., the current observed state, and the future motion intent expressed by the human pose sequence P=(p1, p2,. p. 2.., pM) The goal is to generate egocentric video frames yi that not only look realistic but also visually coherent with their corresponding control pose pi ∈ P
Pose Representation and Encoding
The authors propose a compact and informative pose representation P = [∆h, ∆r, J] ∈ R M×23×6 that jointly captures global camera motion and articulated body dynamics This representation is constructed by computing the relative head transformation between consecutive frames as ∆Hi = HiH−1 i−1, ∀i ∈ [1, M] Furthermore, they compute the relative transformations of all joints with respect to the pelvis at each frame denoted as J ∈ RM×21×6 This allows EgoControl to capture both egocentric camera dynamics and articulated body motion in a unified form
Control Mechanisms
Two control mechanisms are proposed to condition the video generation on the full human body pose representation P The first approach via modulation uses a pair of two-layer MLPs ge and gm to map the flattened pose vector Pflat into an embedding and modulation space, yielding parameters for scaling, shifting, and gating in the diffusion process The second approach via cross-attention involves projecting each frame’s pose vector Pm into a feature space to yield a sequence of M pose tokens cm = PE(P′m), which are then used as context for the cross-attention pathway in the diffusion model The authors show that the combination of both methods gives the best results
Experimental Results and Evaluation
Experiments use Nymeria, a large dataset of human motion captured ‘in the wild’ containing over 1,100 high-resolution egocentric video sequences The evaluation protocol assesses visual quality using metrics like Frechet Video Distance (FVD) and SSIM, global motion control via translation and rotation error, and agent body control by measuring the mean Intersection-over-Union (IoU) between ground truth and generated arm masks Results show that controlling the video generation by the full-body 3D pose yields substantial gains over baselines, with mIoU increasing by nearly 55% relative to head-only control
Limitations and Future Work
The study reveals limitations, noting that the model does not explicitly model fine-grained hand articulation because the training data does not provide explicit hand-pose annotations Furthermore, the training data is biased toward a specific acquisition setup with fisheye-mounted cameras and subjects wearing a motion-capture suit, which may reduce out-of-distribution generalization to everyday viewpoints and clothing Future work should extend pose conditioning to include explicit hand representations and utilize additional training data to broaden camera and clothing variability
Ablation Study Findings
Ablation studies on the control mechanism confirm that AdaLN captures global styling and head motion better but struggles to enforce fine-grained articulated body control, while cross-attention provides more granular control over body alignment The differential formulation for pose representation is shown to be superior to the cumulative representation because it provides a more stable and temporally accurate signal A pelvis-centric coordinate system proves to be a more robust and easier-to-learn representation for articulated body dynamics compared to frame-toframe differential encoding for joints
Conclusion
EgoControl successfully translates global camera dynamics and articulated body motion into temporally coherent, pose-consistent egocentric videos, improving substantially both visual fidelity and motion alignment compared to baselines The model demonstrates that full-body information is critical to have full control over egocentric videos By introducing a compact pose representation and a dedicated control pathway, EgoControl translates global camera dynamics and articulated body motion into temporally coherent, pose-consistent egocentric videos, improving substantially both visual fidelity and motion alignment compared to baselines The study reveals clear limitations that point to promising future work
Acknowledgements
The authors acknowledge funding from the Federal Ministry of Research, Technology and Space, the Ministry of Culture and Science of the State of North Rhine-Westphalia, the Ministry of Science, Research and Arts of the State of Baden-Wurttemberg, the Bavarian State Ministry of Science and the Arts and the Gauss Centre for Supercomputing e.V. (GCS) for funding this project by providing computing time on the Supercomputer JUPITER at Julich Supercomputing Centre (JSC) of Forschungszentrum Julich through the Gauss AI Compute Competition The authors also thank the German AI Service Center WestAI for providing additional computational resources
References
[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. p. 8 arXiv preprint arXiv:2311. p. 815127 4
[6] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang et al. Videocrafter1: Open diffusion models for high-quality video generation. p. 8 arXiv preprint arXiv:2310. p. 819512 6
[7] Dima Damen, Hazel Doughty, Giovanni Maria Farinella et al. Scaling egocentric vision: The epic-kitchens dataset. p. 8 In Proceedings of the European conference on computer vision (ECCV), pages 720–736, 2018 7
[9] Stephanie Fu, Netanel Tamir, Shobhita Sundaram et al. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. p. 8 arXiv preprint arXiv:2306. p. 809344 9
[11] Shenyuan Gao, Jiazhi Yang, Li Chen et al. Vista: A generalizable driving world model with high fidelity and versatile controllability. p. 8 Advances in Neural Information Processing Systems, 37:91560–91596, 2024 1
[30] Lingni Ma, Yuting Ye, Fangzhou Hong et al. Nymeria: A massive collection of multimodal egocentric daily motion in the wild. p. 9 In European Conference on Computer Vision, pages 445–465. p. 9 Springer, 2024 3
[37] William Peebles and Saining Xie. Scalable diffusion models with transformers. p. 9 In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023 1
[47] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach et al. Fvd: A new metric for video generation. p. 5 arXiv preprint, 2019 5
[56] Shiyuan Yang, Liang Hou et al. Direct-a-video: Customized video generation with userdirected camera movement and object motion. p. 4 In ACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024
--- Page 1 ---
EgoControl: Controllable Egocentric Video Generation via 3D Full-Body Poses
Enrico Pallotta Sina Mokhtarzadeh Azar Lars Doorenbos Serdar Ozsoy Umar Iqbal Juergen Gall cvg-bonn.github.io/EgoControl Figure 1 EgoControl generates highly realistic egocentric videos from observed context frames and allows detailed control over the body of the camera wearer. p. 0 The camera view of the generated frames follows the head pose and the visible body parts of the human pose interact with the objects in the scene in a plausible manner. p. 0 Abstract Egocentric video generation with fine-grained control through body motion is a key requirement towards embodied AI agents that can simulate, predict, and plan actions In this work, we propose EgoControl, a pose-controllable video diffusion model trained on egocentric data We train a video prediction model to condition future frame generation on explicit 3D body pose sequences To achieve precise motion control, we introduce a novel pose representation that captures both global camera dynamics and articulated body movements, and integrate it through a dedicated control mechanism within the diffusion process Given a short sequence of observed frames and a sequence of target poses, EgoControl generates temporally coherent and visually realistic future frames that align with the provided pose control Experimental results demonstrate that EgoControl produces high-quality, pose-consistent egocentric videos, paving the way toward controllable embodied video simulation and understanding
--- Page 2 ---
Egocentric (first-person) vision has recently become a central topic in computer vision [38, 46] Unlike third-person imagery, egocentric video captures the sensorimotor experience of an agent, characterized by rapid camera motion, frequent occlusions from hands and objects, and a tight coupling between body pose and visual observations This perspective is critical for problems where perception and action must co-design one another, such as embodied navigation, manipulation, augmented reality, and human–robot collaboration Consequently, the community has invested heavily in large-scale egocentric datasets and benchmarks [7, 14, 30], as well as models that can understand [25] and predict [25] first-person behavior To be useful in these scenarios, generative models must not only be photorealistic but also controllable with fine-grained, physically plausible body-level commands that reflect how a human or a robot would move Existing state-of-the-art generative video models are impressive at content synthesis and, more recently, at conditioning on high-level signals such as text prompts or camera trajectories However, these conditioning modalities do not provide direct, explicit control over the camera wearer’s articulated body
--- Page 3 ---
In the egocentric setting, this is a critical gap: camera motion is produced by the wearer’s head global translation and rotation, while local articulated motions (arms, hands, legs) produce the occlusions and object interactions that define first-person scenes Without the ability to specify full-body pose sequences, a model cannot reliably simulate the visual consequences of specific actions In this work, we introduce EgoControl, the first egocentric video generative model that enables explicit control over the camera wearer’s body through 3D full-body poses
--- Page 4 ---
EgoControl is a latent diffusion model trained to synthesize future egocentric frames conditioned on a short sequence of past observations and a sequence of target body poses We design an informative pose representation that jointly captures global camera motion and articulated body dynamics, and integrate it via a dedicated control pathway within the diffusion process
--- Page 5 ---
Egocentric datasets typically provide frame-synchronized 3D body poses defined in a global reference frame However, since EgoControl is designed to operate from an embodied perspective, we require a pose representation that reflects the agent’s own motion rather than its absolute position in space Each 3D pose is given as a set of J = 23 transformation matrices describing the position and orientation of each joint with respect to the global frame From an egocentric viewpoint, most perceptual changes in the video arise from the movement of the camera, which corresponds to the head motion
--- Page 6 ---
To incorporate full-body state, we also compute the relative transformations of all joints with respect to the pelvis at each frame, denoted as J ∈ RM×21×6, where 21 is the number of joints excluding head and pelvis Finally, we concatenate these components to obtain the full pose representation that will control the generation of the M future frames: P = [∆h, ∆r, J] ∈ R M×23×6
--- Page 7 ---
Figure 2. EgoControl generates future egocentric frames conditioned on the past frames and sequence of human poses. p. 3 We condition the model on the human poses in two ways The purple branch uses the networks ge and gm to obtain modulation vectors that are combined with the ones obtained from the diffusion step The blue branch tokenizes the poses and computes the cross-attention with the encoded visual tokens Both branches are necessary to generate realistic egocentric video frames that are well-aligned with the human pose sequence
--- Page 8 ---
Inspired by [2, 22, 35], we use a modulation and gating approach to update the features in the DiT [37] blocks A pair of two-layer MLPs ge and gm then map this vector to the embedding and modulation space Different shift, scale and gate parameters are predicted for the self-attention, cross-attention, and MLP components in each block
--- Page 9 ---
The resulting outputs of the self-attention, cross-attention, and MLP components are gated and combined through residual connections This mechanism allows P to modulate normalization and residual pathways across all layers As illustrated in Fig. 2, this mechanism is also used to integrate the diffusion step t
--- Page 9 ---
We propose to use two control mechanisms to condition the video generation on the full human body pose representation P The first approach via modulation is described in Sec. 3.2.1 and the second approach via cross-attention is described in Sec. p. 3 3.2.2
--- Page 9 ---
Inspired by [1, 2, 25], we use a modulation and gating approach to update the features in the DiT [37] blocks A pair of two-layer MLPs ge and gm then map this vector to the embedding and modulation space Different shift, scale and gate parameters are predicted for the self-attention, cross-attention, MLP components in each block
--- Page 9 ---
Each component input u is then normalized and modulated via adaptive layer normalization: AdaLN(k) (u; β(k)P, γ(k)P) = LN(u) ⊙ (1 + γ(k)P) + β(k)P The resulting outputs of the self-attention, cross-attention, and MLP components are gated and combined through residual connections This mechanism allows P to modulate normalization and residual pathways across all layers
--- Page 9 ---
In addition to the AdaLN-based modulation, we introduce a fine-grained control pathway by further conditioning the transformer blocks through cross-attention with body pose tokens as illustrated in Fig. 2 This cross-attention pathway provides temporally localized control signals that complement the global modulation from AdaLN, together enabling coherent and controllable egocentric video synthesis
--- Page 9 ---
Each frame’s pose vector Pm ∈ R 23×6 is projected to the model’s feature space as P′m = LayerNorm(GELU(WpPm)) ∈ R Dp, (u) yielding a sequence of M pose tokens These pose tokens are then concatenated and used as context c for the cross-attention in Eq. (9) This cross-attention pathway provides temporally localized control signals that complement the global modulation from AdaLN, together enabling coherent and controllable egocentric video synthesis
--- Page 9 ---
The authors show that the combination of both gives the best results The combination of both methods yields superior performance in both visual quality and motion accuracy, while being similar to using only cross-attention for body control (mIoU)
--- Page 9 ---
The authors show that the combination of both gives the best results
--- Page 9 ---
Figure 4 Pose alignment (mIoU) and SSIM for each of the 32 generated frames.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Full-body pose conditioning allows for precise motion control by translating
body motion into realistic egocentric visual outcomes,
enabling fine-grained action conditioning such asreaching, turning while walking, or performing a hand-object manipulation.
-
The proposed hybrid control mechanism combines modulation and cross-attention to achieve superior performance, specifically noting that the combination yields results
superior in both visual quality and motion accuracy, while being similar to using only cross-attention for body control (mIoU).
-
The model can generate temporally coherent egocentric sequences that follow intended motions while adapting to the scene and objects, as shown in qualitative results where
EgoControl manages to control both the camera view and body movements.
-
The system can achieve high alignment of generated arms with controlling arm poses, with the mIoU metric increasing by nearly 55% relative to head-only control when using full-body pose conditioning.
-
The model demonstrates robustness across different initial contexts, showing that
regardless of the initial state of the person, the generated video is successfully controlled by the pose condition.
Sources
- Cosmos World Foundation Model Platform for Physical AI
- Whole-Body Conditioned Egocentric Video Prediction
- Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets
- MagicPose: Realistic Human Poses and Facial Expressions Retargeting with Identity-aware Diffusion
- VideoCrafter1: Open Diffusion Models for High-Quality Video Generation
- DreamSim: Learning New Dimensions of Human Visual Similarity using Synthetic Data
- AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning
- CameraCtrl: Enabling Camera Control for Text-to-Video Generation
- Imagen Video: High Definition Video Generation with Diffusion Models
- ViPE: Video Pose Engine for 3D Geometric Perception
- EgoM2P: Egocentric Multimodal Multitask Pretraining
- Intention-driven Ego-to-Exo Video Generation
- Movie Gen: A Cast of Media Foundation Models
- SAM 2: Segment Anything in Images and Videos
- GAIA-2: A Controllable Multi-View Generative World Model for Autonomous Driving
- Make-A-Video: Text-to-Video Generation without Text-Video Data
- Wan: Open and Advanced Large-Scale Video Generative Models
- EgoExo-Gen: Ego-centric Video Prediction by Watching Exo-centric Videos
- CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer
- MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs