HumanOrbit: 3D Human Reconstruction as 360° Orbit Generation
| Authors | Keito Suzuki et al. |
| Year | 2026 |
| Field | Computer Vision |
| arXiv | 2602.24148 |
| Download | |
| Categories | cs.CV |
Abstract
We present a method for generating a full 360° orbit video around a person from a single input image. Existing methods typically adapt image-based diffusion models for multi-view synthesis, but yield inconsistent results across views and with the original identity. In contrast, recent video diffusion models have demonstrated their ability in generating photorealistic results that align well with the given prompts. Inspired by these results, we propose HumanOrbit, a video diffusion model for multi-view human image generation. Our approach enables the model to synthesize continuous camera rotations around the subject, producing geometrically consistent novel views while preserving the appearance and identity of the person. Using the generated multi-view frames, we further propose a reconstruction pipeline that recovers a textured mesh of the subject. Experimental results validate the effectiveness of HumanOrbit for multi-view image generation and that the reconstructed 3D models exhibit superior completeness and fidelity compared to those from state-of-the-art baselines.
Engineering Breakdown
Plain English
HumanOrbit solves the problem of generating a full 360° video orbit around a person starting from just a single image. Rather than adapting image diffusion models (which struggle with view consistency and identity preservation), the authors built a video diffusion model that generates continuous, geometrically consistent novel views while keeping the person's appearance intact. The key insight is that recent video diffusion models produce more photorealistic and coherent results than image-based approaches when asked to synthesize multi-view sequences. The method takes one photo as input and outputs a complete rotating video sequence around the subject, with plans to reconstruct 3D geometry from the generated frames.
Core Technical Contribution
The core novelty is applying video diffusion models to multi-view human synthesis instead of adapting image diffusion models. Image diffusion approaches suffer from view inconsistency and identity drift across different angles, but video models naturally enforce temporal and spatial coherence across frames. The authors designed a video diffusion architecture that conditions on the input image and camera rotation parameters to guide the generation of consistent 360° orbits. This is a significant shift in the technical approach—moving from frame-by-frame image generation to sequence-level video generation, which implicitly solves the consistency problem through the model's learned temporal dynamics.
How It Works
The system takes a single input image and camera trajectory parameters (specifying a 360° orbit) and feeds them into a video diffusion model. The diffusion model operates in a latent space, progressively refining noise into coherent video frames across the orbit sequence. The input image is encoded as a conditioning signal so the model understands what person to keep consistent, while the camera trajectory guides where the viewpoint moves at each timestep. The model generates all frames of the orbit jointly rather than independently, which allows it to maintain geometric consistency—if the person's pose and shape are consistent frame-to-frame, the 3D structure implicitly stays valid. After generation, the multi-view frames can be fed into 3D reconstruction pipelines (like NeRF or mesh-fitting algorithms) to produce the final 3D human model.
Production Impact
This enables a completely new product capability: turning a single selfie or portrait into an interactive 3D avatar or digital human without needing multi-camera rigs or 3D scanning equipment. For e-commerce, social media, and metaverse applications, this dramatically lowers the cost of generating 3D content—you no longer need professional 3D artists or expensive capture systems. The main trade-off is inference latency: video diffusion models are computationally expensive, likely requiring 10-30 seconds of generation time per orbit on high-end GPUs, and they demand significant VRAM (probably 16-24GB minimum). Integration is straightforward—ingest a single image, run the diffusion pipeline, store the generated frames or 3D reconstruction—but you'll need GPU infrastructure and need to budget for the generation cost per user request.
Limitations and When Not to Use This
The method inherits all the limitations of diffusion models: slow inference (not real-time), high memory requirements, and difficulty controlling fine details in the output (the user has limited control over pose, expression, or clothing details in the generated orbit). The paper doesn't address how it handles occlusions, extreme poses, or unusual camera trajectories beyond smooth 360° rotations—edge cases like hands near the face or complex arm positions may fail. Identity preservation is claimed but not quantitatively validated; there's no mention of perceptual loss metrics or face recognition accuracy checks across the orbit. The reconstruction step (converting video to 3D) is mentioned but not detailed, so the full pipeline quality depends on downstream 3D fitting methods that may introduce their own errors and artifacts.
Research Context
This work sits at the intersection of video generation (building on recent advances in video diffusion models) and multi-view synthesis (a classic problem in 3D computer vision). It responds directly to the failure modes of prior image-diffusion-based methods like SyncDreamer or similar works that struggle with view consistency. The paper likely contributes to benchmarks like WILD or synthetic multi-view datasets, advancing the state-of-the-art in single-image human reconstruction. This opens a research direction around conditioning video models for constrained generation tasks—how to guide diffusion models to follow specific 3D constraints, camera paths, or geometric requirements while maintaining photorealism.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::