One‑Step Face Video Restoration and 15.7× Faster Streaming Video Models – Xiaomi Papers at ECCV 2026

In a collection of twelve papers accepted to ECCV 2026, Xiaomi's AI and autonomous‑driving teams present advances across visual understanding, multimodal interaction, and autonomous‑driving intelligence.

TIGER: High‑Quality Face Video Restoration

The TIGER framework fuses three priors—Identity, Geometry, and Generative—into a structured three‑prior fusion pipeline. Identity embeddings from reference images preserve the subject’s identity, a decoupled 3‑D parameter space provides temporally consistent geometric priors, and a generative prior models the restoration as a single‑step Rectified Flow, enabling one‑forward‑pass high‑fidelity repair of severely degraded face videos. A large‑scale face‑video restoration dataset supports robust training and standardized evaluation. Experiments show TIGER leads in identity preservation, temporal stability, and visual realism while maintaining efficient inference.

RED‑Aes: Relative Aesthetic Scoring

RED‑Aes reframes image aesthetic assessment from absolute scoring to learning the relative aesthetic difference induced by edits. The RED‑20k dataset pairs original and edited images generated by large image‑editing models, annotated with relative aesthetic differences and reasoning chains. Training first applies contrastive and causal pre‑training to inject aesthetic knowledge, then uses GRPO reinforcement learning to optimize ranking. Results demonstrate the 7B model surpasses GPT‑5 and prior expert models on zero‑shot benchmarks, while a lightweight 2B version outperforms all existing baselines.

VST: Streaming VideoLLM with Real‑Time Thinking

VST introduces a "thinking while watching" paradigm for video LLMs, allowing continuous inference as video streams in. The model generates intermediate thoughts for each new segment, maintaining a dynamic knowledge graph of events, entities, and causal relations. A two‑stage fine‑tuning pipeline (VST‑SFT and VST‑RL) and an automatic video‑KG data synthesis pipeline improve streaming reasoning. Experiments report 79.5% on StreamingBench, 59.3% on OVO‑Bench, and a 15.7× response‑time improvement over Video‑R1 on VideoHolmes, with a 5.4% performance gain.

GAIA: Data‑Flywheel Training for GUI Agents

GAIA builds a data‑flywheel system that harvests high‑quality positive and negative samples from real GUI interactions to train lightweight Critic models for test‑time scaling. During inference, multiple action candidates are sampled and filtered by the Critic, while hard negatives are fed back for continual model improvement. Experiments show over 10% increase in task success rate and a single‑token decision cost, significantly reducing inference overhead.

UniTranslator: End‑to‑End In‑Image Machine Translation

UniTranslator tackles two core challenges of image‑inside translation: semantic mismatch between translation understanding and image generation, and spatial misalignment of rendered text. It introduces an understanding‑generation alignment module with shared latent space and a spatial mask decoder that uses precise text‑region masks to constrain rendering. The system produces fully translated images without intermediate OCR, text removal, or font matching, achieving state‑of‑the‑art performance on multiple benchmarks.

ELVA: Ranking‑Driven Generalizable Multimodal Retrieval

ELVA identifies "grain blindness"—the tendency of multimodal retrievers to ignore fine‑grained attributes, actions, and relations. By incorporating ranking supervision, ELVA forces the model to prefer more relevant candidates (e.g., "person in red running on the beach" over "person in red standing in a park"). Experiments show leading results on the M‑BEIR benchmark and a 13.1% gain over LamRA‑Ret‑7B on the MRBench fine‑grained retrieval benchmark.

CausalDrive: Real‑Time Causal World Models for Autonomous Driving

CausalDrive treats future prediction as a causal modeling problem rather than pure video generation. Using only front‑view images, ego‑vehicle trajectory, and textual prompts, the Context‑Forced DMD architecture learns cause‑effect interactions among traffic participants. Continuous‑flow matching and self‑correcting distillation achieve 12 FPS real‑time simulation, moving autonomous‑driving world models toward social interaction simulation.

SWAM: Spatial‑Perceiving World Action Model for Embodied Navigation

SWAM unifies visual path generation and action planning in a single forward pass. It leverages DepthAnything V3 for pseudo‑depth supervision and generates RGB‑D visual trajectories together with corresponding actions, using a shared decoder. The approach outperforms baselines on RECON, SCAND, and TartanDrive datasets.

MindDrive: Online RL for Vision‑Language‑Action Autonomous Driving

MindDrive separates decision‑making (language instructions) and execution (trajectory points) via a shared base and dual LoRA adapters. After executing an action, the resulting reward is fed back to the language layer, enabling efficient reinforcement learning in discrete language space rather than continuous trajectory space. Experiments confirm effective online RL for VLA models.

DriveVA: Unified Video‑Action Model for Zero‑Shot Driving

DriveVA integrates future video prediction and action token generation within a shared latent generative process, using a DiT‑based decoder to produce both video latents and action tokens. Progressive video continuation enables long‑term consistency. On NAVSIM v1, DriveVA achieves 90.9 PDMS and demonstrates strong zero‑shot cross‑domain generalization to nuScenes and Bench2Drive/CARLA without fine‑tuning.

BeyondDrive: Learning from Hard Negative Trajectories

BeyondDrive generates challenging negative samples that are spatially close to expert trajectories but semantically unsafe. A reverse‑distance loss pushes the model away from these unsafe regions while staying near expert paths, improving safety. Experiments show significant gains across single‑modal and multimodal end‑to‑end driving pipelines, reaching 90.1 EPDMS on NAVSIM v2.

DriveFine: Refinement‑Augmented Masked Diffusion for Robust Driving

DriveFine combines a masked diffusion LLM with plug‑and‑play block‑level mixture‑of‑experts modules that provide self‑refinement capabilities. Expert routing and gradient isolation decouple trajectory generation from optimization, allowing iterative correction. The model attains strong performance and robustness on NAVSIM v1, NAVSIM v2, and the challenging Navhard benchmark, demonstrating reliable planning in complex and long‑tail scenarios.

These works collectively illustrate Xiaomi's systematic AI capability building across visual understanding, multimodal interaction, and autonomous‑driving intelligence, moving from isolated demos toward deployable, real‑world applications.