MobileWan: Closing the Quality Gap for Mobile Video Diffusion

2026-05-05 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Internet of Things (IoT) & Connected Devices · Depth: Expert, extended

Summary

MobileWan, introduced on November 6, 2025, is a novel system that enables a server-scale 5B-parameter Wan2.2 video diffusion transformer to run efficiently on memory-constrained mobile hardware. This development addresses the quality gap where existing mobile models, typically 0.4-1.8B parameters, fall short. MobileWan achieves this through a combination of recurrence distillation, which converts video generation into a chunk-wise autoregressive process with constant-memory attention, and a learnable attention head pruning method optimized with a noise-biased sparsity objective. Further enhancements include sampling-step distillation and memory-optimized VAE decoding. The system generates 5-second 480x832 videos at 16 FPS with 20 seconds end-to-end latency, scoring 83.79 on VBench and outperforming the previous state-of-the-art mobile model, Neodragon, in 80% of user preferences.

Key takeaway

For AI Engineers and MLOps Engineers tasked with deploying large video diffusion models on mobile devices, MobileWan demonstrates that you can achieve server-scale quality without resorting to drastically smaller models. You should explore integrating recurrence distillation and learnable attention head pruning techniques to overcome memory and computational constraints. This approach allows for high-fidelity video generation on devices like the Snapdragon® 8 Gen. 5 NPU, significantly narrowing the quality gap for your on-device generative AI applications.

Key insights

Server-scale 5B-parameter video diffusion models can run on mobile via recurrent reformulation and structured compression.

Principles

• High-quality mobile video does not require small models.
• Recurrence distillation enables constant-memory attention computation.
• Noise-biased training improves pruning robustness for diffusion models.

Method

Combines recurrence distillation for chunk-wise autoregressive processing, learnable attention head pruning with noise-biased optimization, sampling-step distillation (DMD/D-DMD), and memory-optimized VAE decoding with an extended causal look-back window.

In practice

• Convert transformer blocks to recurrent structures for constant memory footprint.
• Employ learnable, noise-biased head pruning for aggressive model compression.
• Extend VAE causal look-back windows to reduce temporal artifacts in generated video.

Topics

• Mobile Video Generation
• Video Diffusion Models
• Transformer Architectures
• Model Compression
• Recurrence Distillation
• Attention Head Pruning
• VAE Optimization

Code references

• ModelTC/lightx2v
• madebyollin/taehv
• Qualcomm-AI-research/fastforward

Best for: Machine Learning Engineer, Computer Vision Engineer, Research Scientist, AI Engineer, MLOps Engineer, AI Scientist

Related on AIssential

• Veda: Scalable Video Diffusion via Distilled Sparse Attention — Generation quality in sparse attention depends on mask alignment with full attention's tile-wise geometry, not just sparsity ratio.
• ComfyUI Update: Stable Video Diffusion on 8GB vram with 25 frames and more. — ComfyUI's latest update enhances video generation, sampling efficiency, and high-resolution image consistency.
• UniTemp: Unlocking Video Generation in Any Temporal Order via Bidirectional Distillation — UniTemp enables flexible, any-direction video generation by overcoming causal model limitations with bidirectional distillation and anchor latents.
• Real-Time AttentionBender: Granular Interactive Network Bending of Video Diffusion Transformers — Real-Time AttentionBender offers granular, interactive control over Video Diffusion Transformers, enhancing creative agency and model understanding.
• Qwen-Image-Flash: Beyond Objective Design — Effective few-step distillation relies more on the training recipe than just distillation objectives.
• Mode Seeking meets Mean Seeking for Fast Long Video Generation — Decoupling local fidelity and long-term coherence enables minute-scale video generation from limited long-form data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.