Research map / Video / Video generation

MobileWan: Closing the Quality Gap for Mobile Video Diffusion

NeurIPS 2026 · Video › Video generation · approach: Efficient video diffusion models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Deploys a 5B Wan2.2 video diffusion transformer on a Snapdragon 8 Gen 5 phone via head pruning, recurrent linear/softmax hybrid attention, step distillation, and an optimized VAE decoder, reaching 83.79 VBench.

Problem

Mobile video diffusion models are limited to roughly 0.4–1.8B parameters because of memory, quadratic attention cost, decoder overhead, and latency, which leaves a large quality gap to server-scale video models.

Approach

Convert a pretrained 5B video DiT into a chunk-wise autoregressive model with hybrid softmax/causal linear attention that runs as an RNN at inference, combined with learned per-head gate pruning (noise-biased sparsity), DMD-based step distillation, and a decoder with extended temporal look-back.

Key results

• MobileWan reaches 83.79 VBench total on 81x480x832 video, the highest among listed mobile models (Neodragon 81.61).
• Generates 5-second 480x832 videos at 16 FPS in about 20 seconds end-to-end on a Snapdragon 8 Gen. 5 NPU; full-recurrence DiT runs a single NFE in 6.6 s.
• In an internal user study of 859 comparisons, MobileWan was preferred over Neodragon 80% of the time.

Surprising findings

• Recurrent hybrid conversion of all 30 blocks dropped the self-attention TFLOPs and memory so that the original 5B DiT, which OOMs on device, runs on mobile hardware.
• Full-model VBench scores were higher for DMD2-style distillation (83.95-84.03) but its dynamic degree rose while D-DMD yielded more pleasant colors; VBench rewards oversaturation, which may not match human preference.

Limitations

• Quality of some scene types such as human faces in the middle distance is suboptimal.
• Step distillation can cause oversaturation or reduced motion, and the decoder plus distillation can produce flicker.
• RNN reformulation sometimes leads to temporal discontinuities.

Experimental setup

Datasets: in-house synthetic videos generated by Wan2.1 14B DAVIS VBench prompts

Benchmarks: VBench

Metrics: VBench total VBench quality VBench semantic latency peak RAM human preference rate

Baselines: Neodragon Wan2.2 5B S2DiT SnapGenV Mobile Video DiT CogVideoX 5B SANA-Video Attention Surgery

Builds on: Wan2.2 distribution matching distillation D-DMD ReHyAt hybrid attention LightX2V VAE

Closest papers by meaning

• SnapGen-V: Generating a Five-Second Video within Five Seconds on a Mobile Device (CVPR 2025 · similarity 0.90)
• From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.90)
• Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.89)
• AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.88)
• OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.88)
• FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention (NeurIPS 2024 · similarity 0.88)