💻 computer science

MobileWan: Closing the Quality Gap for Mobile Video Diffusion

MobileWan introduces a novel framework that enables the efficient deployment of a high-quality 5B-parameter video diffusion transformer on memory-constrained mobile devices through recurrent reformulation, structured attention pruning, and distillation, achieving state-of-the-art generation performance with low latency.

Original authors: Mohsen Ghafoorian, Denis Korzhenkov, Adil Karjauv, Ioannis Lelekas, Noor Fathima, Spyridon Stasis, Hanno Ackermann, Boris van Breugel, Markus Nagel, Fatih Porikli, Animesh Karnewar, Amirhossein Habibi

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohsen Ghafoorian, Denis Korzhenkov, Adil Karjauv, Ioannis Lelekas, Noor Fathima, Spyridon Stasis, Hanno Ackermann, Boris van Breugel, Markus Nagel, Fatih Porikli, Animesh Karnewar, Amirhossein Habibian

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a master chef (a massive, powerful AI model) who can cook incredible, high-definition video meals. The problem is, this chef's kitchen is the size of a football stadium and requires a warehouse full of ingredients. You want to take this chef and put them in a tiny, cramped food truck (your smartphone) to cook the same delicious meals on the go.

Usually, when you try to shrink a giant chef down to fit in a food truck, you have to fire half the staff and use a smaller, cheaper stove. The result? The food is okay, but it's not as tasty or detailed as the stadium version.

MobileWan is the team's solution to this problem. They didn't just shrink the chef; they completely reorganized the kitchen so the giant 5-billion-parameter "stadium chef" (based on the Wan2.2 model) can actually work inside a smartphone without running out of space or time.

Here is how they did it, using simple analogies:

1. The "Memory Black Hole" Problem

Video AI models usually try to remember every single frame of a video at once to keep the motion smooth. On a computer with a huge hard drive, this is fine. On a phone, this is like trying to hold a library of books in your pocket at the same time—it causes the phone to crash (Out of Memory).

The Fix: The "Chunking" Strategy
Instead of trying to remember the whole movie at once, MobileWan breaks the video into small "chunks" (like short scenes). It generates one chunk, saves a tiny summary of it, and then moves to the next.

The Analogy: Imagine writing a story. Instead of keeping every word you've ever written in your head, you write a paragraph, jot down a quick summary on a sticky note, and then move to the next paragraph. You only need to hold the sticky note in your hand, not the whole book. This allows the phone to generate long videos without running out of memory.

2. The "Overworked Staff" Problem

Inside the AI, there are many "attention heads" (think of them as different team members looking at the video). In a 5-billion parameter model, there are hundreds of these team members. Some are doing great work, but many are just repeating what others are doing or doing very little.

The Fix: The "Smart Layoff"
The team used a special method to identify which team members were actually essential and which ones could be let go.

The Analogy: Imagine a committee of 100 people trying to decide on a menu. The team realized that 30 of those people were just nodding along or repeating the same ideas. They quietly removed those 30 people. The remaining 70 people worked harder and smarter, and the menu (the video) still turned out delicious, but the meeting was much faster and required less space.

3. The "Slow Cooking" Problem

Even with a smaller team and a chunking strategy, cooking a 5-second video usually takes the AI 100 steps (like taking 100 tiny bites to eat a meal). This is too slow for a phone.

The Fix: The "Fast-Forward Recipe"
They used a technique called "distillation."

The Analogy: Imagine a student (the mobile model) learning from a master teacher (the big server model). Instead of making the student take 100 small steps to learn a dance, the teacher shows them the whole dance, and the student learns to do it in just 3 big, confident steps. The result is the same dance, but it's done much faster.

4. The "Blurry Replay" Problem

When you play a video back on a phone, the decoder (the part that turns the AI's math into actual pictures) can sometimes make the motion look jerky or glitchy, like a bad video call.

The Fix: The "Better Camera Lens"
They upgraded the decoder to look back further into the past when creating each new frame.

The Analogy: If you are drawing a movie, and you only look at the frame immediately before the current one, your drawing might look shaky. MobileWan's decoder looks back at the last 4 frames to ensure the movement flows smoothly, like a director checking the last few shots to make sure the actor's movement is consistent.

The Result

By combining these tricks, the team managed to fit a massive, high-quality video AI onto a commercial smartphone (powered by a Snapdragon chip).

The Performance: It can generate a 5-second video (480x832 resolution) in about 20 seconds.
The Quality: In tests, the videos were rated almost as high as the massive server versions and were preferred by human users 80% of the time over the previous best mobile model.
The Claim: They proved you don't need to sacrifice quality to get video AI on a phone; you just need to be smarter about how the model thinks and remembers.

In short, MobileWan is like taking a luxury cruise ship and fitting it into a speedboat without losing the luxury experience, by redesigning the engine, trimming the crew, and changing the route.

Technical Summary: MobileWan

Problem Statement

Recent advancements in video diffusion have been driven by scaling transformer-based architectures (DiT) to billions of parameters, significantly improving visual fidelity and motion coherence. However, mobile video generation lags substantially behind server-scale systems due to strict constraints on peak memory, the quadratic complexity of self-attention, decoder overhead, and end-to-end latency. Existing mobile models are typically restricted to small backbones (0.4–1.8B parameters), which inevitably limits generation quality and creates a pronounced gap between mobile and server-scale performance. The core challenge is that bottlenecks are not confined to a single module; simply replacing softmax attention with linear attention often requires retraining from scratch or incurs an expressiveness gap, while hybrid approaches have not yet fully converted the backbone into a low-memory recurrent architecture suitable for mobile deployment.

Methodology

The authors propose MobileWan, a system that demonstrates a server-scale 5B-parameter video diffusion transformer (based on Wan2.2-5B) can be deployed efficiently on memory-constrained mobile hardware through a combination of recurrent reformulation and structured compression. The methodology consists of four primary optimization stages:

1. Learnable Head Pruning

Instead of heuristic block pruning, which the authors found Wan2.2-5B to be highly sensitive to, they introduce a learnable head pruning method.

Mechanism: A learnable scalar gate is introduced for each self-attention head, modulating its contribution. These gates are optimized end-to-end using a sparsity regularizer and a noise-biased learning strategy.
Training Strategy: The model is trained with a high-noise-biased distribution (drawing noise $\sigma \sim \text{LogitNormal}(1.5, 1)$ ) during the gate learning phase. This emphasizes global structure and motion, making the pruning decisions more robust. The process involves a three-phase schedule: soft gate exploration, temperature annealing to binary values, and final fine-tuning with a fixed pruned architecture.
Result: This approach outperforms heuristic pruning, particularly at aggressive pruning ratios, yielding measurable FLOP reductions with limited quality loss.

2. Recurrence Distillation

To address the quadratic memory complexity of self-attention, the authors convert the pretrained transformer into a chunk-wise autoregressive model.

Hybrid Attention: They employ a hybrid attention mechanism combining causal linear attention with local softmax attention. The model attends to a small local token set via softmax for immediate context and uses linear attention for historical context.
RNN Reformulation: At inference time, the causal linear attention allows the model to operate as a Recurrent Neural Network (RNN). It maintains compact running aggregates (state vectors $s_t$ and $z_t$ ) for past chunks rather than storing the full attention matrix.
Benefit: This reduces the memory footprint to be constant regardless of sequence length, enabling the generation of long videos on devices with limited RAM.

3. Step Distillation

To reduce latency, the authors apply Distribution Matching Distillation (DMD) and its variant D-DMD.

Approach: They distill the multi-step teacher model into a few-step student model.
Optimization: Experiments compared standard DMD, D-DMD (splitting the loss into Distribution Matching and Classifier-Free Guidance augmentation terms), and DMD2 (incorporating adversarial loss).
Selection: While DMD2 with 2 steps achieved the highest dynamic degree and VBench scores, the authors noted that D-DMD produced visually appealing samples with slightly lower dynamic degrees. The final system utilizes a configuration optimized for 3 denoising steps (NFEs) to balance quality and speed.

4. Decoder Optimization

Standard Video VAEs use causal 3D convolutions which are computationally expensive.

Modification: The authors build upon the efficient LightX2V decoder but modify its architecture to extend the causal look-back window.
Mechanism: Instead of conditioning on immediate past frames, the prediction of frame $F_t$ is conditioned on a longer temporal context (up to $F_{t-4}$ ).
Result: This reduces temporal artifacts and flickering while maintaining the efficiency of 2D convolutions, significantly improving reconstruction quality (PSNR and VBench) compared to other efficient decoders.

5. Quantization Pipeline

To enable deployment on the Snapdragon® 8 Gen 5 NPU, the model undergoes a two-stage quantization process:

Post-Training Quantization (PTQ): Compresses weights to 8-bit.
Quantization-Aware Training (QAT): A two-stage distillation process (block-wise followed by end-to-end) aligns the quantized model with the full-precision teacher, recovering accuracy lost during PTQ.

Key Results

MobileWan achieves the following performance metrics on a mobile device (Snapdragon® 8 Gen 5 NPU):

Generation Specs: Generates 5-second videos at 480×832 resolution, 16 FPS.
Latency: End-to-end latency of 20 seconds (including 6.65 seconds for the DiT denoising step).
Quality: Achieves a VBench score of 83.79, which is comparable to the server-scale Wan2.2 model (83.12) and significantly outperforms previous mobile state-of-the-art (Neodragon: 81.61).
Memory: The recurrent reformulation reduces the peak memory requirement from "Out of Memory" (OOM) on mobile to 9.9 GB for the DiT and 1.9 GB for the decoder.
User Study: In a blind user study, participants preferred MobileWan over Neodragon 80% of the time. In comparisons against the finetuned server Wan2.2, MobileWan was not preferred in 53% of cases but was preferred in 22%, indicating it is competitive with the server model in half of the cases.

Significance and Claims

The paper claims to bridge the quality gap between mobile and server-scale video generation by challenging the assumption that mobile video generation requires small models.

Scalability: It demonstrates that a 5B-parameter video DiT can be deployed natively on commercial mobile hardware through fine-grained compression and recurrent reformulation.
Efficiency vs. Quality: The work proves that aggressive optimizations (pruning, recurrence, distillation) can be applied without an extreme loss in quality, achieving a VBench score (83.79) that rivals server models.
System Integration: MobileWan is presented as the first 5B-scale video diffusion model deployable on a commercial mobile device, establishing a new state-of-the-art in mobile video generation.

The authors acknowledge limitations, including suboptimal quality for certain scene types (e.g., mid-distance human faces), occasional oversaturation or reduced motion depending on the distillation method, and potential temporal discontinuities or flickering in some generated videos. They intend to address these in future work.