Research map / Video / Video generation
Neodragon: Mobile Video Generation Using Diffusion Transformer
Animesh Karnewar, Denis Korzhenkov, Ioannis Lelekas, Noor Fathima, Adil Karjauv, Mohsen Ghafoorian, Amir Habibian
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace T5-XXL with distilled DT5 + ContextAdapter (Replace-Mode) | Pyramidal-Flow · VBench | VBench total: 80.31 -> 79.64 | No clear effect |
| use LoRA adapters instead of MLP ContextAdapter (ablation) | DT5 text encoder · VBench | VBench total: 79.64 -> 64.74 | Hurt |
| extend (cascade) CA with frozen CE instead of replacing CE (ablation) | DT5 distillation · VBench | VBench total: 79.64 -> 79.16 | Hurt |
| add cosine distance loss to distillation (ablation) | DT5 Replace-Mode · training stability | convergence | Improved |
| swap native decoder for distilled TinyAEHV decoder | Pyramidal-Flow · VBench | VBench total: 80.31 -> 80.25 | No clear effect |
| prune 6 of 24 MMDiT blocks, Stage-1 finetune only | Pyramidal-Flow MMDiT · VBench | VBench total: 80.31 -> 78.39 | Hurt |
| add Stage-2 full-teacher distillation after Stage-1 finetune (ablation) | 18-block MMDiT · VBench | VBench total: 78.39 -> 80.21 | Improved |
| prune 8 of 24 blocks with Stage-1 only (ablation) | MMDiT · VBench | VBench total: 80.31 -> 74.59 | Hurt |
| prune blocks (24 -> 18) for latency | MMDiT on Hexagon NPU · latency | per-step latency: 1.15s -> 0.74s | Improved |
| apply Stage-2 directly to pruned model without Stage-1 (ablation) | pruned MMDiT · VBench | VBench total | Hurt |
| use Next-Block mapping with data FM loss only plus feature loss instead of all losses/simple mapping (ablation) | 18-block MMDiT Stage-2 · VBench | VBench total: 80.21 -> 80.35 | Mixed |
| use Pyramidal DMD step distillation with 1-1-1 schedule vs native 1-1-1 (ablation) | Pyramidal-Flow · VBench | VBench total: 59.62 -> 76.48 | Improved |
Closest papers by meaning
- MobileWan: Closing the Quality Gap for Mobile Video Diffusion (NeurIPS 2026 · similarity 0.91)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.87)
- AsymRnR: Video Diffusion Transformers Acceleration with Asymmetric Reduction and Restoration (ICML 2025 · similarity 0.87)
- SnapGen-V: Generating a Five-Second Video within Five Seconds on a Mobile Device (CVPR 2025 · similarity 0.87)
- LinGen: Towards High-Resolution Minute-Length Text-to-Video Generation with Linear Computational Complexity (CVPR 2025 · similarity 0.87)
- GenTron: Diffusion Transformers for Image and Video Generation (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
