Research map / Image generation / Text-to-image diffusion
MobileDiffusion: Instant Text-to-Image Generation on Mobile Devices
Yang Zhao, Zhisheng Xiao, Yanwu Xu, Haolin Jia, Tingbo Hou
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 |
|---|---|---|---|
| set W_K = W_V in self-attention (ablation) | MobileDiffusion UNet · text-to-image quality | model quality: about 5% parameter reduction | No clear effect |
| replace gelu with swish in GLU (ablation) | MobileDiffusion UNet · text-to-image quality | metrics and perceptual quality | No clear effect |
| finetune softmax attention into relu attention (ablation) | MobileDiffusion UNet · text-to-image quality | visual quality | No clear effect |
| trim feed-forward expansion ratio from 4 to 3 (ablation) | MobileDiffusion UNet · MS-COCO FID | FID: +0.27 FID, 10% fewer parameters | Hurt |
| use larger separable-conv kernels (7x7, 9x9) instead of 3x3 (ablation) | MobileDiffusion separable conv block · text-to-image quality | quality | No clear effect |
| use LoRA adversarial fine-tuning instead of full-parameter fine-tuning (ablation) | UFOGen one-step model · MS-COCO FID | FID: 11.67 -> 13.45 | Hurt |
| replace diffusion reconstruction loss with distillation loss (ablation) | UFOGen one-step model · MS-COCO FID | FID: 11.67 -> 11.20 | Improved |
| replace diffusion loss with distillation loss for downstream applications (ablation) | UFOGen one-step model · downstream conditional generation | application quality | Hurt |
| use MobileDiffusion architecture instead of SD UNet (386M vs 862M params) | text-to-image diffusion · MS-COCO FID-30K, 50 step DDIM | FID-30K: 9.62 -> 8.65 | Improved |
| reduce to one-step generation with MobileDiffusion | MobileDiffusion · HPS v2 | average score: 27.30 -> 26.65 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- UFOGen — UFOGen: You Forward Once Large Scale Text-to-Image Generation via Diffusion GANs (CVPR 2024) · better on 1 of 1 shared result
- SnapFusion — SnapFusion: Text-to-Image Diffusion Model on Mobile Devices within Two Seconds (NeurIPS 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SnapGen: Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training (CVPR 2025 · similarity 0.91)
- UFOGen: You Forward Once Large Scale Text-to-Image Generation via Diffusion GANs (CVPR 2024 · similarity 0.87)
- SwiftBrush: One-Step Text-to-Image Diffusion Model with Variational Score Distillation (CVPR 2024 · similarity 0.86)
- R&B: Region and Boundary Aware Zero-shot Grounded Text-to-image Generation (ICLR 2024 · similarity 0.86)
- Towards Understanding the Working Mechanism of Text-to-Image Diffusion Model (NeurIPS 2024 · similarity 0.86)
- InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation (ICLR 2024 · similarity 0.86)
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.
