Research map / Image generation / Text-to-image diffusion
NanoFLUX: Distillation-Driven Compression of Large Text-to-Image Generation Models for Mobile Devices
Ruchika Chavhan, Malcolm Chadwick, Alberto Gil Couto Pimentel Ramos, Luca Morreale, Mehdi Noroozi, Abhinav Mehrotra
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 |
|---|---|---|---|
| prune heads, 12B to 3B DiT (step C2) | FLUX.1-Schnell · GenEval | score: 66.0 -> 54.1 | Hurt |
| compress DiT from 12B to 1.8B | FLUX.1-Schnell · One-IG | score: 43.1 -> 43.2 | No clear effect |
| add progressive token downsampling to 2.5B DiT (ablation) | Step C3 2.5B · HPSv3 | score: 10.68 -> 10.56 | Hurt |
| add progressive token downsampling to 1.8B DiT (ablation) | Step C4 1.8B · GenEval | score: 53.1 -> 49.8 | Hurt |
| add progressive token downsampling to 1.8B DiT (ablation) | Step C4 1.8B · latency on Samsung S25U (s) | seconds: 2.75 -> 2.40 | Improved |
| distill T5-XXL into T5-Large text encoder | Step C3 2.5B · DPG | score: 82.6 -> 76.3 | Hurt |
| distill T5-XXL into T5-Large text encoder | Step C3 2.5B · HPSv3 | score: 10.68 -> 10.45 | Hurt |
| replace pruning with block merging in single-stream blocks 7-12 (ablation) | 2.5B DiT · HPSv3 | score: 9.66 -> 10.66 | Improved |
| prune double-stream blocks instead of single-stream blocks (ablation) | 2.5B DiT · HPSv3 | score: 9.66 -> 9.41 | Hurt |
| add static AdaLN coefficients precomputed from 2 samples (ablation) | 2.5B DiT · HPSv3 | score: 10.66 -> 10.54 | No clear effect |
| use block-wise distillation loss instead of output-only distillation (ablation) | T5-Large for 2.5B DiT · HPSv3 | score: 10.37 -> 10.45 | Improved |
| use block-wise distillation loss instead of output-only distillation (ablation) | T5-Large for 1.8B DiT · HPSv3 | score: 10.21 -> 10.45 | Improved |
Closest papers by meaning
- SnapGen: Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training (CVPR 2025 · similarity 0.89)
- InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation (ICLR 2024 · similarity 0.87)
- Denoising Autoregressive Transformers for Scalable Text-to-Image Generation (ICLR 2025 · similarity 0.87)
- Meissonic: Revitalizing Masked Generative Transformers for Efficient High-Resolution Text-to-Image Synthesis (ICLR 2025 · similarity 0.87)
- KOALA: Empirical Lessons Toward Memory-Efficient and Fast Diffusion Models for Text-to-Image Synthesis (NeurIPS 2024 · similarity 0.87)
- MobileWan: Closing the Quality Gap for Mobile Video Diffusion (NeurIPS 2026 · 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.
