Swift-Image: Exploring the Performance Frontier of Compact Unified Image Generation and Editing Models

Forum topic · 小凯 · 2026-08-22

Summary

Swift-Image is a compact unified model for text-to-image generation, single-image editing, and multi-image editing, introduced in the arXiv paper 2608.20334 by Taihang Hu, Zhao Wang, Zuan Gao, and Tao Liu. The work investigates how far a relatively small visual generator can be pushed under a restricted compute budget through systematic training engineering. Swift-Image uses an efficient 6B single-stream DiT with a progressive training pipeline that evolves from broad semantic coverage to higher resolution, stronger visual quality, and unified generation-editing supervision. Post-training combines parallel expert reinforcement learning with multi-teacher on-policy distillation to reduce interference between heterogeneous objectives, while a Prompt Enhancer translates user requests into generator-aligned visual specifications, decoupling high-level reasoning from pixel-level rendering. Structural pruning and few-step distillation yield 3B and accelerated variants. With only 6B parameters and 243K GPU training hours, Swift-Image reportedly achieves leading comprehensive performance among evaluated open-source models; the compressed 3B variant shows nearly no loss, and few-step distillation further improves editing performance with greatly reduced sampling steps.

Overview

Field: Computer Vision (CV) Authors: Taihang Hu, Zhao Wang, Zuan Gao, Tao Liu Published: 2026-08-22 arXiv: 2608.20334

Abstract

Swift-Image is a compact unified model designed for text-to-image generation, single-image editing, and multi-image editing. The central question of the paper is how far a relatively small visual generator can be pushed under a constrained compute budget through systematic training engineering.

Key aspects of the approach:

Architecture: An efficient 6B single-stream DiT (Diffusion Transformer).
Progressive training pipeline: Training evolves from broad semantic coverage toward higher resolution, stronger visual quality, and unified generation-editing supervision.
Post-training: Parallel expert reinforcement learning followed by multi-teacher on-policy distillation mitigates interference between heterogeneous objectives.
Prompt Enhancer: Translates user requests into generator-aligned visual specifications, decoupling high-level reasoning from pixel-level rendering.
Compression: Structural pruning and few-step distillation produce 3B and accelerated variants.

Results

With only 6B parameters and 243K GPU training hours, Swift-Image achieves the leading overall performance among evaluated open-source models.
The compressed 3B model retains performance with almost no loss.
Few-step distillation further improves overall editing performance while drastically reducing the number of sampling steps.

---

*Automatically collected on 2026-08-22*

Paper link: https://arxiv.org/abs/2608.20334