Swift-Image: A Compact 6B Unified Model Pushing the Performance Frontier of Image Generation and Editing

Forum topic · 小凯 · 2026-08-24

Summary

Swift-Image is a compact unified visual generation model that handles text-to-image generation, single-image editing, and multi-image editing within one architecture. The system uses an efficient 6B-parameter single-stream Diffusion Transformer (DiT) trained with a progressive pipeline that moves from broad semantic coverage to higher resolutions, 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. A Prompt Enhancer decouples high-level reasoning from pixel-level rendering, translating user requests into generator-aligned visual specifications. For deployment, structural pruning and few-step distillation yield 3B and accelerated variants. Despite using only 6B parameters and 243K GPU training hours, Swift-Image achieves leading overall performance among evaluated open-source models; the compressed 3B model retains nearly all capability, and few-step distillation further improves editing quality with fewer sampling steps. The paper also documents practical lessons on architecture, data curriculum, post-training, prompt enhancement, and model compression. arXiv: 2608.20334.

Overview

Swift-Image (arXiv:2608.20334) is a compact unified model covering text-to-image generation, single-image editing, and multi-image editing. The central question: how far can a relatively small visual generator be pushed through systematic training engineering under a constrained computational budget?

Key Points

Architecture: An efficient 6B-parameter single-stream DiT (Diffusion Transformer).
Progressive training pipeline: 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, mitigating interference between heterogeneous objectives.
Prompt Enhancer: Decouples high-level reasoning from pixel-level rendering, translating user requests into generator-aligned visual specifications.
Efficient deployment: Structural pruning and few-step distillation produce 3B and accelerated variants.

Results

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

The work also summarizes practical engineering experience across architecture design, data curriculum, post-training, prompt enhancement, and model compression.