DreamLite

In this paper, we propose DreamLite, a compact unified on-device diffusion model (0.39B) that supports both T2I generation and text-guided image editing within a single network. DreamLite is built on a pruned mobile U-Net backbone and unifies conditioning through In-Context spatial concatenation in the latent space. To stabilize the training of this compact model, we introduce a Task-Progressive Joint pretraining strategy that sequentially targets T2I, editing, and joint tasks. After SFT and RL, DreamLite outperforms existing on-device models and remaining competitive with several server-side models in both generation and editing tasks. By employing step distillation, we further achieve 4-step inferencing, enabling our DreamLite could generate or edit a 1024 × 1024 image in ~3s (using 4-bit Qwen VL and fp16 VAE+UNet) on a iPhone17 pro.

Our contributions are summarized as follows:

• We propose, to the best of our knowledge, the first unified on-device model that supports both text-to-image generation and text-based image editing, eliminating the need to deploy two separate models.
• We introduce an in-context conditioning mechanism for UNet to unify generation and editing, and propose a task-progressive joint pretraining scheme (i.e., T2I → Edit → Unified Joint Training) to stably train the model.
• DreamLite achieves competitive performance on standard benchmarks and consistently outperforms prior mobile models. After deployment on mobile device, DreamLite could generate or edit a 1024 × 1024 image in less than 5s.

Table 1. Comparison with existing methods on GenEval, DPG, ImgEdit and GEdit-EN Benchmarks.

Method	Params	GenEval ↑	DPG ↑	ImgEdit ↑	GEdit-EN-Q ↑
FLUX.1-Dev / Kontext	12B	0.67	84.0	3.76	6.79
BAGEL	7B	0.82	85.1	3.42	7.20
OmniGen2	4B	0.80	83.6	3.44	6.79
LongCat-Image / Edit	6B	0.87	86.6	4.49	7.55
DeepGen1.0	2B	0.83	84.6	4.03	7.54
SANA-1.6B	1.6B	0.67	84.8	-	-
MEISSONIC	1B	0.54	65.3	-	-
VIBE	1.6B	-	-	3.85	7.28
SANA-0.6B	0.6B	0.64	83.6	-	-
SnapGen++ (small)	0.4B	0.66	85.2	-	-
EditMGT	0.96B	-	-	2.89	6.33
DreamLite (Ours)	0.39B	0.72	85.8	4.11	6.88

Table 2. Ablation study on GenEval and ImgEdit benchmarks. "TPJ" denotes "Task-progressive Joint".

Experiments	Mechanism	Training Stage	GenEval ↑	ImgEdit ↑
Text-to-image Pretraining			0.70	-
Condition Mechanism	Pix2Pix	T2I → Edit	0.56	3.67
	Pix2Pix	T2I → Edit → Unified	0.61	3.65
Training Strategy	In-context	T2I → T2I	0.65	-
	In-context	T2I → Edit	0.64	3.88
	In-context	T2I → Unified	0.65	3.14
	In-context	T2I → Edit → Unified	0.71	3.94
Reinforcement Learning	In-context	TPJ Pretrain → RLHF	0.72	4.11
Step Distillation	In-context	TPJ Pretrain → RLHF → DMD	0.70	3.8

@article{feng2026dreamlite,
  title={DreamLite: A Lightweight On-Device Unified Model for Image Generation and Editing},
  author={Kailai Feng and Yuxiang Wei and Bo Chen and Yang Pan and Hu Ye and Songwei Liu and Chenqian Yan and Yuan Gao},
  journal={arXiv preprint arXiv:2603.28713},
  year={2026}
}