Diffusion models have made significant progress in both text-to-image (T2I) generation and text-guided image editing. However, these models are typically built with billions of parameters, leading to high latency and increased deployment challenges. While on-device diffusion models improve efficiency, they largely focus on T2I generation and lack support for image editing. 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. It concatenates images horizontally as input, using a (target | blank) configuration for generation tasks and (target | source) for editing tasks. 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 high-quality SFT and reinforcement learning, DreamLite achieves GenEval (0.72) for image generation and ImgEdit (4.11) for image editing, outperforming existing on-device models and remaining competitive with several server-side models. By employing step distillation, we further reduce denoising processing to just 4 steps, enabling our DreamLite could generate or edit a 1024 x 1024 image in less than 1s on a Xiaomi 14 smartphone. To the best of our knowledge, DreamLite is the first unified on-device diffusion model that supports both image generation and image editing.
Figures & tables
Figure 1: Generation (left) and Editing (right) examples of DreamLite.
Figure 2: Overview of DreamLite architecture. It consists of three primary modules: a UNet backbone, a Variational Autoencoder (VAE), and a text encoder. To support unified generation and editing, an in-context conditioning framework is adopted that unifies both tasks at the input level.
Figure 3: Architectural evolution of DreamLite. We calculate the Parameters (Params) and GFLOPs of the U-Net.
Category
Source / Content Type
Samples
Part A: Generation Data (20.0M)
General Perception
COYO, LAION, JoourneyDB, etc. .
14.8M
Human & Portrait
High-quality human, Portrait subsets
0.4M
Graphic Design
Text content, Canva subsets
0.4M
Scene Text
Dense/Large scene text
1.5M
Artistic Styles
Midjourney style prompts
0.6M
Table 1: Detailed distribution of the 21.7M total samples.
Model
Param.
Single Obj.
Two Obj.
Counting
Colors
Position
Color Attr.
Overall
FLUX.1-Dev [ 24 ]
12B
0.99
0.81
0.79
0.74
0.20
0.47
0.67
BAGEL [ 10 ]
7B
0.99
0.94
0.81
0.88
0.64
0.63
0.82
OmniGen2 [ 52 ]
4B
1.00
0.95
0.64
0.88
0.55
0.76
0.80
Z-Image [ 6 ]
6B
1.00
0.94
0.78
0.93
0.62
0.77
0.84
LongCat-Image [ 48 ]
6B
0.99
0.98
0.86
0.86
0.75
0.73
0.87
DeepGen1.0 [ 50 ]
2B
0.99
0.97
0.68
0.90
0.71
0.71
0.83
Table 2: Evaluation on GenEval benchmark. All images are generated at 1024×1024 resolution. “Param.” denotes backbone parameters.
Model
Param.
Global
Entity
Attribute
Relation
Other
Overall
FLUX.1-Dev [ 24 ]
12B
82.1
89.5
88.7
91.1
89.4
84.0
BAGEL [ 10 ]
7B
88.9
90.4
91.3
90.8
88.7
85.1
OmniGen2 [ 52 ]
4B
88.8
88.8
90.2
89.4
90.3
83.6
Z-Image [ 6 ]
6B
93.4
91.2
93.2
92.2
91.5
88.1
LongCat-Image [ 48 ]
6B
89.1
92.5
92.0
93.3
87.5
86.6
DeepGen1.0 [ 50 ]
2B
84.2
91.0
88.6
94.0
85.3
84.6
Table 3: Evaluation on DPG benchmark. All images are generated at 1024×1024 resolution. “Param.” denotes backbone parameters.
Model
Param.
Add
Remove
Replace
Adjust
BG
Style
Extract
Action
Hybrid
Overall
Kontext-Dev [ 23 ]
12B
4.04
2.92
4.28
3.89
4.00
4.43
2.71
4.52
3.08
3.76
BAGEL [ 10 ]
7B
3.81
3.16
3.85
3.59
3.39
4.51
1.58
4.25
2.67
3.42
OmniGen2 [ 52 ]
4B
3.57
3.20
3.74
3.06
3.57
4.81
1.77
4.68
2.52
3.44
Z-Image [ 6 ]
6B
4.40
4.13
4.57
4.14
4.14
4.85
4.30
4.50
3.63
4.30
LongCat-Image [ 48 ]
6B
4.66
4.68
4.75
4.64
4.35
4.81
3.90
4.83
3.80
4.49
DeepGen1.0 [ 50 ]
2B
4.52
4.43
4.42
4.09
4.38
4.48
1.75
4.57
3.61
4.03
Table 4: Evaluation on ImgEdit benchmark. “Param.” denotes backbone parameters.
Model
Param.
QSC
QPQ
QO
Kontext-Dev [ 23 ]
12B
6.61
7.24
6.79
BAGEL [ 10 ]
7B
7.66
7.08
7.20
OmniGen2 [ 52 ]
4B
7.11
7.22
6.79
LongCat-Image [ 48 ]
6B
7.87
7.49
7.55
DeepGen1.0 [ 50 ]
2B
7.83
7.47
7.54
VIBE [ 1 ]
1.6B
7.61
7.34
7.28
Table 5: Evaluation on GEdit-EN benchmark. “Param.” denotes backbone parameters. QSC , QPQ , and QO denote Semantic Consistency, Perceptual Quality and Overall score evaluated by Qwen-VL-2.5 respectively.
Figure 4: Comparison of image generation. All input prompts are listed at the bottom of the images.
Figure 5: Comparison of image editing. All source images and instructions are from ImgEdit Benchmark.
Exp.
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 Recipe
In-context
T2I → T2I
0.65
-
In-context
T2I → Edit
0.64
3.88
In-context
T2I → Unified
0.65
3.14
Table 6: Ablation study on GenEval and ImgEdit benchmarks. “TPJ” denotes “Task-progressive Joint”.
Figure 6: Ablation study on generation (left) and editing (right) tasks. From left to right in each task: results without RL, our full model, and our model after 4-step distillation.
Figure 7: DreamLite generation and editing examples on mobile device.
Component
Snapdragon 8 Gen3
Dimensity 9300
Model Size
VAE Encoder (fp16)
22.17
26.16
2.45 MB
U-Net (w8a8, per step)
103.84
122.53
389.00 MB
VAE Decoder (fp16)
22.17
26.16
2.45 MB
Total (4 steps, U-Net)
415.36
490.12
393.90 MB
Table 7: On-device inference latency (ms) for DreamLite components at 1024×1024 resolution. Measurements are conducted on W8A8 quantized models.
The remarkable generation quality of modern diffusion models often comes at the cost of massive parameter counts, which necessitate server-side inference with significant computational costs and potential privacy risks. Consequently, there is growing momentum toward developing efficient on-device alternatives. While recent efforts have optimized text-to-image models for mobile hardware, they remain relatively bulky, typically ranging from 0.5B to 1B parameters. We present BlazeEdit, a highly efficient, generalist image-to-image diffusion model tailored for on-device deployment. By identifying that many practical image editing tasks do not require text-based guidance, we eliminate the text-conditioning components and develop a multi-task architecture that consolidates object removal, outpainting, tone correction, relighting, and sticker generation into a single, compact model of only 195M parameters. BlazeEdit achieves a substantial reduction in download size and memory overhead while maintaining competitive generation quality. It completes a full inference pass in just 290ms on a Pixel 10, delivering a seamless, privacy-preserving, and lightning-fast experience for generalist image editing on the edge.
We present Swift-Image, a compact unified model for text-to-image generation, single-image editing, and multi-image editing. Our goal is to explore how far a relatively small visual generator can be pushed through systematic training engineering under a constrained computational budget. Swift-Image adopts an efficient 6B single-stream DiT and a progressive training pipeline that evolves from broad semantic coverage to higher resolution, stronger visual quality, and unified generation-editing supervision. For post-training, we employ parallel expert reinforcement learning followed by multi-teacher on-policy distillation to alleviate interference among heterogeneous objectives. We further decouple high-level reasoning from pixel-level rendering with a Prompt Enhancer that translates user requests into generator-aligned visual specifications. For efficient deployment, structural pruning and few-step distillation produce 3B and accelerated variants. Swift-Image achieves leading aggregate performance among evaluated open-source models with only 6B parameters and 243K GPU training hours; the compressed 3B model incurs nearly no loss, while few-step distillation further improves aggregate editing performance with substantially fewer sampling steps. Our study also summarizes practical lessons for architecture, data curriculum, post-training, prompt enhancement, and model compression.
Recent advances in diffusion transformers (DiTs) have set new standards in image generation, yet remain impractical for on-device deployment due to their high computational and memory costs. In this work, we present an efficient DiT framework tailored for mobile and edge devices that achieves transformer-level generation quality under strict resource constraints. Our design combines three key components. First, we propose a compact DiT architecture with an adaptive global-local sparse attention mechanism that balances global context modeling and local detail preservation. Second, we propose an elastic training framework that jointly optimizes sub-DiTs of varying capacities within a unified supernetwork, allowing a single model to dynamically adjust for efficient inference across different hardware. Finally, we develop Knowledge-Guided Distribution Matching Distillation, a step-distillation pipeline that integrates the DMD objective with knowledge transfer from few-step teacher models, producing high-fidelity and low-latency generation (e.g., 4-step) suitable for real-time on-device use. Together, these contributions enable scalable, efficient, and high-quality diffusion models for deployment on diverse hardware.