Organizations: The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China. · The University of Tokyo, Japan. · Microsoft, Beijing. · Hong Kong Baptist University, Hong Kong.
The growing demand for text-to-image generation has led to rapid advances in generative modeling. Recently, text-to-image diffusion models trained with flow matching algorithms, such as FLUX, have achieved remarkable progress and emerged as strong alternatives to conventional diffusion models. At the same time, inference-time enhancement strategies have been shown to improve the generation quality and text-prompt alignment of text-to-image diffusion models. However, these techniques are mainly applicable to conventional diffusion models and usually fail to perform well on flow models. To bridge this gap, we propose Reflective Flow Sampling (RF-Sampling), a theoretically-grounded and training-free inference enhancement framework explicitly designed for flow models, especially for the CFG-distilled variants (i.e., models distilled from CFG guidance techniques), like FLUX. Departing from heuristic interpretations, we provide a formal derivation proving that RF-Sampling implicitly performs gradient ascent on the text-image alignment score. By leveraging a linear combination of textual representations and integrating them with flow inversion, RF-Sampling allows the model to explore noise spaces that are more consistent with the input prompt. Extensive experiments across multiple benchmarks demonstrate that RF-Sampling consistently improves both generation quality and prompt alignment. Moreover, RF-Sampling is also the first inference enhancement method that can exhibit test-time scaling ability to some extent on FLUX.
Figures & tables
Fig. 1 : Qualitative comparisons with three representative flow models. Images for each prompt are synthesized using the same random seed. More visualization results are in Appendix -H .
Fig. 2 : RF-Sampling outperforms standard sampling with the same time consumption and significantly enhances the performance of FLUX-Lite and FLUX-Dev. With the increase of inference time, RF-Sampling consistently performs well, validating the scalability of our method. (Breakdown is shown in Appendix Tab. XV )
Fig. 3 : Illustration of RF-Sampling. Compared to previous methods, RF-Sampling employs interpolation on text embeddings similar to the traditional CFG, thereby enhancing the model’s generation quality and making it more suitable for flow diffusion models, especially CFG-distilled models.
Animation
Concept-art
Painting
Photo
Average
Model
Method
AES(↑)
HPSv2(↑)
AES(↑)
HPSv2(↑)
AES(↑)
HPSv2(↑)
AES(↑)
HPSv2(↑)
AES(↑)
HPSv2(↑)
Standard
5.9474
30.93
6.1926
28.59
6.4161
28.84
5.4077
27.66
5.9909
29.01
GI [ 31 ]
5.9814
26.23
6.2188
23.48
6.2188
23.61
5.3417
23.81
5.9401
24.28
Z-Sampling [ 13 ]
5.8729
30.58
6.0427
27.58
6.2579
28.21
5.4394
27.92
5.9032
28.57
CFG++ [ 32 ]
5.8329
29.81
6.0969
27.41
6.3206
27.81
5.3969
27.04
5.9118
28.02
CFG-Zero* [ 17 ]
5.9743
31.22
6.2066
29.27
6.4280
29.22
5.4190
27.65
6.0061
29.34
TABLE I : Main experiments on HPDv2 [ 30 ] dataset across 3 different models. The experiments show the consistent superior performance compared with previous methods, highlighting the effectiveness of our RF-Sampling. Note that other baselines are not applicable to FLUX.
Pick-a-Pic
DrawBench
Model
Method
PickScore(↑)
ImageReward(↑)
AES(↑)
HPSv2(↑)
PickScore(↑)
ImageReward(↑)
AES(↑)
HPSv2(↑)
Standard
21.99
85.13
5.9435
29.32
22.60
86.02
5.4591
27.76
GI
21.19
28.94
5.9534
24.63
22.11
47.53
5.4279
23.96
Z-Sampling
21.73
89.03
5.9091
28.84
22.55
92.05
5.4784
28.06
CFG++
21.79
85.17
5.8821
28.50
22.54
81.80
5.3757
27.18
CFG-Zero*
21.88
86.78
5.9536
29.37
22.66
91.90
5.4511
28.10
TABLE II : Main experiments on Pick-a-Pic [ 33 ] and DrawBench [ 34 ] datasets across 3 different models. Obviously, our proposed RF-Sampling exhibits superior performance across 4 different metrics. Note that other baselines are not applicable to FLUX.
Fig. 4 : The winning rate of RF-Sampling over other methods on SD3.5. The standard sampling (baseline) winning rate defaults to 50%. The results reveal the superiority of RF-Sampling in synthesizing images with good quality.
Fig. 5 : The winning rate of RF-Sampling over other methods on FLUX. The standard sampling (baseline) winning rate defaults to 50%. The results reveal the superiority of RF-Sampling in synthesizing images with good quality.
Fig. 6 : Ablation of the gap between shigh and slow . When the gap of shigh - slow increases within a certain range, the quality of synthesized images improves. The dotted lines represents the performance of the standard method. This indicates that within a certain range of values, RF-Sampling perform better than the standard one, demonstrating the robustness of it.
Fig. 7 : Ablation study on the effect of βlow and βhigh . No β means that we do not implement the interpolation weight in Eqn. 2 . The results reveal that following the high-weight denoising → low-weight inversion paradigm can enchance the quality of synthesized images. The dotted lines represents the performance of the standard method. This indicates that within a certain range of values, RF-Sampling perform better than the standard one, demonstrating the robustness of it.
Method
FID ( ↓ )
IS ( ↑ )
Standard
35.08
150.07
RF-Sampling
33.12
155.21
TABLE III : Comparison of FID and IS between standard sampling and RF-Sampling on ImageNet-1K. We use FLUX-Lite with inference steps 28, combining the nunchaku [ 42 ] sampling acceleration framework, to generate 5,000 samples (5 images per class). RF-Sampling achieves a lower FID and a higher IS than the standard one, demonstrating its ability to better align with the real data distribution while maintaining high-quality and diverse image generation.
Fig. 8 : Visualization of the sampling trajectories sampled by our method and the standard approach.
Fig. 9 : Ablation study of standard guidance scale w . With the increase of the standard guidance scale w , the quality of the synthesized images degrades a lot. The results guarantee that the performance gain of RF-Sampling is not introduced by the increase of the amplifying weight s .
Fig. 10 : Robustness to the RF-Sampling steps. The horizontal axis shows the ratio of RF-Sampling operations during the whole inference steps. As the ratio increases, generation quality improves, indicating effective semantic information gain throughout the whole path. The dotted lines represents the performance of the standard method. This indicates that within a certain range of values, RF-Sampling perform better than the standard one, demonstrating the robustness of it.
Fig. 11 : We explore the influence of merge ratio γ on Pick-a-Pic dataset. The results across 4 metrics reveal that γ=0.5 is a better choice, where the synthesized images are the best. The dotted lines represents the performance of the standard method. This indicates that within a certain range of values, RF-Sampling perform better than the standard one, demonstrating the robustness of it.
Fig. 12 : We combine our proposed methods with existing LoRAs in FLUX community. Our RF-Sampling can be directly applied to the corresponding downstream tasks. The synthesized images validate the effectiveness and generalizability of our method.
Model
Method
PickScore(↑)
ImageReward(↑)
AES(↑)
HPSv2(↑)
Standard
21.91
86.64
6.3224
30.12
RF-Sampling
22.05
99.21
6.5379
31.16
Standard + Nunchaku
22.07
95.94
6.2303
30.47
FLUX-Lite (28 steps)
RF-Sampling + Nunchaku
22.23
102.35
6.4171
30.86
Standard
22.06
97.47
6.2464
30.49
RF-Sampling
22.19
100.90
6.3113
31.06
TABLE IV : Orthogonal experiments with Nunchaku [ 42 ] , a sampling acceleration method for FLUX. The results demonstrate the generalizability of RF-Sampling to sampling acceleration.
Model
Method
PickScore(↑)
ImageReward(↑)
AES(↑)
HPSv2(↑)
0
21.95
88.58
6.4346
29.90
α=1
∅
21.95
87.64
6.4576
30.05
0
21.88
91.86
6.4667
29.61
FLUX-Lite (28 steps)
α=2
∅
22.05
99.21
6.5379
31.16
0
22.11
98.82
6.2566
30.00
α=1
∅
22.19
100.90
6.3113
31.06
TABLE V : We conduct ablation studies of the value of α and the form of cuncond on Pick-a-Pic dataset. It is noticed that as α increases, the quality of synthesized images improves, albeit at an inevitable cost to computational time. The form of cuncond may be either the null prompt embedding ∅ or the zero-padded prompt embedding 0 . In relation to the form of cuncond , the null prompt embedding ∅ yields superior results, as it carries a greater amount of unconditional semantic information.
Fig. 13 : Visualization of synthesized images with different s scales. shigh>slow serves as a necessary condition for achieving superior image synthesis quality. When shigh−slow<0 , RF-Sampling shows minimal advantage over the standard method with poor text generation accuracy and blurred details. However, as shigh−slow increases to positive values, RF-Sampling exhibits dramatic improvements in text rendering precision, visual detail clarity, and overall image coherence, ultimately generating high-fidelity outputs that significantly outperform Standard approaches. The results indicate that the parameter relationship shigh>slow acts as a crucial control mechanism that enables RF-Sampling to leverage its full potential for complex text-to-image generation tasks.
Fig. 14 : Visualization of synthesized images with different β scales. By applying the interpolation weight β , the model can synthesize higher-quality, more detailed, and visually appealing images that better align with user expectations for complex prompts, especially when βhigh>βlow .
Fig. 15 : Image editing experiments on FLUX-Kontext Bench [ 38 ] . Compared to the standard sampling, RF-sampling enables a more precise understanding of the given instruction, thereby achieving accurate image editing. For more examples, please see Appendix Fig. 21 .
Method
UMT-FVD (↓)
UMTScore (↑)
GPT4o-MTScore (↑)
MTScore(↑)
Standard
264.84
2.7053
3.4797
0.41497
RF-Sampling
229.49
2.9095
3.5302
0.43671
TABLE VI : We extend our method to the video generation task. Due to the computational budget, we utilize Wan2.1-T2V-1.3B [ 41 ] . The results on ChronoMagic-Bench-150 [ 37 ] across 4 metrics show the promising scalability of our method to the video generation task.
Fig. 16 : Visualizations of the sampling trajectories of RF-Sampling and the standard method. we randomly select two ImageNet classes [ 43 ] (Ambulance and Zebra) and visualize their respective data distributions. For each class, we randomly sample 6 Gaussian noises and process them through both standard diffusion sampling and RF-Sampling methods using the prompt format "a photo of class in ImageNet." The results reveal that RF-Sampling trajectories consistently demonstrate stronger convergence towards the real data distribution compared to standard method trajectories.
Attribute Binding
Object Relationship
Model
Method
Color(↑)
Shape(↑)
Texture(↑)
2D-Spatial(↑)
3D-Spatial(↑)
Non-Spatial(↑)
numeracy(↑)
Complex(↑)
Overall(↑)
Standard
0.7511
0.5709
0.7119
0.2927
0.3751
0.3166
0.6078
0.3846
0.5013
SD3.5 (28 steps)
RF-Sampling
0.7817
0.5885
0.7241
0.2864
0.3974
0.3174
0.6121
0.3844
0.5119
Standard
0.7030
0.4154
0.4887
0.2258
0.3710
0.3030
0.5564
0.3365
0.4249
FLUX-Lite (28 steps)
RF-Sampling
0.7613
0.4725
0.5970
0.2420
0.4042
0.3070
0.6090
0.3649
0.4698
Standard
0.7535
0.5018
0.6167
0.2783
0.3866
0.3078
0.6052
0.3706
0.4775
TABLE VII : We evaluate the effectiveness of RF-Sampling on T2I-CompBench [ 36 ] across 3 diffusion models. The results validate the effectiveness and generalizability of our method.
Model
Method
Single(↑)
Two(↑)
Counting(↑)
Colors(↑)
Positions(↑)
Color Attribution(↑)
Overall(↑)
Standard
0.97
0.91
0.75
0.85
0.21
0.53
0.70
SD3.5 (28 steps)
RF-Sampling
0.99
0.91
0.72
0.89
0.19
0.54
0.71
Standard
0.90
0.57
0.52
0.71
0.11
0.36
0.53
FLUX-Lite (28 steps)
RF-Sampling
0.93
0.62
0.59
0.73
0.18
0.42
0.58
Standard
0.99
0.80
0.78
0.77
0.23
0.50
0.68
FLUX-Dev (50 steps)
RF-Sampling
0.99
0.82
0.76
0.80
0.25
0.50
0.69
TABLE VIII : We evaluate the effectiveness of RF-Sampling on GenEval [ 35 ] across 3 diffusion models. The results show the superiority over the standard.
Fig. 17 : We directly extend our proposed method to video generation task on Wan2.1-T2V-1.3B. The visualizations show the superiority of our proposed method compared with standard sampling.
Fig. 18 : Visual results of FLUX-Lite with guidance scale w=1 . The generated images remain semantically aligned with the input text prompts, demonstrating that the model’s output is still conditionally generated even at the minimum guidance scale. This empirically verifies that CFG-distilled models like FLUX do not possess a true unconditional generation mode, and setting w=1 does not produce unconditional outputs.
Fig. 19 : Impact of Merge Ratio γ on generation quality. The inverted U-shaped curves across all metrics confirm the existence of an optimal step size, balancing gradient alignment and manifold constraints. FLUX-Dev shows significantly higher robustness to large γ values than FLUX-Lite, attributed to the smoother latent manifold of the larger model. Dotted curve lines represent quadratic fits to the data.
Method
PickScore(↑)
HPSv2(↑)
AES(↑)
ImageReward(↑)
Standard
21.91
30.12
6.3224
86.84
Round 1
RF-Sampling
22.05
31.16
6.5379
99.21
Standard
21.95
30.33
6.3473
93.73
Round 2
RF-Sampling
22.04
30.82
6.5231
100.81
Standard
21.94
30.20
6.3608
99.42
Round 3
RF-Sampling
21.99
30.63
6.5133
103.45
TABLE IX : To demonstrate the robustness of our method, we conducted repeated experiments on the Pick-a-Pic using FLUX-Lite with different random seeds. The results show that our approach consistently outperformed the standard method across varying random seeds, highlighting the robustness of RF-Sampling.
Method
PickScore(↑)
HPSv2(↑)
AES(↑)
ImageReward(↑)
Standard
21.99
29.32
5.9435
85.13
High Embedding Mix (s=9,β=0.7)
21.99
29.32
5.9435
85.13
Low Embedding Mix (s=−1,β=0.3)
21.99
29.32
5.9435
85.13
RF-Sampling
21.99
29.90
5.9981
101.50
TABLE X : Ablation on reflection component. We replace the full reflection step with a simple linear interpolation between embeddings (using Eqn. 2 ) under two distinct mixing weights. The results show that both linear variants fail to improve over the standard baseline, performing identically across all metrics. This demonstrates that the model-driven reflection is essential, as simpler heuristics cannot achieve the performance gains of our full RF-Sampling.
Method
PickScore(↑)
HPSv2(↑)
AES(↑)
ImageReward(↑)
s/img( ↓ )
Standard (28 steps)
21.99
29.32
5.9435
85.13
29.93
Standard (28 × 3 = 84 steps)
21.96
29.60
5.9109
89.87
67.06
Best-of-5
22.21
30.58
5.9849
106.69
154.17
Best-of-3
21.94
30.14
5.9642
100.40
97.63
RF-Sampling
21.99
29.90
5.9981
101.50
65.04
TABLE XI : Comparison of our RF-Sampling with Best-of-N method. RF-Sampling achieves a better trade-off between performance and efficiency: it outperforms standard sampling in all metrics and is competitive with Best-of-N methods. While Best-of-5 achieves the highest performance, it requires more than double the time per image compared to RF-Sampling. RF-Sampling outperforms Best-of-3 in PickScore, AES and ImageReward with approximately 1.5 times faster. These results demonstrate the effectiveness of our method in achieving high performance with reduced computational cost.
Method
PickScore(↑)
HPSv2(↑)
AES(↑)
ImageReward(↑)
s/img ( ↓ )
Standard (28 steps)
21.99
29.32
5.9435
85.13
29.93
GI (28 steps)
21.19
24.63
5.9534
28.94
31.33
CFG++ (28 steps)
21.79
28.50
5.8821
85.17
32.46
CFG-Zero* (28 steps)
21.88
29.37
5.9536
86.78
28.91
Standard (28 × 3 = 84 steps)
21.96
29.60
5.9109
89.87
67.06
GI (28 × 3 = 84 steps)
21.25
25.27
5.9335
28.16
67.04
TABLE XII : Comparison under equivalent computational budget. To demonstrate the effectiveness of our method, we compare RF-Sampling against baselines using 84 steps (28×3), matching the total inference time. The results show that RF-Sampling almost outperforms all baseline methods across different metrics while maintaining comparable time per image, demonstrating its effectiveness under a fair computational setting.
Method
Metric
Standard
Aesthetic + Random
+ ZO-2
+ Path-2
RF-Sampling
NFEs
50
2880
2880
2880
50 × 3 = 150
ImageReward
99.73
101.21
98.42
97.13
106.21
TABLE XIII : Quantitative comparisons with [ 54 ] on DrawBench. RF-Sampling requires only 150 NFEs, far fewer than the baseline methods (2880 NFEs), yet still achieves the top results in both ImageReward and AES, demonstrating the dual advantages of our method in both efficiency and effectiveness.
Method
Color(↑)
Shape(↑)
Texture(↑)
Spatial(↑)
Numeracy(↑)
Complex(↑)
Overall(↑)
Standard
0.7535
0.5018
0.6167
0.2783
0.6052
0.3706
0.5210
Aesthetic + Random (1920 NFEs)
0.7518
0.5219
0.5926
0.2893
0.6059
0.3572
0.5199
RF-Sampling (50 × 3 = 150 NFEs)
0.7761
0.5323
0.6422
0.2687
0.6082
0.3733
0.5335
TABLE XIV : Quantitative comparisons with [ 54 ] on T2I-CompBench. RF-Sampling requires only 150 NFEs, far fewer than the baseline methods (1920 NFEs), yet almost achieves the top results across different dimensions, demonstrating the dual advantages of our method in both efficiency and effectiveness.
Model
Method
NFEs
HPSv2(↑)
AES(↑)
s/img ( ↓ )
28
30.12
6.3224
34.63
50
30.39
6.3045
46.60
Standard
75
30.46
6.2864
60.61
7 × 5 + 21 = 56
30.84
6.4397
49.63
14 × 5 + 14 = 84
30.98
6.4736
64.57
21 × 5 + 7 = 112
31.04
6.5032
76.84
TABLE XV : Detailed breakdown of Fig. 2 , including step counts (NFEs) and wall time. As shown in the table below, RF-Sampling outperforms standard sampling with the same time consumption and significantly enhances the performance of FLUX-Lite and FLUX-Dev. With the increase of inference time, RF-Sampling consistently performs well, validating the scalability of our method.
Fig. 20 : We combine our proposed methods with existing LoRAs in FLUX community. Our RF-Sampling can be directly applied to the corresponding downstream tasks, validating the generalizability of our method.
Fig. 21 : We extend our proposed methods to image editing tasks on FLUX-Kontext. Our RF-Sampling can be directly applied to the corresponding downstream tasks, validating the effectiveness of our method.
Fig. 22 : The winning rate of RF-Sampling over other methods on SD3.5 on Pick-a-Pic dataset. The standard sampling (baseline) winning rate defaults to 50%
Fig. 23 : The winning rate of RF-Sampling over other methods on SD3.5 on DrawBench dataset. The standard sampling (baseline) winning rate defaults to 50%.
Fig. 24 : The winning rate of RF-Sampling over other methods on SD3.5 on the animation subset of HPD v2 dataset. The standard sampling (baseline) winning rate defaults to 50%.
Fig. 25 : The winning rate of RF-Sampling over other methods on SD3.5 on the photo subset of HPD v2 dataset. The standard sampling (baseline) winning rate defaults to 50%.
Fig. 26 : The winning rate of RF-Sampling over other methods on SD3.5 on the concept-art subset of HPD v2 dataset. The standard sampling (baseline) winning rate defaults to 50%.
Fig. 27 : The winning rate of RF-Sampling over other methods on SD3.5 on the painting subset of HPD v2 dataset. The standard sampling (baseline) winning rate defaults to 50%..
Fig. 28 : The winning rate of RF-Sampling over the standard one on FLUX-Lite and FLUX-Dev on Pick-a-Pic and DrawBench datasets. The standard sampling (baseline) winning rate defaults to 50%.
Fig. 29 : The winning rate of RF-Sampling over the standard one on FLUX-Lite on the 4 subsets of HPD v2 datasets. The standard sampling (baseline) winning rate defaults to 50%.
Fig. 30 : The winning rate of RF-Sampling over the standard one on FLUX-Dev on the 4 subsets of HPD v2 datasets. The standard sampling (baseline) winning rate defaults to 50%.
Fig. 31 : Synthesized images of FLUX-Lite on anime subset of HPD v2.
Fig. 32 : Synthesized images of FLUX-Lite on photography subset of HPD v2.
Fig. 33 : Synthesized images of FLUX-Lite on painting subset of HPD v2.
Fig. 34 : Synthesized images of FLUX-Lite on concept-art subset of HPD v2.
Fig. 35 : Synthesized images of FLUX-Lite on GenEval.
Fig. 36 : Synthesized images of FLUX-Lite on Pick-a-Pic and DrawBench.
Fig. 37 : Synthesized images of FLUX-Dev on anime subset of HPD v2.
Fig. 38 : Synthesized images of FLUX-Dev on photography subset of HPD v2.
Fig. 39 : Synthesized images of FLUX-Dev on painting subset of HPD v2.
Fig. 40 : Synthesized images of FLUX-Dev on concept-art subset of HPD v2.
Fig. 41 : Synthesized images of FLUX-Dev on GenEval.
Fig. 42 : Synthesized images of FLUX-Dev on Pick-a-Pic and DrawBench.
Hardware-agnostic strategies for accelerating text-to-image diffusion, such as timestep distillation and feature caching, can reduce inference time without custom kernels or system-level optimization. Among them, multi-resolution generation strategies have recently received broad attention, attaining more than 5x speedup without any training. However, the design of performing upsampling in the latent space, together with the selective modification of partial regions, causes these methods to exhibit noticeable blurring or artifacts. To this end, we propose MrFlow, a training-free multi-resolution acceleration strategy for pretrained flow-matching models built upon a staged low-to-high-resolution pipeline. MrFlow first rapidly generates the main structure at low resolution, then performs super-resolution in the pixel space using a lightweight pretrained GAN-based model, subsequently injects low-strength noise to enable high-frequency resampling, and finally refines the details at high resolution. Quantitative and qualitative results on FLUX.1-dev and Qwen-Image show that MrFlow exploits the quadratic token reduction and reduced step requirement of low-resolution sampling to achieve 10x end-to-end acceleration while keeping OneIG within a 1% gap relative to that before acceleration, significantly surpassing other training-free acceleration strategies, and requiring no training or runtime dynamic identification whatsoever. MrFlow can further be directly combined orthogonally with pre-trained timestep distillation strategies, achieving even higher generation acceleration of up to 25x.
Xingyu Zheng, Xianglong Liu, Yifu Ding +4
State Key Laboratory of Complex & Critical Software Environment, Beihang University · School of Computer Science and Engineering, Beihang University · Nanyang Technological University +5
Text-to-image diffusion models are capable of generating high-quality images, but suboptimal pre-trained text representations often result in these images failing to align closely with the given text prompts. Classifier-free guidance (CFG) is a popular and effective technique for improving text-image alignment in the generative process. However, CFG introduces significant computational overhead. In this paper, we present DIstilling CFG by sharpening text Embeddings (DICE) that replaces CFG in the sampling process with half the computational complexity while maintaining similar generation quality. DICE distills a CFG-based text-to-image diffusion model into a CFG-free version by refining text embeddings to replicate CFG-based directions. In this way, we avoid the computational drawbacks of CFG, enabling high-quality, well-aligned image generation at a fast sampling speed. Furthermore, examining the enhancement pattern, we identify the underlying mechanism of DICE that sharpens specific components of text embeddings to preserve semantic information while enhancing fine-grained details. Extensive experiments on multiple Stable Diffusion v1.5 variants, SDXL, and PixArt-α demonstrate the effectiveness of our method. Code is available at https://github.com/zju-pi/dice.
Zhenyu Zhou, Defang Chen, Can Wang +2
Zhejiang University, State Key Laboratory of Blockchain and Data Security · University at Buffalo, State University of New York
Text-to-image users often provide concise and underspecified prompts, whereas generative models benefit from detailed textual conditions for reliable instruction following. Existing systems bridge this gap with Prompt Enhancers (PEs) that rewrite raw prompts at inference time, introducing additional latency and leaving prompt elaboration external to the generator. We instead view enhanced prompts as privileged training information and ask whether their benefits can be internalized. We propose Prompt-Enhanced On-Policy Self-Distillation (PE-OPSD) for text-to-image flow-matching models. During training, a raw-prompt student follows its own generation trajectory, while an enhanced-prompt teacher provides vector-field targets at the states visited by the student. This on-policy supervision distills the behavior induced by enhanced prompts into the raw-prompt student without requiring additional text--image pairs. At inference, both the PE and teacher are removed, and the student generates directly from raw prompts. Across multiple model families, PEs, and benchmarks, PE-OPSD achieves the strongest aggregate prompt fidelity among the evaluated baselines, yields positive aggregate visual appeal gains, and retains the base-model inference efficiency.
Mingfeng Lin, Chengfei Cai, Lin Xu +3
Harbin Institute of Technology (Shenzhen) · Zhejiang University · Peking University +1