A deployable text-to-image diffusion model must hold several capabilities at once, such as instruction following, text rendering and human preference, and must sample in a few steps. Capabilities and few-step sampling are built in separate stages: a full-step model first consolidates capabilities through multi-teacher on-policy distillation, then serves as a fixed teacher for step distillation. Every full-step model we evaluate scores lower once compressed, and the second stage cannot begin until the first finishes. We introduce StepX-OPD, which merges the two stages: one student acquires every capability while learning to sample in a few steps, so frozen specialists supervise it on the states its own rollout visits rather than on a longer trajectory it never runs. Because a decomposition of the rollout's KL weights velocity mismatch most heavily near the clean end, we also regress the student's velocity on the low-noise part of the rollout. Separately, we find that capability integration makes models write prompt words into images on prompts that never asked for text, and we give a variant that lowers it. We evaluate on SD3.5-Medium and FLUX.1-dev across instruction following, text rendering, image quality and human preference. On both architectures, StepX-OPD attains the highest normalized average score among the four-step methods compared and exceeds the forty-step Flow-OPD model, and on SD3.5-Medium it trains with about 37% fewer GPU-hours than sequential DiffusionOPD followed by step distillation.
Figures & tables
Figure 1: Training cost for the displayed optimization stages (left) and four-step performance across nine benchmarks after fixed axis-wise normalization (right).
Base Model
Model
NFEs
GenEval ↑
OCR ↑
DeQA ↑
PickScore ↑
Average ↑
SD3.5-Medium
Base
80
0.666
0.529
4.045
21.79
0.7105
GenEval Teacher
80
0.946
0.597
3.965
21.66
0.7923
OCR Teacher
80
0.656
0.924
3.849
21.66
0.7957
PickScore Teacher
80
0.744
0.629
4.166
23.30
0.7756
Flow-OPD
80
0.905
0.885
4.268
22.66
0.8788
DiffusionOPD
80
0.945
0.930
4.113
23.68
0.9022
Table 1: Comparison of text-to-image scores across methods, at forty and at four sampling steps, for both base architectures evaluated. The lower per-architecture block reports the same models after step distillation.
Figure 2: Qualitative comparison result from benchmark.
Figure 3: Qualitative comparison result from generated prompt.
Figure 4: Text that models render into the image. The left panel shows generations on the No text called for prompt group; the right panel counts the generations whose detected on-image text reproduces a word from the prompt, on that same group and on Text content specified , drawn on separate scales.
Rollout grid
Re-noising source
GenEval ↑
OCR ↑
DeQA ↑
PickScore ↑
Average ↑
Fixed
Segment endpoint
0.7230
0.4499
4.003
22.42
0.7090
Randomized
Segment endpoint
0.7523
0.5776
4.129
22.46
0.7548
Randomized
Interpolated
0.8198
0.7968
4.076
22.47
0.8240
Table 2: Ablation of the three supervision-point constructions. Best results are in bold.
Teacher selection
GenEval ↑
OCR ↑
DeQA ↑
PickScore ↑
Average ↑
Source routing
0.9442
0.8986
4.170
22.98
0.8901
Random teacher
0.9241
0.7388
4.203
22.37
0.8410
Uniform teacher average
0.9292
0.6738
4.179
22.18
0.8230
Table 3: Ablation of how each mini-batch selects its teacher, reported as full-benchmark scores. Best results are in bold.
Table 8Table 9
Appendix figures & tables18 assets
Supplementary material from the paper’s appendix.
Appendix
Base Model
Model
NFEs
CLIPScore ↑
HPSv2.1 ↑
Aesthetic ↑
ImageReward ↑
UnifiedReward ↑
SD3.5-Medium
Base
80
1.000
0.289
5.552
1.081
3.701
GenEval Teacher
80
1.020
0.277
5.541
1.222
3.783
OCR Teacher
80
0.994
0.274
5.390
1.015
3.579
PickScore Teacher
80
1.004
0.325
5.954
1.396
3.964
DeQA Teacher
80
1.011
0.301
5.622
1.293
3.903
DiffusionOPD
80
1.033
0.304
5.676
1.288
3.874
Appendix
Table 8: Comparison of additional aligned quality and preference measures across methods, for both base architectures evaluated.
Initialization
Teachers
GenEval ↑
OCR ↑
DeQA ↑
PickScore ↑
Average ↑
Averaged-specialist merge
Flow-OPD
0.885
0.874
4.073
22.61
0.8610
DiffusionOPD
0.903
0.896
4.164
22.89
0.8779
SFT-augmented merge
Flow-OPD
0.932
0.886
4.077
22.75
0.8772
DiffusionOPD
0.944
0.899
4.170
22.98
0.8901
Appendix
Table 9: Ablation of two initialization recipes crossed with two multi-teacher on-policy distillation teacher families, reported as full-benchmark scores.
Student configuration
Source ratio
GenEval ↑
OCR ↑
DeQA ↑
PickScore ↑
Average ↑
Fresh-LoRA student
1:1:1
0.752
0.5776
4.129
22.46
0.7548
1:2:1
0.762
0.683
4.145
22.42
0.7839
Merged-init. student
1:1:1
0.944
0.899
4.170
22.98
0.8901
1:2:1
0.869
0.815
4.095
22.53
0.8427
Appendix
Table 10: Ablation of the training-source sampling ratio (GenEval:OCR:PickScore) within each of two student configurations.
Model
Text rendered
Prompt-word recurrence
PickScore ↑
SD3.5-Medium base
166
21
0.838
SD3.5-Medium SFT (GenEval)
283
40
0.838
SD3.5-Medium DiffusionOPD
699
218
0.912
SD3.5-Medium Flow-OPD
238
66
0.872
FLUX.1-dev base
177
8
0.844
FLUX.1-dev DiffusionOPD
266
44
0.907
Appendix
Table 11: Comparison of text rendered, the count of generations with any detected on-image text, and prompt-word recurrence, the count among those whose detected text reproduces a word from the prompt, across models, on PickScore benchmark prompts that do not call for rendered text, together with each model’s PickScore.
No text expected
Text-bearing object named
Scene-typical text
Specified text content
Model
Text rendered
Prompt-word recurrence
Pick- Score ↑
Text rendered
Prompt-word recurrence
Pick- Score ↑
Text rendered
Prompt-word recurrence
Pick- Score ↑
Text rendered
Prompt-word recurrence
Pick- Score ↑
SD3.5-Medium base
166
21
0.838
149
54
0.821
74
2
0.842
234
126
0.860
SD3.5-Medium SFT (GenEval)
283
40
0.838
162
78
0.820
103
12
0.840
247
137
0.872
SD3.5-Medium DiffusionOPD
699
218
0.912
273
178
0.889
198
44
0.909
247
159
0.928
SD3.5-Medium Flow-OPD
238
66
0.872
159
77
0.850
80
11
0.870
255
155
0.895
FLUX.1-dev base
177
8
0.844
66
12
0.823
84
3
0.848
225
115
0.871
Appendix
Table 12: Comparison of text rendered, the count of generations with any detected on-image text, and prompt-word recurrence, the count among those whose detected text reproduces a word from the prompt, together with PickScore, across models, for four prompt groups defined by whether the described scene calls for rendered text.
Figure 5: Count of generated images whose detected text reproduces a word drawn from the prompt, for seven models on the prompt group that calls for no rendered text. Colour marks the kind of post-training each model received.
Prompt-word recurrence
Model
Slogan register
Descriptive, nothing to write on
OCR ↑
GenEval ↑
Base + TDM
0.168
0.000
0.335
0.653
DiffusionOPD + TDM
0.782
0.000
0.833
0.838
Flow-OPD + TDM
0.675
0.000
0.751
0.875
SFT + TDM
0.610
0.000
0.725
0.932
StepX-OPD , w/o precise teacher
0.820
0.000
0.899
0.944
Appendix
Table 13: Comparison of prompt-word recurrence, the fraction of generations whose detected on-image text reproduces words from the prompt, across four-step models under two prompt registers, alongside each model’s OCR and GenEval scores from the full evaluation.
Figure 6: Prompt-word recurrence against PickScore, resolved to individual prompts. Each point is one prompt on one model: the vertical coordinate is the fraction of that prompt’s three seeds whose detected on-image text reproduces a word from the prompt, and the horizontal coordinate is that prompt’s PickScore averaged over the same three seeds. Panels and colours are the three SD3.5-Medium models of Fig. , on the prompt group that figure draws from. Points carry a small vertical jitter so that points sharing a vertical value are visible as separate marks.
Figure 7: Text seven SD3.5-Medium models render into the image, on seven of the prompts of the Fig. subset for which DiffusionOPD produces an image whose text reproduces a word from the prompt. Columns are prompts (1)–(7), printed in full below the block; rows are the untrained base, four single-reward experts, and two published multi-capability models. Fig. continues with prompts (8)–(14).
Figure 8: Text seven SD3.5-Medium models render into the image, continuing Fig. with prompts (8)–(14) of the same fourteen. Columns are those prompts, printed in full below the block; rows are the untrained base, four single-reward experts, and two published multi-capability models.
Figure 9: (a) A black apple and a green backpack on a sunlit wooden bench, shallow depth of field, warm afternoon light; (b) A brass diving helmet and a coil of blue rope on a weathered dock, sea spray, low sun behind; (c) Three wooden rowing boats moored on a still alpine lake at dawn, mountains mirrored in the water; (d) Three ripe pomegranates split open on a dark table, seeds glistening, dramatic side light, baroque still life.
Figure 10: (a) Four seagulls in flight against a stormy grey sky above cliffs, wind-torn spray below; (b) Five jellyfish drifting in dark water, translucent bells glowing blue, fine tentacles trailing; (c) Three drones hovering above a terraced rice field at sunrise, mist between the terraces; (d) Four crows lined up on a snowy fence rail, black against white, flat overcast light.
Figure 11: (a) Stacked wooden animal totem with bear at the bottom, eagle in the middle and fish on top, with weathered paint, forest bokeh and warm light; (b) A hummingbird hovering in front of a red trumpet flower with blurred green foliage behind, high-speed macro; (c) An old typewriter with a sheet of paper rolled in, the typed line reading “CHAPTER ONE”, dust motes in a slanting sunbeam; (d) A hand-stitched sampler framed on a wall reading “HOME SWEET HOME”, late afternoon light, faded floral wallpaper behind.
Figure 12: (a) A construction site hoarding painted with “COMING SOON”, scaffolding above, crane silhouette against an orange evening sky; (b) An enamel mug on a campfire stone printed with “WILDERNESS”, flames glowing, pine forest dark behind, blue hour; (c) A vast, sun-baked desert landscape with a lone highway stretching into the distance. A weathered road sign stands by the side, warning “Next Gas 100 Miles”, under a blazing sky; (d) A green sign that says “Very Deep Learning” and is at the edge of the Grand Canyon.
No text expected
No text called for
Initialization
Teachers
Precise teacher
Text rendered
Prompt-word recurrence
Text rendered
Prompt-word recurrence
SFT-augmented merge
DiffusionOPD
Off
304
94
96
20
On
298
84
92
19
Averaged-specialist merge
DiffusionOPD
Off
284
101
93
24
On
292
87
86
21
Averaged-specialist merge
Flow-GRPO
Off
190
71
60
17
Appendix
Table 14: Comparison of text rendered and prompt-word recurrence with precise-teacher supervision off and on, across three distillation configurations, on the prompt group with no text expected and on its subset with no text called for. Bold marks the lower prompt-word recurrence count within each configuration line.
Initialization
GenEval ↑
OCR ↑
DeQA ↑
PickScore ↑
Average ↑
Merge (3 teacher)
0.8709
0.8603
4.064
22.53
0.8527
Merge w/o OCR
0.8822
0.8253
4.078
22.62
0.8483
Merge w DEQA
0.8691
0.8346
4.086
22.58
0.8474
Appendix
Table 15: Ablation of soup-composition variants of the initialization, reported as full-benchmark scores.
vMSE query
GenEval ↑
OCR ↑
DeQA ↑
PickScore ↑
Average ↑
Continuous sampled τ
0.8664
0.7296
4.101
22.65
0.8219
Per-sample inference-grid point
0.8449
0.7123
4.098
22.57
0.8113
Appendix
Table 16: Comparison of two velocity-regression query designs, reported as full-benchmark scores.
Figure 13: Ablation of the weight on the low-noise velocity-regression term. The left panel gives the Average score and the right panel rescales each of its four components to its own range across the five settings.
On-policy distillation (OPD) has become an effective approach for consolidating multiple task-specialized image generation models into a single student. However, existing OPD methods optimize the student mainly to match the teacher's output velocity, making the teacher the upper limit of the optimization objective. While output-level supervision alone leaves the student's blockwise representation evolution underconstrained, which weakens the transfer of capabilities that must be progressively developed across layers. We propose STEP-OPD, an on-policy distillation framework for image generation that extends the student's learning target beyond the teacher and introduces explicit constraints on its internal representation evolution. Instead of treating the teacher as the final target, we use the velocity difference between each task-specific teacher and the shared base model as a direction for further learning and add a scaled version of this difference to the teacher velocity. In addition, we align the direction and magnitude of representation changes between the student and teacher, enabling the student to learn how representations are progressively transformed across network blocks. Experiments on compositional alignment, text rendering, and human preference show that our method consistently improves Standard OPD methods. In particular, it increases the GenEval score of DiffusionOPD from 0.927 to 0.961, while also improving OCR and all preference-based metrics. The resulting unified student surpasses the corresponding single-task teachers across all three capability groups, showing that output extrapolation enables beyond-teacher learning. And representation change alignment provides complementary guidance for the student's internal transformations.
Qingyan Wei, Guangzhao Li, Xiaobing Tu +5
Shanghai Jiao Tong University · Shanghai Innovation Institute · Alibaba Group
The landscape of high-performance image generation models is currently shifting from the inefficient multi-step ones to the efficient few-step counterparts (e.g, Z-Image-Turbo and FLUX.2-klein). However, these models present significant challenges for direct continuous supervised fine-tuning. For example, applying the commonly used fine-tuning technique would compromise their inherent few-step inference capability. To address this, we propose D-OPSD, a novel training paradigm for step-distilled diffusion models that enables on-policy learning during supervised fine-tuning. We first find that the modern diffusion models, where the LLM/VLM serves as the encoder, can inherit its encoder's in-context capabilities. This enables us to formulate the training as an on-policy self-distillation process. Specifically, during training, we make the model act as both the teacher and the student with different contexts, where the student is conditioned only on the text feature, while the teacher is conditioned on the multimodal feature of both the text prompt and the target image. Training minimizes the two predicted distributions over the student's own roll-outs. By optimizing on the model's own trajectory and under its own supervision, D-OPSD enables the model to learn new concepts, styles, etc., without sacrificing the original few-step capacity.
Dengyang Jiang, Xin Jin, Dongyang Liu +9
1The Hong Kong University of Science and Technology · 2Z-Image Team, Alibaba Group · 4The Chinese University of Hong Kong +1
Leading open text-to-image models often carry complementary strengths: one may lead on preference-aligned aesthetics while another follows compositional instructions more faithfully. However, differences in their autoencoders and noise schedules make it difficult to transfer these strengths across models. In this paper, we present Poly-OPD, a framework that can consolidate complementary strengths of heterogeneous teachers into a single compact flow-matching student. To bridge the incompatible latent spaces of different teachers, Poly-OPD performs on-policy distillation through a pixel bridge. Each student-generated image is re-encoded by a selected teacher's encoder and refined from a noise level matched by magnitude under the teacher's noise schedule. The resulting target is further matched to the student in frozen DINOv2 space, enabling supervision across incompatible latent spaces. To retain complementary capabilities without cross-teacher interference, Poly-OPD uses a gradient compatibility diagnostic to organize its adapters: attention LoRA modules are shared across teachers, whereas feed-forward adapters remain teacher-specific. During distillation, a gap-aware curriculum devotes more training to compositional categories where the student still falls short of the teacher. As each gap narrows, training shifts toward categories with larger remaining gaps. By distilling FLUX.1-dev and Z-Image into a 2.5B SD3.5-Medium student, Poly-OPD improves GenEval from 67.3 to 73.3, surpassing both larger teachers, and raises DrawBench HPSv3 from 9.34 to 11.35, consolidating both strengths within a switchable model.
Siming Fu, Haojun Xu, Ruizhe He +9
1Joy Future Academy · 2Zhejiang University · 3Beihang University