Self-Adversarial One Step Generation via Condition Shifting

Deyuan Liu; Peng Sun; Yansen Han; Zhenglin Cheng; Chuyan Chen; Tao Lin

cs.CVApr 14, 2026

Self-Adversarial One Step Generation via Condition Shifting

Authors: Deyuan Liu, Peng Sun, Yansen Han, Zhenglin Cheng, Chuyan Chen, Tao Lin

Organizations: Westlake University · Zhejiang University · Shanghai Innovation Institute · Peking University

Abstract

The push for efficient text to image synthesis has moved the field toward one step sampling, yet existing methods still face a three way tradeoff among fidelity, inference speed, and training efficiency. Approaches that rely on external discriminators can sharpen one step performance, but they often introduce training instability, high GPU memory overhead, and slow convergence, which complicates scaling and parameter efficient tuning. In contrast, regression based distillation and consistency objectives are easier to optimize, but they typically lose fine details when constrained to a single step. We present APEX, built on a key theoretical insight: adversarial correction signals can be extracted endogenously from a flow model through condition shifting. Using a transformation creates a shifted condition branch whose velocity field serves as an independent estimator of the model's current generation distribution, yielding a gradient that is provably GAN aligned, replacing the sample dependent discriminator terms that cause gradient vanishing. This discriminator free design is architecture preserving, making APEX a plug and play framework compatible with both full parameter and LoRA based tuning. Empirically, our 0.6B model surpasses FLUX-Schnell 12B (20 $\times$ more parameters) in one step quality. With LoRA tuning on Qwen-Image 20B, APEX reaches a GenEval score of 0.89 at NFE=1 in 6 hours, surpassing the original 50-step teacher (0.87) and providing a 15.33 $\times$ inference speedup. Code is available https://github.com/LINs-lab/APEX.

Figures & tables

Figure 1: An overview of generated images.

Figure 2: Qualitative Analysis between APEX and existing methods under different NFEs.

Methods		NFEs	Throughput	Latency	Params	FID $\downarrow$	CLIP $\uparrow$	GenEval $\uparrow$
Methods		NFEs	(samples/s)	(s)	(B)	FID $\downarrow$	CLIP $\uparrow$	GenEval $\uparrow$
Few Step Distillation Models	SDXL-LCM Luo et al. (2023)	2	2.89	0.40	0.9	18.11	27.51	0.44
	PixArt-LCM Chen et al. (2024c)	2	3.52	0.31	0.6	10.33	27.24	0.42
	SD3.5-Turbo Esser et al. (2024)	2	1.61	0.68	8.0	51.47	25.59	0.53
	PCM Wang et al. (2024a) †	2	2.62	0.56	0.9	14.70	27.66	0.55
	SDXL-DMD2 Yin et al. (2024a) †	2	2.89	0.40	0.9	7.61	28.87	0.58

Table 1: System level comparison of efficiency and quality. Speeds are on a single A100 ( BF16 ). Throughput is samples/s (batch=10); latency is seconds (batch=1). GenEval is the primary quality metric; FID/CLIP are reported for completeness. The best and second best entries are highlighted. † indicates methods requiring distinct models per NFE. Notation: Blue =full tuning; Red =LoRA; X.B =trainable params (B);

r

=LoRA rank.

Model	Single	Two	Counting	Colors	Position	Attribute	Overall $\uparrow$
Model	Object	Object	Counting	Colors	Position	Binding	Overall $\uparrow$
Show-o ( Xie et al., 2024b )	0.95	0.52	0.49	0.82	0.11	0.28	0.53
Emu3-Gen ( Wang et al., 2024b )	0.98	0.71	0.34	0.81	0.17	0.21	0.54
PixArt- $\alpha$ ( Chen et al., 2024d )	0.98	0.50	0.44	0.80	0.08	0.07	0.48
SD3 Medium ( Esser et al., 2024 )	0.98	0.74	0.63	0.67	0.34	0.36	0.62
FLUX.1 [Dev] ( BlackForest, 2024 )	0.98	0.81	0.74	0.79	0.22	0.45	0.66

Table 2: Quantitative Evaluation results on GenEval.

Model	Global	Entity	Attribute	Relation	Other	Overall $\uparrow$
SD v1.5 ( Rombach et al., 2022 )	74.63	74.23	75.39	73.49	67.81	63.18
PixArt- $\alpha$ ( Chen et al., 2024d )	74.97	79.32	78.60	82.57	76.96	71.11
Lumina-Next ( Zhuo et al., 2024 )	82.82	88.65	86.44	80.53	81.82	74.63
SDXL ( Podell et al., 2023 )	83.27	82.43	80.91	86.76	80.41	74.65
Hunyuan-DiT ( Li et al., 2024b )	84.59	80.59	88.01	74.36	86.41	78.87
Janus ( Wu et al., 2025b )	82.33	87.38	87.70	85.46	86.41	79.68

Table 3: Quantitative evaluation results on DPGBench.

Model	Cultural	Time	Space	Biology	Physics	Chemistry	Overall $\uparrow$
SD v1.5 ( Rombach et al., 2022 )	0.34	0.35	0.32	0.28	0.29	0.21	0.32
SDXL ( Podell et al., 2023 )	0.43	0.48	0.47	0.44	0.45	0.27	0.43
SD3.5-Large Esser et al. (2024)	0.44	0.50	0.58	0.44	0.52	0.31	0.46
PixArt- $\alpha$ ( Chen et al., 2024d )	0.45	0.50	0.48	0.49	0.56	0.34	0.47
Playground-v2.5 ( Li et al., 2024a )	0.49	0.58	0.55	0.43	0.48	0.33	0.49
FLUX.1 [Dev] ( BlackForest, 2024 )	0.48	0.58	0.62	0.42	0.51	0.35	0.50

Table 4: Quantitative evaluation results on WISE.

Model	ShareGPT-4o			Blip-3o
Model	2Ksteps	8Ksteps	10Ksteps	2Ksteps	8Ksteps	10Ksteps
APEX 0.6B	0.37	0.67	0.73	0.71	0.77	0.81
APEX 1.6B	0.36	0.70	0.73	0.27	0.78	0.83
	0.4Ksteps	1Ksteps	2Ksteps	0.4Ksteps	1Ksteps	2Ksteps
APEX 20B ( r=32 )	0.19	0.33	0.62	0.83	0.84	0.83
APEX 20B ( r=64 )	0.21	0.35	0.61	0.73	0.85	0.84

Table 5: Effect of training data and steps on GenEval Overall (NFE=1). We compare ShareGPT-4o and BLIP-3o across training steps for APEX 0.6B/1.6B, and LoRA tuned Qwen-Image 20B with ranks r=32 / r=64 . All runs use global batch size 64 .

Weighting Ratio ( $\lambda_{p}:\lambda_{e}$ )	APEX 0.6B	APEX 1.6B	APEX 20B ( $r{=}32$ )
1.0 : 0.0 ( $\mathcal{L}_{\text{fake}}$ Only)	0.32	0.35	0.42
0.0 : 1.0 ( $\mathcal{L}_{\text{mix}}$ Only)	0.63	0.66	0.69
1.0 : 0.5	0.72	0.71	0.81
1.0 : 1.0 (Ours)	0.77	0.76	0.83
1.0 : 2.0	0.74	0.75	0.82

Table 6: Ablation on the weights of

\mathcal{L}_{\text{fake}}

vs.

\mathcal{L}_{\text{mix}}

. We report GenEval Overall (NFE=1) for different weighting ratios (

\lambda_{p}{:}\lambda_{e}

). The dataset is BLIP-3o. Training steps are 8K for 0.6B/1.6B models and 0.4K for the 20B (LoRA) model. Best per model in bold .

$a$ $\backslash$ $b$	0.0	0.1	1.0	10.0
$-1.0$	0.76	0.73	0.74	0.74
$-0.5$	0.75	0.79	0.81	0.70
$\phantom{-}0.5$	0.29	0.37	0.30	0.73

Table 7: Effect of condition-shifting hyperparameters on GenEval Overall (NFE=1). Moderate negative scaling (

a\in\{-1.0,-0.5\}

) yields the most robust gains.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3: Qualitative Comparison of 512x512 in APEX 20B LoRA for NFE=1.

Figure 4: Qualitative Comparison of 512x512 in APEX 20B LoRA for NFE=1.

Figure 5: Qualitative Comparison of 512x512 in APEX 20B LoRA for NFE=1.

Figure 6: Qualitative Comparison of 512x512 in APEX 0.6B LoRA for NFE=1.

Figure 7: Qualitative Comparison of 512x512 in APEX 20B LoRA for NFE=1.

Figure 8: Qualitative Comparison of 512x512 in APEX 20B LoRA for NFE=1.

Figure 9: Qualitative Comparison of 512x512 in APEX 20B LoRA for NFE=1.

Figure 10: Qualitative Comparison of 512x512 in APEX 20B Full Parameter Tuning for NFE=1.

Figure 11: Qualitative Comparison of 512x512 in APEX 20B Full Parameter Tuning for NFE=1.

Figure 12: Qualitative Comparison of 512x512 in APEX 20B Full Parameter Tuning for NFE=1.

Figure 13: Qualitative Comparison of 512x512 in Qwen-Image Lightning LoRA for NFE=1.

Figure 14: Qualitative Comparison of 512x512 in Qwen-Image Lightning LoRA for NFE=1.

Figure 15: Qualitative Comparison of 512x512 in Qwen-Image Lightning LoRA for NFE=1.

Figure 16: Qualitative Comparison of 512x512 in 20B Full Parameter Tuning of APEX methods and Synthetic dataset from NFE=1 to NFE=20.

Figure 17: Qualitative Comparison of 512x512 in 20B Full Parameter Tuning of APEX methods and BLIP-3o dataset from NFE=1 to NFE=20.

Figure 18: Qualitative Comparison of 512x512 in 20B Full Parameter Tuning of sCM methods and BLIP-3o dataset from NFE=1 to NFE=20.

Figure 19: Qualitative Comparison of 512x512 in 20B Full Parameter Tuning of CTM methods and BLIP-3o dataset from NFE=1 to NFE=20.

Figure 20: Qualitative Comparison of 512x512 in 20B Full Parameter Tuning of MeanFlow methods and BLIP-3o dataset from NFE=1 to NFE=20.

Explore similar work

Denoising, Fast and Slow: Difficulty-Aware Adaptive Sampling for Image Generation
Apr 21, 2026Johannes Schusterbauer, Ming Gui, Yusong Li +3Diffusion Model Sampling Diffusion Sampling
Setting-Matched and Semantics-Scaled Benchmarking of One-Step Generative Models Against Multistep Diffusion and Flow Models
Mar 15, 2026Advaith Ravishankar, Serena Liu, Mingyang Wang +11One-Step Generative Modeling Image Generation Evaluation
Harvard AI and Robotics Lab, Harvard University
Extending One-Step Image Generation from Class Labels to Text via Discriminative Text Representation
Apr 20, 2026Chenxi Zhao, Chen Zhu, Xiaokun Feng +6Flow Matching Representation Learning
College of Computer Science, Nankai University · AMAP, Alibaba Group

Apr 21, 2026J/K move · Enter open · S save

Denoising, Fast and Slow: Difficulty-Aware Adaptive Sampling for Image Generation

Johannes Schusterbauer, Ming Gui, Yusong Li, Pingchuan Ma+2

Diffusion Model Sampling Diffusion Sampling Diffusion-Based Image Generation Diffusion Timestep Scheduling Diffusion Model Inference Acceleration

Open paper AskPDF arXiv

Diffusion- and flow-based models usually allocate compute uniformly across space, updating all patches with the same timestep and number of function evaluations. While convenient, this ignores the heterogeneity of natural images: some regions are easy to denoise, whereas others benefit from more refinement or additional context. Motivated by this, we explore patch-level noise scales for image synthesis. We find that naively varying timesteps across image tokens performs poorly, as it exposes the model to overly informative training states that do not occur at inference. We therefore introduce a timestep sampler that explicitly controls the maximum patch-level information available during training, and show that moving from global to patch-level timesteps already improves image generation over standard baselines. By further augmenting the model with a lightweight per-patch difficulty head, we enable adaptive samplers that allocate compute dynamically where it is most needed. Combined with noise levels varying over both space and diffusion time, this yields Patch Forcing (PF), a framework that advances easier regions earlier so they can provide context for harder ones. PF achieves superior results on class-conditional ImageNet, remains orthogonal to representation alignment and guidance methods, and scales to text-to-image synthesis. Our results suggest that patch-level denoising schedules provide a promising foundation for adaptive image generation.

Abstract

Explore similar work

Denoising, Fast and Slow: Difficulty-Aware Adaptive Sampling for Image Generation

Setting-Matched and Semantics-Scaled Benchmarking of One-Step Generative Models Against Multistep Diffusion and Flow Models

Extending One-Step Image Generation from Class Labels to Text via Discriminative Text Representation