In this work, we analyze the optimization dynamics of generative fine-tuning. We observe that under the Flow Matching framework, the standard MSE objective can be formulated as a Quadratic Form governed by a dynamically evolving Neural Tangent Kernel (NTK). This geometric perspective reveals a latent Data Interaction Matrix, where diagonal terms represent independent sample learning and off-diagonal terms encode residual correlation between heterogeneous features. Although standard training implicitly optimizes these cross-term interferences, it does so without explicit control; moreover, the prevailing data-homogeneity assumption may constrain the model's effective capacity. Motivated by this insight, we propose Semantic Granularity Alignment (SGA), using Text-to-Image synthesis as a testbed. SGA engineers targeted interventions in the vector residual field to mitigate gradient conflicts. Evaluations across DiT and U-Net architectures confirm that SGA advances the efficiency-quality trade-off by accelerating convergence and improving structural integrity.
Figures & tables
Figure 1
Figure 1: Qualitative comparison between the Baseline and our SGA method across three GDA domains. For each row, the reference images (left) define the target domain; the Baseline (middle) fails to preserve domain-specific attributes, while SGA (right) faithfully captures the target domain characteristics.
Figure 2: Structure of the Data Interference Matrix Ω . Diagonal entries (blue) represent independent learning within each sub-manifold. Off-diagonal entries encode cross-scale interactions: constructive (synergy, green) or destructive (conflict, red).
Figure 3: The optimization landscape LFM=α⊤Ωα of CFM fine-tuning. Path 1 (red): Underfitting Region. Path 3 (blue): OOD Region. Only Path 2 (green) navigates the narrow Goal Region.
Figure 4: Failure analysis of standard fine-tuning. All outputs exhibit dominant prior characteristics regardless of target domain—a prior capture phenomenon that traps the model in the Underfitting Region of Fig. 3 .
Figure 5: Overview of the Hierarchical Semantic Decomposition (H-SD) pipeline. Raw images are parsed by a general-purpose detector ( e.g . , multimodal models, YOLO, or Grounding DINO) into three granularities ( YMacro , YMeso , YMicro ), filtered via IoU-based redundancy elimination, and post-processed into aspect-ratio-preserving resolution buckets.
Figure 6: Optimization trajectories on the H-SD-augmented loss landscape. H-SD + Standard Opt. (orange): after signal amplification via H-SD, vanilla CFM training still suffers from scale-biased gradient updates, causing persistent oscillation and drift toward the OOD boundary. SGA (blue): by adding Tuple-wise co-sampling and Scale-Adaptive Modulation, SGA stabilizes the descent and converges to the Goal Region. The residual correlation sign ⟨Δξ,Δη⟩ determines whether cross-scale interactions are constructive ( >0 , stable) or destructive ( <0 , oscillation).
Figure 7: Qualitative Comparison (FLUX-only). Domains δ , ε , and ζ , evaluated exclusively on FLUX (portrait layout).
Figure 8: Qualitative Comparison (Cross-Architecture). Domains α – γ , evaluated on both SDXL and FLUX. Reference images (left) define each target domain.
Figure 9: Quantitative Evaluation — LLM Judge (GPT 5.2). Aggregate rank distribution (left) and per-dataset 1st-place rate heatmap (right). SGA variants achieve favourable rankings across both architectures and all stylistic domains.
Figure 10: Quantitative Evaluation — Human Evaluation. Matching layout to Fig. 9 . Human rankings corroborate the automated assessment, with SGA variants accumulating the majority of top rankings.
FLUX (6 datasets)
SDXL (3 datasets)
Metric
Baseline
SGA
Baseline
SGA
CLIP-I
0.686
0.701
↑
0.715
0.717
↑
CLIP-T
0.248
0.253
↑
0.296
0.298
↑
DINO-I
0.425
0.427
↑
0.492
0.539
↑
Table 1: Calibration Metrics (averaged across datasets). Higher is better for all metrics. ↑ denotes improvement of SGA over the Baseline.
FLUX
SDXL
Variant
Human
LLM
Human
LLM
Full SGA
54
55
48
61
w/o Tuple-wise Opt.
27
25
16
14
w/o Scale-Adaptive Mod.
19
20
36
25
Table 2: Ablation Study: 1st-place rate (%) across evaluator types. Higher is better.
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Hyperparameter
Value
Backbone
FLUX.1-dev [ 6 ]
Adaptation Method
DoRA [ 37 ]
LoRA Rank
32
LoRA Alpha
16
Batch Size
8
Learning Rate
1×10−4
Appendix
Table 3: FLUX training hyperparameters. All domains use the same configuration.
Domain
N1 (hours)
N2 (hours)
α
14
21
β
12
18
γ
14
21
δ
9
13.5
ε
8
12
ζ
14
21
Appendix
Table 4: Estimated GPU hours per domain (FLUX). N1 and N2 denote the two evaluation checkpoints. Actual models used for evaluation are the closest saved checkpoints to the listed times.
Hyperparameter
Value
Backbone
Animagine XL 3.1 [ 47 ]
Optimizer
Lion
U-Net Learning Rate
3×10−5
Text Encoder Learning Rate
3×10−6
Batch Size
2
Linear layers
Appendix
Table 5: Animagine XL 3.1 (SDXL) training hyperparameters.
Domain
N1 (hours)
N2 (hours)
α
2
3
β
2
3
γ
1.5
2.5
Appendix
Table 6: Estimated GPU hours per domain (Animagine XL 3.1). The same checkpoint selection protocol as FLUX applies: actual models used are the closest saved checkpoints to the listed times, with deviation not exceeding 30 minutes.
Figure 11: Reference images defining the target domain for all ablation experiments.
Figure 12: FLUX ablation (Group 1). Full SGA maintains domain fidelity; removing Scale-Adaptive Modulation introduces subtle but visible stylistic drift.
Figure 13: FLUX ablation (Group 2). The w/o Scale-Adaptive Modulation variant exhibits increased OOD tendencies in fine-grained detail rendering.
Figure 14: FLUX ablation (Group 3). All three variants produce structurally coherent outputs—consistent with FLUX’s global attention mechanism—but the Full SGA variant best preserves the target domain’s micro-level characteristics.
Figure 19: SDXL ablation (Group 2). Full SGA best preserves limb coherence, consistent with CNN-based models requiring explicit structural regularization.
Figure 20: SDXL ablation (Group 3). SGA removal consistently causes the most significant anatomical degradation, supporting its critical role in compensating for CNN’s limited global awareness.
Architecture
Foundation Models
Community Derivatives
U-Net
SD 1.5 / 2.1 [ 51 ] ,
Pony Diffusion V6 [ 48 ] , Animagine XL [ 8 ] ,
SDXL [ 47 ]
NoobAI-XL [ 32 ] , Illustrious XL [ 43 ]
MM-DiT
PixArt- α [ 11 ] , SD 3.5 [ 17 ] ,
—
AuraFlow [ 18 ] , FLUX.1 [ 6 ] , FLUX.2 [ 7 ]
Appendix
Table 7: Open-source diffusion model ecosystem. Representative foundation models and their community-driven derivatives, organized by backbone architecture.
Existing Flow Matching (FM) text-to-image models suffer from two critical bottlenecks under multi-task alignment: the reward sparsity induced by scalar-valued rewards, and the gradient interference arising from jointly optimizing heterogeneous objectives, which together give rise to a 'seesaw effect' of competing metrics and pervasive reward hacking. Inspired by the success of On-Policy Distillation (OPD) in the large language model community, we propose Flow-OPD, the first unified post-training framework that integrates on-policy distillation into Flow Matching models. Flow-OPD adopts a two-stage alignment strategy: it first cultivates domain-specialized teacher models via single-reward GRPO fine-tuning, allowing each expert to reach its performance ceiling in isolation; it then establishes a robust initial policy through a Flow-based Cold-Start scheme and seamlessly consolidates heterogeneous expertise into a single student via a three-step orchestration of on-policy sampling, task-routing labeling, and dense trajectory-level supervision. We further introduce Manifold Anchor Regularization (MAR), which leverages a task-agnostic teacher to provide full-data supervision that anchors generation to a high-quality manifold, effectively mitigating the aesthetic degradation commonly observed in purely RL-driven alignment. Built upon Stable Diffusion 3.5 Medium, Flow-OPD raises the GenEval score from 63 to 92 and the OCR accuracy from 59 to 94, yielding an overall improvement of roughly 10 points over vanilla GRPO, while preserving image fidelity and human-preference alignment and exhibiting an emergent 'teacher-surpassing' effect. These results establish Flow-OPD as a scalable alignment paradigm for building generalist text-to-image models. The codes and weights will be released in: https://github.com/CostaliyA/Flow-OPD .
Zhen Fang, Wenxuan Huang, Yu Zeng +8
University of Science and Technology of China · University of California, Los Angeles · 3The Chinese University of Hong Kong +1
This paper focuses on the alignment of flow matching models with human preferences. A promising way is fine-tuning by directly backpropagating reward gradients through the differentiable generation process of flow matching. However, backpropagating through long trajectories results in prohibitive memory costs and gradient explosion. Therefore, direct-gradient methods struggle to update early generation steps, which are crucial for determining the global structure of the final image. To address this issue, we introduce LeapAlign, a fine-tuning method that reduces computational cost and enables direct gradient propagation from reward to early generation steps. Specifically, we shorten the long trajectory into only two steps by designing two consecutive leaps, each skipping multiple ODE sampling steps and predicting future latents in a single step. By randomizing the start and end timesteps of the leaps, LeapAlign leads to efficient and stable model updates at any generation step. To better use such shortened trajectories, we assign higher training weights to those that are more consistent with the long generation path. To further enhance gradient stability, we reduce the weights of gradient terms with large magnitude, instead of completely removing them as done in previous works. When fine-tuning the Flux model, LeapAlign consistently outperforms state-of-the-art GRPO-based and direct-gradient methods across various metrics, achieving superior image quality and image-text alignment.
Zhanhao Liang, Tao Yang, Jie Wu +2
1The Australian National University · 2ByteDance Seed
Despite the success of foundation models in image synthesis, achieving fine-grained spatial control remains a significant challenge, typically requiring auxiliary structural adapters or complex inference-time steering. While modern vision-language (VLM) text encoders possess the latent capacity to represent spatial coordinates, generative models often fail to leverage this information. We demonstrate that the signal-to-noise ratio (SNR) distribution is a critical determinant of spatial grounding in flow-matching models. Using a novel synthetic benchmark of geometric primitives, we isolate structural mechanics from textural complexity inherent in natural imagery and introduce a rigorous color-aware layout metric to quantify spatial fidelity. We show that training on a centered SNR distribution results in poor layout adherence (25.9%). Instead, we identify high-noise regimes as a "critical learning phase" for spatial grounding. Shifting the training focus toward these stages increases layout adherence to 80.8%, effectively reaching the 83.6% reconstruction ceiling of the underlying variational autoencoder. We find that the necessity for this shift scales with resolution and aspect ratio variance, and that the conservative noise schedules employed by models like FLUX.1 are insufficient for precise layout adherence. Finally, we show that these findings translate to real-world imagery via the COCO dataset (1.0 vs. 57.3 Average Precision), where state-of-the-art layout adherence is unlocked through distribution tuning alone, without architectural modifications.
Martin Thißen, Elke Hergenröther
Department of Computer Science, Darmstadt University of Applied Sciences, Darmstadt, Germany