Diffusion Probe: Generated Image Result Prediction Using CNN Probes

Benlei Cui; Bukun Huang; Zhizeng Ye; Xuemei Dong; Tuo Chen; Hui Xue; Dingkang Yang; Longtao Huang

cs.CVFeb 27, 2026

Diffusion Probe: Generated Image Result Prediction Using CNN Probes

Authors: Benlei Cui, Bukun Huang, Zhizeng Ye, Xuemei Dong, Tuo Chen, Hui Xue, Dingkang Yang, Longtao Huang,

Organizations: Yuvion Team, Alibaba Group · Laboratory for Statistical Monitoring and Intelligent Governance of Common Prosperity, School of Statistics and Data Science, Zhejiang Gongshang University · Southeast University · College of Intelligent Robotics and Advanced Manufacturing, Fudan University ·

Abstract

Text-to-image (T2I) diffusion models lack an efficient mechanism for early quality assessment, leading to costly trial-and-error in multi-generation scenarios such as prompt iteration, agent-based generation, and flow-grpo. We reveal a strong correlation between early diffusion cross-attention distributions and final image quality. Based on this finding, we introduce Diffusion Probe, a framework that leverages internal cross-attention maps as predictive signals. We design a lightweight predictor that maps statistical properties of early-stage cross-attention extracted from initial denoising steps to the final image's overall quality. This enables accurate forecasting of image quality across diverse evaluation metrics long before full synthesis is complete. We validate Diffusion Probe across a wide range of settings. On multiple T2I models, across early denoising windows, resolutions, and quality metrics, it achieves strong correlation (PCC > 0.7) and high classification performance (AUC-ROC > 0.9). Its reliability translates into practical gains. By enabling early quality-aware decisions in workflows such as prompt optimization, seed selection, and accelerated RL training, the probe supports more targeted sampling and avoids computation on low-potential generations. This reduces computational overhead while improving final output quality. Diffusion Probe is model-agnostic, efficient, and broadly applicable, offering a practical solution for improving T2I generation efficiency through early quality prediction. Code: https://github.com/Alibaba-VELLDEPTH/DiffusionProbe

Figures & tables

Figure 1: Illustration of early cross-attention dispersion. Here, we present the prompt, the corresponding four cross-attention activation maps in the early denoising stage, and the final generated image. Compared to other tokens, the cross-attention activation maps of the “bird” token shows significant sparsity in spatial distribution.

Figure 2: Overview of the Diffusion Probe framework. Our framework takes as input the early-stage cross-attention feature maps (derived from the CrossAttn module at a probed timestep $t$ ) and the TimeStep Embedding . A lightweight network processes these inputs, ultimately outputting a quality score prediction for the final generated image ( $x_{0}$ ). This predicted score is learned to align with a specified ground-truth Metric (e.g., aesthetic, semantic coherence) evaluated on the fully synthesized image. The Diffusion Probe then serves as a versatile tool to enable various downstream applications, such as Prompt Optimization, Seed Selection, and Efficient-GRPO training.

Figure 3: Visualization of the Cross-Attention Map for the token ”cat” within a FluxTransformerBlock. The image illustrates the spatial attention distribution for the text token ”cat” generated from the prompt “A cat holding a sign that says hello world”. Regions with intense red patterns indicate high attention scores, demonstrating where the model’s focus is directed in response to the specified token.

Figure 4: Early-stage cross-attention maps reveal object rendering fidelity. (Top) For the prompt “Woman carrying a bunch of bananas on top of her hat”, the model successfully renders all objects, resulting in sharp, focused attention maps for each token. (Bottom) In contrast, for “A child … surrounded by … building blocks in a playroom”, the model fails to generate the building blocks , and the corresponding attention map becomes highly diffuse. This demonstrates that attention statistics serve as a reliable early indicator of object-level generation success or failure (maps extracted at step $t=5$ , layer 19, as detailed in Figure 3 ).

Base Model	Resolution	Step	SRCC $\uparrow$	AUC-ROC $\uparrow$	KTC $\uparrow$	PCC $\uparrow$
SDXL [ 25 ]	1024 $\times$ 1024	1	0.49	0.53	0.35	0.48
		5	0.73	0.86	0.57	0.72
		10	0.76	0.89	0.61	0.75
		15	0.75	0.89	0.60	0.74
FLUX [ 18 ]	1024 $\times$ 1024	1	0.52	0.62	0.38	0.50
		5	0.76	0.88	0.60	0.75

Table 1: Diffusion Probe’s Predictive Accuracy for External Image Quality Metrics across Diffusion Steps (at 1024

\times

1024 Resolution). This table evaluates the predictive accuracy of our Diffusion Probe’s internal scores in aligning with established external image quality metrics. We quantify prediction performance using four distinct rank-based and correlation metrics. The Step column indicates the diffusion step from which the probe’s attention features were extracted for prediction. Higher values indicate superior predictive alignment.

Model	Task	Method	CLIP Score $\uparrow$	ImageReward $\uparrow$	Aesthetic Score $\uparrow$
SDXL	Prompt Opt.	Baseline	28.31	0.71	5.13
		+ Diffusion Probe (Ours)	30.24	0.72	5.29
		+ LLM	30.80	0.73	5.34
	Seed Sel.	Baseline (Random)	28.31	0.71	5.13
	Seed Sel.	+ Diffusion Probe (Ours)	30.14	0.82	5.43
FLUX	Prompt Opt.	Baseline	31.37	1.02	5.67

Table 2: Performance comparison of Diffusion Probe in Prompt Optimization and Seed Selection tasks on SDXL and FLUX models. All metrics are reported as higher is better (

\uparrow

Figure 5: Comparison of the PickScore during training steps for our method (”Ours”) versus the baseline method (”Origin”) applied to Flow-GRPO. The plot demonstrates that our approach enhances the stability and convergence speed of the training process, as evidenced by the smoother fluctuations and faster rise in the PickScore across training steps. This indicates a more consistent and efficient learning process when using our method.

Figure 6: Diffusion Probe as an effective filter for aesthetically poor generations. By leveraging training on image quality labels, the probe accurately flags images with common defects (e.g., distorted anatomy, poor composition), assigning them low scores.

Task	Workflow	Cost Breakdown	Total FLOPS (T)	Total Latency (s)
Reference	Single Generation	1 $\times$ Full Gen.	1877.56	14.70
Reference	Single Gen. + 1 $\times$ Pred.	1 $\times$ Full Gen. + 1 $\times$ Pred.	1877.57	14.75
Seed Selection	Naive Oversampling	10 $\times$ Full Gen.	18775.60	147.00
Seed Selection	Diffusion Probe Guided (Ours)	1 $\times$ Full Gen. + 10 $\times$ Pred.	5280.43	42.62
Prompt Optimization	Naive Comparison	4 $\times$ Full Gen.	7,510.25	58.00
Prompt Optimization	Diffusion Probe Guided (Ours)	1 $\times$ Full Gen. + 4 $\times$ Pred.	3026.42	28.29

Table 3: Computational Cost Analysis. This table compares the computational cost of naive brute-force workflows against our Diffusion Probe-guided approach. ’Single Generation’ serves as the baseline, representing the cost of one complete image generation (14.70s), while the subsequent row highlights the negligible overhead of a single probe prediction (+0.05s). We evaluate two scenarios: a 10-candidate Seed Selection and a 4-candidate Prompt Optimization , demonstrating significant savings in both latency and FLOPS.

Base Model	Resolution	Step	SRCC $\uparrow$	AUC-ROC $\uparrow$	KTC $\uparrow$	PCC $\uparrow$
FLUX	1024 $\times$ 1024	3	0.69	0.83	0.53	0.68
		4	0.75	0.88	0.60	0.74
		6	0.77	0.90	0.62	0.70
		7	0.74	0.87	0.58	0.72
	512 $\times$ 512	3	0.61	0.78	0.46	0.60
		4	0.66	0.81	0.50	0.64

Table 4: Diffusion Probe’s Predictive Accuracy for External Image Quality Metrics across Diffusion Steps near 5, evaluated on two resolutions (1024

\times

1024 and 512

\times

512). This table reports the predictive alignment between the Diffusion Probe’s internal scores and standard external image quality metrics under different input resolutions. Higher values indicate better predictive accuracy.

Base Model	Resolution	Step	Label Category	SRCC $\uparrow$	AUC-ROC $\uparrow$	KTC $\uparrow$	PCC $\uparrow$
FLUX	1024 $\times$ 1024	5	Aesthetic Score	0.74	0.82	0.58	0.73
FLUX	1024 $\times$ 1024	5	CLIP-SCORE	0.77	0.84	0.62	0.76

Table 5: Diffusion Probe’s Predictive Accuracy for External Image Quality Metrics across Diffusion Step 5 (FLUX Model, 1024

\times

1024 Resolution) with Different Label Categories. This table reports the predictive alignment between the Diffusion Probe’s internal scores and external image quality metrics, trained under two different label categories: Aesthetic Score and CLIP-SCORE. Higher values indicate better predictive accuracy.

Model	Res.	$N$	Ext. $t$	Met.	SRCC	AUC	PCC	KTC
SDXL	768	25	5	IR	0.70	0.81	0.51	0.69
SDXL	768	25	5	CLP	0.63	0.74	0.42	0.61
FLUX	2048	25	5	IR	0.79	0.91	0.64	0.76
			5	CLP	0.72	0.82	0.56	0.71
			21	IR	0.76	0.88	0.61	0.72
			22	IR	0.75	0.89	0.60	0.73

Table 6: Robustness across Architectures, Resolutions, and Steps. Our probe is fixed (trained on FLUX). Results across IR and CLIP metrics.

Prompt	Res.	$N$	Ext. $t$	SRCC	AUC	PCC	KTC	User
Simple	2048	10	2	0.73	0.82	0.71	0.58	79.2%
		25	5	0.72	0.87	0.71	0.56	83.5%
		50	10	0.74	0.88	0.72	0.57	81.3%
Complex	2048	10	2	0.69	0.79	0.67	0.53	76.0%
		25	5	0.68	0.81	0.66	0.52	79.1%
		50	10	0.70	0.83	0.68	0.54	77.5%

Table 7: Generalization and Human Alignment. Zero-shot performance across prompt complexities and sampling steps.

Figure 7: Normal case illustrating generated images and their corresponding cross-attention maps when the image generation is successful with no quality issues. In these cases, the cross-attention maps are well-focused, highlighting the specific regions of the image that correspond to the key features of the prompt. This indicates that when the generated image quality is high, the attention mechanism remains concentrated on the relevant visual attributes, reflecting the model’s proper alignment with the textual description.

Figure 8: We list some cases of generation failures. When the generated image has attributes that do not match the prompt, the corresponding cross-attention map visualization exhibits a dissipation characteristic.

Figure 9: More qualitative results about Diffusion Probe

Figure 10: More qualitative results about Diffusion Probe

T2I Diffusion Models

Text-to-image (T2I) diffusion models generate images from prompts by learning to reverse noise under text conditioning.

See all 403 papers

Toward Early Quality Assessment of Text-to-Image Diffusion Models
Mar 3, 2026Huanlei Guo, Hongxin Wei, Bingyi JingT2I Generation Evaluation T2I Diffusion Models
Southern University of Science and Technology · The Chinese University of Hong Kong, Shenzhen · Shenzhen Loop Area Institute
Naïve PAINE: Lightweight Text-to-Image Generation Improvement with Prompt Evaluation
Mar 12, 2026Joong Ho Kim, Nicholas Thai, Souhardya Saha Dip +2T2I Generation T2I Generation Evaluation
LSU ATHENA Lab, Baton Rouge, LA, U.S.A. · LSU Vision Lab, Baton Rouge, LA, U.S.A.
DynEval: Holistic Evaluations of T2I Generative Models in the Wild
Jul 13, 2026Shyam Marjit, Dheeraj Baiju, Anuj Shikarkhane +3VLM Evaluation Automated Evaluation
Indian Institute of Science · VCL, IISc · Hugging Face

Mar 3, 2026J/K move · Enter open · S save

Toward Early Quality Assessment of Text-to-Image Diffusion Models

Huanlei Guo, Hongxin Wei, Bingyi Jing

Southern University of Science and Technology · The Chinese University of Hong Kong, Shenzhen · Shenzhen Loop Area Institute

T2I Generation Evaluation T2I Diffusion Models Diffusion Sampling

Open paper AskPDF arXiv

Recent text-to-image (T2I) diffusion and flow-matching models can produce highly realistic images from natural language prompts. In practical scenarios, T2I systems are often run in a ``generate--then--select'' mode: many seeds are sampled and only a few images are kept for use. However, this pipeline is highly resource-intensive since each candidate requires tens to hundreds of denoising steps, and evaluation metrics such as CLIPScore and ImageReward are post-hoc. In this work, we address this inefficiency by introducing Probe-Select, a plug-in module that enables efficient evaluation of image quality within the generation process. We observe that certain intermediate denoiser activations, even at early timesteps, encode a stable coarse structure, object layout and spatial arrangement--that strongly correlates with final image fidelity. Probe-Select exploits this property by predicting final quality scores directly from early activations, allowing unpromising seeds to be terminated early. Across diffusion and flow-matching backbones, our experiments show that early evaluation at only 20% of the trajectory accurately ranks candidate seeds and enables selective continuation. This strategy reduces sampling cost by over 60% while improving the quality of the retained images, demonstrating that early structural signals can effectively guide selective generation without altering the underlying generative model. Code is available at https://github.com/Guhuary/ProbeSelect.

Abstract

T2I Diffusion Models

Explore similar work

Toward Early Quality Assessment of Text-to-Image Diffusion Models

Naïve PAINE: Lightweight Text-to-Image Generation Improvement with Prompt Evaluation

DynEval: Holistic Evaluations of T2I Generative Models in the Wild