EasyControlEdge: A Foundation-Model Fine-Tuning for Edge Detection

Hiroki Nakamura; Hiroto Iino; Masashi Okada; Tadahiro Taniguchi

cs.CVFeb 18, 2026

EasyControlEdge: A Foundation-Model Fine-Tuning for Edge Detection

Authors: Hiroki Nakamura, Hiroto Iino, Masashi Okada, Tadahiro Taniguchi

Organizations: Panasonic Holdings Corporation · Kyoto University

Abstract

We propose EasyControlEdge, adapting an image-generation foundation model to edge detection. In real-world edge detection (e.g., floor-plan walls, satellite roads/buildings, and medical organ boundaries), crispness and data efficiency are crucial, yet producing crisp raw edge maps with limited training samples remains challenging. Although image-generation foundation models perform well on many downstream tasks, their pretrained priors for data-efficient transfer and iterative refinement for high-frequency detail preservation remain underexploited for edge detection. To enable crisp and data-efficient edge detection using these capabilities, we introduce an edge-specialized adaptation of image-generation foundation models. To better specialize the foundation model for edge detection, we incorporate an edge-oriented objective with an efficient pixel-space loss. At inference, we introduce guidance based on unconditional dynamics, enabling a single model to control the edge density through a guidance scale. Experiments on BSDS500, NYUDv2, BIPED, and CubiCasa compare against state-of-the-art methods and show consistent gains, particularly under no-post-processing crispness evaluation and with limited training data.

Figures & tables

Figure 1: Overview of EasyControlEdge. The left side shows the training flow and the right side shows the inference flow. We train only a condition-injection LoRA on a frozen DiT-based foundation model (Sec. 2.2 ) with edge-oriented objectives (Sec. 2.3 ), and control edge density at inference via classifier-free guidance with scale $\gamma$ (Sec. 2.4 ).

Methods	BSDS500				NYUDv2				BIPED
	SEval		CEval		SEval		CEval		SEval		CEval
	ODS	OIS	ODS	OIS	ODS	OIS	ODS	OIS	ODS	OIS	ODS	OIS
EDTER [ 20 ]	0.824	0.841	0.698	0.706	0.774	0.789	0.430	0.457	0.893	0.898	–	–
UAED [ 34 ]	0.829	0.847	0.722	0.731	–	–	–	–	–	–	–	–
EdgeSAM-S [ 28 ]	0.838	0.852	–	–	0.783	0.797	–	–	–	–	–	–
EDMB [ 13 ]	0.837	0.851	–	–	0.774	0.787	–	–	0.903	0.907	–	–

Table 1: Quantitative comparisons on BSDS500, NYUDv2, and BIPED.

Method	10% (20 images)
Method	SEval		CEval
	ODS	OIS	ODS	OIS
DiffusionEdge	0.853	0.854	0.851	0.857
GED	0.862	0.874	0.746	0.771
Ours ( $K=5$ )	0.878	0.883	0.875	0.883
Ours	0.877	0.883	0.875	0.883

Table 2: Quantitative comparison on the BIPED dataset. The column headers denote the percentage of training samples.

Figure 2: Qualitative results. Top row shows the results on NYUDv2 and bottom row shows the results on CubiCasa.

Figure 3: Qualitative comparison of different inference steps $K$ on BIPED. Increasing $K$ sharpens edges and recovers fine details.

Method	100%		10%		1%
Method	IoU	F-Score	IoU	F-Score	IoU	F-Score
DiffusionEdge	0.793	0.881	0.757	0.858	0.642	0.775
GED	0.771	0.867	0.738	0.846	0.699	0.819
Ours	0.760	0.860	0.767	0.864	0.739	0.846
Ours (w/ CFG)	0.794	0.882	0.780	0.873	0.742	0.848

Table 3: Quantitative comparison of wall detection on the CubiCasa dataset (4,200 training images). The column headers denote the percentage of training samples used to train the models.

Figure 4: Mean Brightness vs

\gamma

	BIPED		CubiCasa
	CEval
	ODS	OIS	IoU	F-score
w/o $\mathcal{L}_{\mathrm{pix}}$	0.879	0.881	0.785	0.877
w $\mathcal{L}_{\mathrm{pix}}$	0.887	0.893	0.794	0.882

Table 4: The effectiveness of

\mathcal{L}_{\mathrm{pix}}

Figure 5: Qualitative effect of guidance scale $\gamma$ on BIPED.

Figure 6: Qualitative effect of guidance scale $\gamma$ on CubiCasa.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Additional qualitative results on BSDS500, NYUDv2, BIPED, and CubiCasa datasets. Our method consistently produces crisp edge maps across diverse scene types.

Figure 9: The generated edge maps of CubiCasa dataset from 10/50 to 50/50 steps. As the number of steps increases, boundaries stabilize.

Edge Detection

See all 10 papers

MatchED: Crisp Edge Detection Using End-to-End, Matching-based Supervision
Feb 24, 2026Bedrettin Cetinkaya, Sinan Kalkan, Emre AkbasEdge Detection Edge Inference
Dept. of Computer Engineering and METU ROMER Robotics Center, Ankara, Turkey
MEMO: Human-like Crisp Edge Detection Using Masked Edge Prediction
Mar 21, 2026Jiaxin Cheng, Yue Wu, Yicong ZhouEdge Detection Edge Inference
Department of Computer and Information Science, University of Macau, Macau, China · Capital One, USA
Amped: Adaptive Multi-stage Non-edge Pruning for Edge Detection
Mar 29, 2026Yuhan Gao, Xinqing Li, Xin He +4Edge Detection Edge Inference
VCIP, College of Computer Science, Nankai University · Academy for Advanced Interdisciplinary Studies, Nankai University · School of Computer Science and Engineering, Tianjin University of Technology +3

Feb 24, 2026J/K move · Enter open · S save

MatchED: Crisp Edge Detection Using End-to-End, Matching-based Supervision

Bedrettin Cetinkaya, Sinan Kalkan, Emre Akbas

Dept. of Computer Engineering and METU ROMER Robotics Center, Ankara, Turkey

Edge Detection Edge Inference

Open paper AskPDF arXiv

Generating crisp, i.e., one-pixel-wide, edge maps remains one of the fundamental challenges in edge detection, affecting both traditional and learning-based methods. To obtain crisp edges, most existing approaches rely on two hand-crafted post-processing algorithms, Non-Maximum Suppression (NMS) and skeleton-based thinning, which are non-differentiable and hinder end-to-end optimization. Moreover, all existing crisp edge detection methods still depend on such post-processing to achieve satisfactory results. To address this limitation, we propose \MethodLPP, a lightweight, only

\sim

21K additional parameters, and plug-and-play matching-based supervision module that can be appended to any edge detection model for joint end-to-end learning of crisp edges. At each training iteration, \MethodLPP performs one-to-one matching between predicted and ground-truth edges based on spatial distance and confidence, ensuring consistency between training and testing protocols. Extensive experiments on four popular datasets demonstrate that integrating \MethodLPP substantially improves the performance of existing edge detection models. In particular, \MethodLPP increases the Average Crispness (AC) metric by up to 2--4

\times

compared to baseline models. Under the crispness-emphasized evaluation (CEval), \MethodLPP further boosts baseline performance by up to 20--35% in ODS and achieves similar gains in OIS and AP, achieving SOTA performance that matches or surpasses standard post-processing for the first time. Code is available at https://cvpr26-matched.github.io.

Abstract

Edge Detection

Explore similar work

MatchED: Crisp Edge Detection Using End-to-End, Matching-based Supervision

MEMO: Human-like Crisp Edge Detection Using Masked Edge Prediction

Amped: Adaptive Multi-stage Non-edge Pruning for Edge Detection