GeoComposer: Geometry-Grounded Photographic Composition Instruction

Shuangzhi Li; Qiaoqiao Jia; Xingxin Chen; Guile Wu; Dongfeng Bai

cs.CVSep 22, 2026

GeoComposer: Geometry-Grounded Photographic Composition Instruction

Authors: Shuangzhi Li, Qiaoqiao Jia, Xingxin Chen, Guile Wu, Dongfeng Bai

Organizations: Huawei Noah’s Ark Lab · University of Alberta · University of Waterloo

Abstract

Photographic composition aims to provide visual guidance for improving the framing, viewpoint, and spatial arrangement of an image. Early methods primarily rely on image cropping to enhance composition, which is restricted to the viewpoint and spatial arrangement of the input image. Recent methods have explored image understanding and editing to improve composition, but they mainly focus on instruction following and aesthetic quality, overlooking the importance of 3D scene geometry consistency for photographic composition. In this work, we propose GeoComposer, a novel geometry-grounded photographic composition framework that analyzes the composition of a given image to generate textual guidance and synthesizes a visual exemplar that enhances the composition of the given image. To promote geometry-grounded composition, we propose a geometry-aware representation learning mechanism that leverages geometric priors from a visual geometry foundation model to shape the intermediate representations of the composition editing model. This mechanism preserves both global structural relationships and local fine-grained correspondences for geometry-grounded composition. Furthermore, we propose a reinforcement learning strategy guided by a hybrid reward that jointly optimizes instruction following, aesthetic quality, and geometric consistency. This enables the model to generate visual exemplars that faithfully follow the composition instructions while remaining visually appealing and geometrically consistent. Extensive experiments show the superiority of our approach over state-of-the-art methods, highlighting its effectiveness in generating visually appealing and geometrically consistent composition.

Figures & tables

Figure 2: The overall framework of the proposed GeoComposer.

	Composition		Image Quality		Geo. Consistency		Text Guidance		Human
Method	WinRate (%) $\uparrow$	PF-ass $\uparrow$	QAlign $\uparrow$	DeQA $\uparrow$	Geo-Suc (%) $\uparrow$	MET3R $\downarrow$	src $\rightarrow$ gen $\uparrow$	src $\rightarrow$ GT $\uparrow$	Overall $\uparrow$
BAGEL	46.5	2.588	2.823	3.598	63.2	0.246	2.69	2.57	43.3
Step1X-Edit	52.1	2.704	2.870	3.496	52.8	0.231	2.69	2.30	–
Qwen-Image-Edit	73.9	2.987	3.272	3.869	54.4	0.257	3.16	2.60	59.4
FLUX.2	67.9	2.948	3.133	3.820	61.0	0.233	–	–	40.6
PhotoFramer	68.9	2.957	3.043	3.812	61.7	0.231	3.21	2.69	63.3

Table 1: Quantitative comparison with state-of-the-art methods on the constructed validation set. The best and second-best results are marked in bold and underline , respectively. “–” denotes unavailable results, i.e., pure image generators emit no reasoning text for text-guidance evaluation, and Step1X-Edit has no human-study result.

Figure 3: Qualitative comparison with state-of-the-art methods.

Und.	Stage I	Stage II	QAlign $\uparrow$	DeQA $\uparrow$	MET3R $\downarrow$	PF-ass $\uparrow$
			3.272	3.869	0.257	2.987
$\checkmark$			3.316	3.925	0.275	3.050
$\checkmark$	$\checkmark$		3.282	3.934	0.222	3.043
$\checkmark$	$\checkmark$	$\checkmark$	3.381	4.012	0.184	3.133

Table 2: Ablation study of components of GeoComposer. “Und.” is the composition understanding model, “Stage I” denotes geometry-aware representation learning, and “Stage II” denotes hybrid reward-guided reinforcement learning.

Model	QAlign $\uparrow$	DeQA $\uparrow$	MET3R $\downarrow$	PF-ass $\uparrow$
Qwen3-VL	3.370	3.992	0.201	3.096
PhotoFramer	3.375	3.938	0.191	3.051
Ours	3.381	4.012	0.184	3.133

Table 3: Ablation study of composition understanding model.

Supervision	QAlign $\uparrow$	DeQA $\uparrow$	MET3R $\downarrow$	PF-ass $\uparrow$
L2	3.260	3.890	0.249	3.011
Local	3.250	3.900	0.240	3.020
Loc.+Glob.	3.282	3.934	0.222	3.043

Table 4: Ablation study of the supervision design in geometry-aware representation learning, evaluated after Stage I. “Loc.+Glob.” denotes local correspondence combined with global structural supervision.

Reward	QAlign $\uparrow$	DeQA $\uparrow$	MET3R $\downarrow$	PF-ass $\uparrow$
Full (Ours)	3.381	4.012	0.184	3.133
$-$ Instruction	3.249	3.626	0.182	3.012
$-$ Quality	3.375	3.931	0.197	3.127
$-$ Geometry	3.405	3.957	0.303	3.014

Table 5: Ablation study of the hybrid reward, where each row removes one reward component from the full hybrid reward.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Subset	Pairs	Source
Shift	13,152	CPC, GAIC
Zoom-in	10,454	CPC, FLMS, GAIC, CUHK-ICD
View change	10,523	Unsplash
Total	34,129	–

Appendix

Table 6: Per-task statistics of the constructed dataset. The view-change poor images are synthesized from Unsplash expert photographs.

Figure 4: Additional qualitative comparison with state-of-the-art methods on diverse scenes. In each row, the leftmost image is the poorly-composed input and the rightmost is the result of GeoComposer.

Method	QAlign $\uparrow$	DeQA $\uparrow$	MET3R $\downarrow$	PF-ass $\uparrow$	Cons. $\uparrow$
BAGEL	2.208	4.063	0.266	2.452	2.263
Qwen-Image-Edit	2.323	4.016	0.242	2.456	3.337
Step1X-Edit	2.087	3.777	0.231	2.327	2.407
FLUX.2	2.247	4.027	0.259	2.563	–
PhotoFramer	2.299	4.182	0.453	2.698	3.423
Nano Banana Pro	2.277	4.082	0.177	2.470	–

Appendix

Table 7: Comparison on the DL3DV-300 benchmark for viewpoint change. Cons. denotes Text Cons. (src

\rightarrow

gen); “–” denotes pure image generators excluded from text-guidance evaluation. The best and second-best results are marked in bold and underline , respectively.

Figure 5: Word distribution of the textual guidance produced by each composition understanding model over the validation set. Word size is proportional to frequency.

Und. Model	src $\rightarrow$ gen $\uparrow$	src $\rightarrow$ GT $\uparrow$	Words
Raw Qwen3-VL	3.16	2.60	210.6
PhotoFramer	3.21	2.69	100.9
Ours	3.45	3.28	58.6

Appendix

Table 8: Comparison of composition understanding models on text consistency and the average number of words per generated guidance. The best and second-best consistency scores are marked in bold and underline , respectively.

$\lambda_{\mathrm{glo}}$	QAlign $\uparrow$	DeQA $\uparrow$	MET3R $\downarrow$	PF-ass $\uparrow$
0.05	3.310	3.908	0.242	3.022
0.10	3.282	3.934	0.222	3.043
0.20	3.274	3.926	0.235	3.038

Appendix

Table 9: Ablation of the global structural supervision weight

\lambda_{\mathrm{glo}}

Layers	QAlign $\uparrow$	DeQA $\uparrow$	MET3R $\downarrow$	PF-ass $\uparrow$
23	3.267	3.933	0.256	3.038
18,23	3.204	3.926	0.241	3.068
12,18,23	3.282	3.934	0.222	3.043

Appendix

Table 10: Ablation of the VGGT layers whose patch features serve as the target of the local and global supervision in Stage I.

$\lambda_{\mathrm{attn}}{:}\lambda_{\mathrm{feat}}$	QAlign $\uparrow$	DeQA $\uparrow$	MET3R $\downarrow$	PF-ass $\uparrow$
0:1	3.294	3.990	0.185	3.063
3:7	3.381	4.012	0.184	3.133
5:5	3.377	4.006	0.206	3.143

Appendix

Table 11: Ablation of the attention and feature relevance weights

\lambda_{\mathrm{attn}}{:}\lambda_{\mathrm{feat}}

in the directed geometry relevance score.

$w_{s}{:}w_{t}$	QAlign $\uparrow$	DeQA $\uparrow$	MET3R $\downarrow$	PF-ass $\uparrow$
0:1	3.344	3.993	0.209	3.124
3:7	3.381	4.012	0.184	3.133
5:5	3.258	3.802	0.176	2.999
7:3	3.239	3.757	0.174	2.988

Appendix

Table 12: Ablation of the source and target weights

w_{s}{:}w_{t}

in the geometry-consistency reward. The two source-heavy settings reach a lower MET3R by under-editing, at a clear cost in composition and quality.

Image Editing

Image editing changes selected visual content while preserving the parts of an image that should remain unchanged.

See all 315 papers

Mar 26, 2026

Geo $^\textbf{2}$ : Geometry-Guided Cross-view Geo-Localization and Image Synthesis

Cross-view geo-spatial learning consists of two important tasks: Cross-View Geo-Localization (CVGL) and Cross-View Image Synthesis (CVIS), both of which rely on establishing geometric correspondences between ground and aerial views. Recent Geometric Foundation Models (GFMs) have demonstrated strong capabilities in extracting generalizable 3D geometric features from images, but their potential in cross-view geo-spatial tasks remains underexplored. In this work, we present Geo^2, a unified framework that leverages Geometric priors from GFMs (e.g., VGGT) to jointly perform geo-spatial tasks, CVGL and bidirectional CVIS. Despite the 3D reconstruction ability of GFMs, directly applying them to CVGL and CVIS remains challenging due to the large viewpoint gap between ground and aerial imagery. We propose GeoMap, which embeds ground and aerial features into a shared 3D-aware latent space, effectively reducing cross-view discrepancies for localization. This shared latent space naturally bridges cross-view image synthesis in both directions. To exploit this, we propose GeoFlow, a flow-matching model conditioned on geometry-aware latent embeddings. We further introduce a consistency loss to enforce latent alignment between the two synthesis directions, ensuring bidirectional coherence. Extensive experiments on standard benchmarks, including CVUSA, CVACT, and VIGOR, demonstrate that Geo^2 achieves state-of-the-art performance in both localization and synthesis, highlighting the effectiveness of 3D geometric priors for cross-view geo-spatial learning.

Yancheng Zhang, Xiaohan Zhang, Guangyu Sun +3

Institute of Artificial Intelligence, University of Central Florida · University of Vermont

Cross-View Geo-Localization Geospatial Representation Learning Image Generation Geometric Representation Learning

Ask PDF

Apr 19, 2026

UniGeo: Unifying Geometric Guidance for Camera-Controllable Image Editing via Video Models

Camera-controllable image editing aims to synthesize novel views of a given scene under varying camera poses while strictly preserving cross-view geometric consistency. However, existing methods typically rely on fragmented geometric guidance, such as only injecting point clouds at the representation level despite models containing multiple levels, and are mainly based on image diffusion models that operate on discrete view mappings. These two limitations jointly lead to geometric drift and structural degradation under continuous camera motion. We observe that while leveraging video models provides continuous viewpoint priors for camera-controllable image editing, they still struggle to form stable geometric understanding if geometric guidance remains fragmented. To systematically address this, we inject unified geometric guidance across three levels that jointly determine the generative output: representation, architecture, and loss function. To this end, we propose UniGeo, a novel camera-controllable editing framework. Specifically, at the representation level, UniGeo incorporates a frame-decoupled geometric reference injection mechanism to provide robust cross-view geometry context. At the architecture level, it introduces geometric anchor attention to align multi-view features. At the loss function level, it proposes a trajectory-endpoint geometric supervision strategy to explicitly reinforce the structural fidelity of target views. Comprehensive experiments across multiple public benchmarks, encompassing both extensive and limited camera motion settings, demonstrate that UniGeo significantly outperforms existing methods in both visual quality and geometric consistency.

Hong Jiang, Wensong Song, Zongxin Yang +2

ReLER, CCAI, Zhejiang University · DBMI, HMS, Harvard University

Multi-View Consistency Diffusion Model Guidance Novel View Synthesis Image Editing

Ask PDF

Mar 17, 2026

Interact3D: Compositional 3D Generation of Interactive Objects

Recent breakthroughs in 3D generation have enabled the synthesis of high-fidelity individual assets. However, generating 3D compositional objects from single images--particularly under occlusions--remains challenging. Existing methods often degrade geometric details in hidden regions and fail to preserve the underlying object-object spatial relationships (OOR). We present a novel framework Interact3D designed to generate physically plausible interacting 3D compositional objects. Our approach first leverages advanced generative priors to curate high-quality individual assets with a unified 3D guidance scene. To physically compose these assets, we then introduce a robust two-stage composition pipeline. Based on the 3D guidance scene, the primary object is anchored through precise global-to-local geometric alignment (registration), while subsequent geometries are integrated using a differentiable Signed Distance Field (SDF)-based optimization that explicitly penalizes geometry intersections. To reduce challenging collisions, we further deploy a closed-loop, agentic refinement strategy. A Vision-Language Model (VLM) autonomously analyzes multi-view renderings of the composed scene, formulates targeted corrective prompts, and guides an image editing module to iteratively self-correct the generation pipeline. Extensive experiments demonstrate that Interact3D successfully produces promising collsion-aware compositions with improved geometric fidelity and consistent spatial relationships.

Hui Shan, Keyang Luo, Ming Li +4

Zhejiang University, China · Shanghai Innovation Institute, China · Westlake University, China +4

3D Asset Generation 3D Scene Generation 3D Shape Generation

Ask PDF

Abstract

Image Editing

Explore similar work

Geo2^\textbf{2}2: Geometry-Guided Cross-view Geo-Localization and Image Synthesis

UniGeo: Unifying Geometric Guidance for Camera-Controllable Image Editing via Video Models

Interact3D: Compositional 3D Generation of Interactive Objects

Geo $^\textbf{2}$ : Geometry-Guided Cross-view Geo-Localization and Image Synthesis