Reasoning-Informed Visual Editing

Xue Yang; Peiyuan Zhang; Yilun Zhu; Qihao Yang; Mingxin Liu; Xiangyu Zhao; Ziqian Fan; Zhaokai Wang

cs.CVOct 8, 2026

Reasoning-Informed Visual Editing

Authors: Xue Yang, Peiyuan Zhang, Yilun Zhu, Qihao Yang, Mingxin Liu, Xiangyu Zhao, Ziqian Fan, Zhaokai Wang, +7 more

Organizations: Shanghai Jiao Tong University · Southeast University · South China University of Technology · Microsoft Research Asia · Fudan University · East China Normal University

Abstract

Large Multi-modality Models (LMMs) have made significant progress in visual understanding and generation, but still face challenges in visual editing, particularly in following complex instructions, preserving appearance consistency, and supporting flexible input formats. To study this gap, we introduce RISEBench, the first benchmark for evaluating Reasoning-Informed viSual Editing (RISE), and extend it to RISEBench++, a more comprehensive and fine-grained benchmark for this emerging task. RISEBench++ extends the taxonomy into a hierarchical scheme spanning six reasoning dimensions: Temporal, Causal, Spatial, Logical, and Counterfactual Reasoning, together with Hybrid Reasoning integrating multiple reasoning types across multi-turn edits. These dimensions are further decomposed into 12 subcategories and 65 fine-grained task types. We expand input formats to include multi-image conditioning and scale the benchmark to 1000 human-annotated test cases, released in English and Chinese. We also improve our evaluation framework, assessing Instruction Reasoning, Appearance Consistency, and Visual Plausibility with human judges and an LMM-as-a-judge approach for more reliable and calibrated judgements. Beyond benchmarking, we introduce RISE-Agent, a training-free agentic framework integrating reasoning-driven planning, tool-augmented execution, and verifier-guided refinement, outperforming most strong existing approaches across diverse RISE tasks. We evaluate 58 visual editing approaches, including 34 open-source models, 19 closed-source models, and 5 agentic methods. The results reveal substantial challenges in reasoning-based visual editing, with even the strongest evaluated approach, GPT-Image-2.5 Sunburst, achieving only 56.6% accuracy. RISEBench++ highlights the limitations of contemporary editing models, provides insights, and indicates future directions for reasoning-aware visual editing.

Figures & tables

Figure 1 : Overview of RISEBench++. We present illustrative example questions from each of the six major reasoning categories: temporal, causal, spatial, logical, counterfactual, and hybrid reasoning, each demanding profound image understanding and reasoning capabilities.

Figure 2 : Task distribution and word clouds of RISEBench++. RISEBench++ comprises 1,000 tasks organized into a three-level taxonomy of six reasoning dimensions: Temporal (200), Causal (200), Spatial (200), Logical (200), Counterfactual (100), and Hybrid (100) Reasoning.

Benchmark	Reasoning	Multi-turn	Multi-image	Agent Support	# Task Types	# Samples	# Evaluated Models	Language
MagicBrush [ 93 ]	✗	✓	✗	✗	9	1,053	7	EN
ImgEdit-Bench [ 89 ]	✗	✓	✗	✗	14	811	7	EN
AnyEdit [ 91 ]	✗	✗	✓	✗	25	1,250	7	EN
I2I-Bench [ 68 ]	✗	✗	✓	✗	10	1,000	11	EN
UniREditBench [ 21 ]	✓	✗	✗	✗	18	2,700	16	EN
WiseEdit [ 53 ]	✓	✗	✓	✗	4	1,220	22	EN, ZH

Table 1 : Comparison with existing image editing benchmarks.

Figure 3 : Evaluation metrics and pipeline of RISEBench++. RISEBench++ evaluates generated images across Instruction Reasoning , Appearance Consistency , and Visual Plausibility . Gemini-3-Flash evaluates the first two using dimension-specific prompts, while Gemini-3.1-Flash-Lite evaluates the third.

Figure 4 : Overview of RISE-Agent. RISE-Agent follows a closed-loop workflow of reasoning-driven planning, tool-augmented execution, and verifier-guided refinement, bridging target-state inference with precise visual realization.

Figure 5 : Comparison across three evaluation dimensions on RISEBench++ (EN). Nano Banana Pro achieves the highest instruction reasoning score, while GPT-Image-2.5 Sunburst leads in appearance consistency and visual plausibility.

Model	Multi- img	Reasoning Tasks						Overall
		Temporal	Causal	Spatial	Logical	Counterfactual	Hybrid	Single	Multi	Total
Closed-source Models
GPT-Image-2.5 Sunburst [ 52 ]	✓	67.5	69.0	64.0	27.0	72.0	39.0	56.9	52.6	56.6
Nano Banana Pro [ 20 ]	✓	63.0	71.0	51.5	36.0	65.0	38.0	55.4	44.9	54.6
Nano Banana 2 [ 20 ]	✓	67.5	67.0	45.0	28.5	62.0	30.0	51.3	44.9	50.8
GPT-Image-2 [ 52 ]	✓	59.5	64.5	60.5	15.0	58.0	21.0	48.2	43.6	47.8

Table 2 : Overall performance on RISEBench++ (EN). Models are grouped into closed-source, open-source, and agentic methods. Bold and underlined values indicate the best and second-best results within each group, respectively.

Model	Multi- img	Reasoning Tasks						Overall
		Temporal	Causal	Spatial	Logical	Counterfactual	Hybrid	Single	Multi	Total
Closed-source Models
GPT-Image-2.5 Sunburst [ 52 ]	✓	62.5	70.5	63.0	29.0	69.0	36.0	56.0	50.0	55.5
Nano Banana Pro [ 20 ]	✓	62.5	67.0	50.5	38.5	50.0	27.0	52.4	39.7	51.4
Nano Banana 2 [ 20 ]	✓	62.0	69.0	50.0	31.5	57.0	28.0	51.7	42.3	51.0
Luma Uni-1 [ 44 ]	✓	65.5	67.0	53.5	17.0	56.0	24.0	49.1	42.3	48.6

Table 3 : Overall performance on RISEBench++ (CN). Models are grouped into closed-source, open-source, and agentic methods. Bold and underlined values indicate the best and second-best results within each group, respectively.

Model	Multi- img	Temporal			Causal			Spatial		Logical		Counterfactual
		CSK	CtxK	DK	CSK	CtxK	DK	Rel.	Struct.	ERD	IPI	CFI	ACG
Closed-source Models
GPT-Image-2.5 Sunburst [ 52 ]	✓	72.9	67.3	60.3	78.8	55.0	66.3	63.2	64.9	32.5	18.2	71.4	73.3
Nano Banana Pro [ 20 ]	✓	68.2	63.5	55.6	77.5	65.0	67.5	50.9	52.1	41.5	27.3	61.4	73.3
Nano Banana 2 [ 20 ]	✓	77.6	63.5	57.1	76.3	50.0	66.3	47.2	42.6	38.2	13.0	58.6	70.0
GPT-Image-2 [ 52 ]	✓	61.2	67.3	50.8	76.3	60.0	55.0	60.4	60.6	21.1	5.2	52.9	70.0

Table 4 : Performance across subcategories on RISEBench++ (EN). Models are likewise grouped into closed-source, open-source, and agentic methods; bold and underlined mark the best and second-best in each group.

Model	Multi- img	Temporal			Causal			Spatial		Logical		Counterfactual
		CSK	CtxK	DK	CSK	CtxK	DK	Rel.	Struct.	ERD	IPI	CFI	ACG
Closed-source Models
GPT-Image-2.5 Sunburst [ 52 ]	✓	69.4	65.4	50.8	85.0	62.5	60.0	62.3	63.8	32.5	23.4	64.3	80.0
Nano Banana Pro [ 20 ]	✓	68.2	63.5	54.0	77.5	62.5	58.8	54.7	45.7	42.3	32.5	54.3	40.0
Nano Banana 2 [ 20 ]	✓	74.1	51.9	54.0	83.8	57.5	60.0	47.2	53.2	35.0	26.0	51.4	70.0
Luma Uni-1 [ 44 ]	✓	81.2	48.1	58.7	75.0	60.0	62.5	52.8	54.3	21.1	10.4	58.6	50.0

Table 5 : Performance across subcategories on RISEBench++ (ZH). Models are likewise grouped into closed-source, open-source, and agentic methods; bold and underlined mark the best and second-best in each group.

Figure 6 : Radar charts of fine-grained task accuracy (%) on RISEBench++ (EN). Results for seven models across 65 task types in 12 subcategories.

Judge	Instruction Reasoning		Appearance Consistency		Visual Plausibility		Overall
Judge	MAE $\downarrow$	Spearman $\uparrow$	MAE $\downarrow$	Spearman $\uparrow$	MAE $\downarrow$	Spearman $\uparrow$	MAE $\downarrow$	Spearman $\uparrow$
GPT-4.1	0.88	0.64	0.62	0.51	0.35	0.35	0.60	0.73
GPT-5	0.73	0.74	0.66	0.59	0.53	0.27	0.52	0.81
Gemini-3.1-Flash-Lite	0.71	0.71	0.76	0.47	0.27	0.41	0.60	0.74
Gemini-3-Flash	0.48	0.84	0.58	0.61	0.39	0.37	0.46	0.85
Seed-2.0-Lite	0.64	0.81	0.88	0.61	0.52	0.39	0.59	0.82

Table 6 : Comparison of different judge models on instruction reasoning, appearance consistency, visual plausibility, and overall evaluation. Lower MAE is better, while higher Spearman correlation is better.

Figure 7 : Examples of several different models’ outputs on RISEBench++. The analyzed models demonstrate distinct characteristics in their responses.

Components			Performance on RISEBench++
Planner	Tools	Verifier	Temporal	Causal	Spatial	Logical	Counterfactual	Hybrid	Overall
–	–	–	16.0	16.5	22.0	0.5	17.0	2.0	12.9
✓	–	–	58.0	50.5	48.0	14.0	51.0	13.0	40.5
✓	✓	–	50.5	56.0	37.5	32.0	61.0	12.0	42.6
✓	✓	✓	58.5	58.0	37.5	39.0	66.0	28.0	48.0

Table 7 : Ablation study on key components of RISE-Agent, with results reported as accuracy (%). FLUX.2 Klein 9B without RISE-Agent components serves as the baseline. Best and second-best results are shown in bold and underline , respectively.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Subcategory	Task Type
Temporal Reasoning (200, 32)
Common-Sense Knowledge (85)	Life Progression (24)
	Material Transformation (9)
	Natural Environmental Changes (13)
	Physical Aging (20)
	Social Activities (8)

Appendix

Table 8 : Taxonomy and Statistics of RISEBench++.

Model	Factual Knowledge				Conceptual Knowledge			Procedural Knowledge			Overall Score
Model	Attribute	Spatial	Temporal	Average	Social Sci.	Natural Sci.	Average	Logical	Instruction	Average	Overall Score
Closed-source Models
GPT-Image-2.5 Sunburst [ 52 ]	95.76	91.92	93.92	94.60	90.93	91.34	91.24	89.09	97.17	93.14	92.89
GPT-Image-2.5 Flare [ 52 ]	94.95	89.58	93.81	93.56	89.90	90.06	90.02	86.87	98.61	92.76	91.93
Nano Banana Pro [ 20 ]	85.61	87.17	88.06	86.36	85.25	83.64	84.03	81.42	93.22	86.48	85.31
Nano Banana [ 20 ]	78.97	80.17	86.49	80.48	83.70	76.34	78.11	63.96	90.44	75.31	78.17

Appendix

Table 9 : Detail performance on KRIS-Bench.

Human	Logical Reasoning			Temporal Reasoning			Causal Reasoning			Spatial Reasoning			Counterfactual Reasoning			Hybrid Reasoning
Score	Reas.	Cons.	Plau.	Reas.	Cons.	Plau.	Reas.	Cons.	Plau.	Reas.	Cons.	Plau.	Reas.	Cons.	Plau.	Reas.	Cons.	Plau.
1	75.0%	20.0%	0.0%	27.5%	0.0%	0.0%	25.0%	5.0%	2.5%	50.0%	8.3%	0.0%	45.8%	0.0%	0.0%	50.0%	40.0%	25.0%
2	0.0%	0.0%	0.0%	0.0%	2.5%	0.0%	2.5%	0.0%	0.0%	5.6%	8.3%	0.0%	0.0%	4.2%	0.0%	15.0%	0.0%	0.0%
3	0.0%	5.0%	0.0%	15.0%	20.0%	0.0%	7.5%	12.5%	5.0%	8.3%	30.6%	0.0%	4.2%	12.5%	8.3%	15.0%	10.0%	0.0%
4	5.0%	2.5%	0.0%	5.0%	17.5%	2.5%	0.0%	7.5%	7.5%	0.0%	2.8%	5.6%	4.2%	12.5%	0.0%	20.0%	15.0%	5.0%
5	20.0%	72.5%	100.0%	52.5%	60.0%	97.5%	65.0%	75.0%	85.0%	36.1%	50.0%	94.4%	45.8%	70.8%	91.7%	0.0%	35.0%	70.0%

Appendix

Table 10 : Sample and human score distributions across reasoning categories for human alignment evaluation.

Model	Multi img	Temporal			Causal			Spatial			Logical			Counterfactual			Hybrid			Overall
		IR	AC	VP	IR	AC	VP	IR	AC	VP	IR	AC	VP	IR	AC	VP	IR	AC	VP	IR	AC	VP
Closed-source Models
GPT-Image-2.5 Sunburst [ 52 ]	✓	81.0	98.0	97.0	83.1	96.3	97.6	81.3	96.6	99.0	27.0	95.0	—	78.3	99.5	99.5	58.0	91.7	97.4	68.1	96.3	98.0
Nano Banana Pro [ 20 ]	✓	80.2	92.6	98.4	89.0	94.2	97.8	74.9	88.1	98.6	38.0	79.0	—	74.8	88.5	90.8	59.3	80.8	93.5	69.8	87.7	96.8
Nano Banana 2 [ 20 ]	✓	84.5	95.9	97.1	86.6	92.4	97.4	75.5	87.4	97.8	29.0	87.0	—	78.5	97.3	98.0	55.3	88.6	94.5	68.5	91.1	97.2
GPT-Image-2 [ 52 ]	✓	76.6	98.9	97.5	80.5	96.8	95.6	76.3	94.5	98.1	16.0	92.0	—	66.5	98.8	100.0	49.3	88.9	98.1	61.5	95.2	97.6

Appendix

Table 11 : Detailed evaluation dimension scores on RISEBench++ (EN).

Model	Multi img	Temporal			Causal			Spatial			Logical			Counterfactual			Hybrid			Overall
		IR	AC	VP	IR	AC	VP	IR	AC	VP	IR	AC	VP	IR	AC	VP	IR	AC	VP	IR	AC	VP
Closed-source Models
GPT-Image-2.5 Sunburst [ 52 ]	✓	78.1	96.3	96.9	82.3	96.3	96.6	80.4	95.7	99.0	29.0	98.5	—	77.5	98.5	100.0	58.8	94.4	98.4	67.6	96.7	97.9
Nano Banana Pro [ 20 ]	✓	84.5	92.2	98.3	86.8	92.9	95.3	75.9	87.9	95.7	39.0	84.5	—	76.0	87.5	99.3	58.0	83.8	93.5	70.6	88.6	96.5
Nano Banana 2 [ 20 ]	✓	84.3	93.9	96.5	88.3	94.1	97.5	75.1	87.4	95.0	32.5	84.0	—	80.5	93.0	99.5	56.0	78.3	97.4	69.7	88.9	96.8
Luma Uni-1 [ 44 ]	✓	83.0	95.3	96.0	81.1	94.2	93.9	75.0	89.9	96.6	18.0	83.0	—	70.0	92.8	96.3	48.0	78.5	88.3	63.2	89.5	94.9

Appendix

Table 12 : Detailed evaluation dimension scores on RISEBench++ (ZH).

Model	Multi- img	Temporal
		T01	T02	T03	T04	T05	T06	T07	T08	T09	T10	T11	T12	T13	T14	T15	T16	T17	T18	T19
Closed-source Models
GPT-Image-2.5 Sunburst [ 52 ]	✓	58.3	88.9	76.9	80.0	87.5	71.4	50.0	60.0	50.0	73.3	100.0	50.0	50.0	66.7	58.3	66.7	40.0	75.0	66.7
Nano Banana Pro [ 20 ]	✓	62.5	100.0	76.9	70.0	50.0	42.9	75.0	66.7	64.3	60.0	62.5	75.0	25.0	66.7	66.7	50.0	60.0	75.0	33.3
Nano Banana 2 [ 20 ]	✓	62.5	88.9	76.9	90.0	100.0	57.1	75.0	60.0	71.4	53.3	75.0	25.0	37.5	55.6	66.7	55.6	100.0	50.0	66.7
GPT-Image-2 [ 52 ]	✓	58.3	88.9	69.2	65.0	62.5	42.9	0.0	53.3	64.3	86.7	62.5	50.0	50.0	66.7	50.0	44.4	40.0	75.0	33.3

Appendix

Table 13 : Fine-grained task accuracy (%) on RISEBench++ (EN). Part I: Temporal reasoning.

Model	Multi- img	Causal														Spatial (Relation)
		C01	C02	C03	C04	C05	C06	C07	C08	C09	C10	C11	C12	C13	C14	S01	S02	S03	S04	S05	S06	S07	S08	S09
Closed-source Models
GPT-Image-2.5 Sunburst [ 52 ]	✓	80.6	85.2	68.2	50.0	60.0	30.0	80.0	56.0	69.6	100.0	100.0	50.0	66.7	40.0	63.6	72.7	45.5	58.3	76.9	80.0	72.2	44.4	16.7
Nano Banana Pro [ 20 ]	✓	74.2	74.1	86.4	40.0	80.0	60.0	80.0	64.0	73.9	91.7	100.0	50.0	33.3	40.0	81.8	45.5	54.5	58.3	53.8	46.7	55.6	33.3	0.0
Nano Banana 2 [ 20 ]	✓	71.0	74.1	86.4	40.0	40.0	70.0	50.0	64.0	73.9	91.7	66.7	50.0	0.0	50.0	54.5	54.5	36.4	41.7	46.2	60.0	55.6	33.3	16.7
GPT-Image-2 [ 52 ]	✓	71.0	77.8	81.8	60.0	60.0	60.0	60.0	36.0	69.6	83.3	100.0	0.0	100.0	30.0	72.7	54.5	54.5	66.7	76.9	60.0	61.1	55.6	16.7

Appendix

Table 14 : Fine-grained task accuracy (%) on RISEBench++ (EN). Part II: Causal reasoning and Spatial Relation.

Model	Multi- img	Spatial (Structure)					Logical							Counterfactual
		S10	S11	S12	S13	S14	L01	L02	L03	L04	L05	L06	L07	F01	F02	F03	F04	F05	F06	F07	F08	F09	F10	F11
Closed-source Models
GPT-Image-2.5 Sunburst [ 52 ]	✓	79.2	73.3	65.4	52.6	40.0	25.0	29.6	40.5	31.4	12.8	33.3	15.0	87.0	66.7	87.5	42.9	75.0	57.1	50.0	50.0	90.0	63.6	66.7
Nano Banana Pro [ 20 ]	✓	58.3	66.7	50.0	36.8	50.0	29.2	66.7	45.9	25.7	25.6	27.8	30.0	78.3	55.6	87.5	42.9	12.5	85.7	25.0	50.0	80.0	81.8	55.6
Nano Banana 2 [ 20 ]	✓	41.7	53.3	34.6	42.1	50.0	33.3	51.9	45.9	22.9	12.8	27.8	0.0	65.2	55.6	100.0	28.6	12.5	71.4	50.0	75.0	80.0	81.8	44.4
GPT-Image-2 [ 52 ]	✓	79.2	73.3	50.0	57.9	30.0	16.7	22.2	27.0	17.1	10.3	0.0	0.0	52.2	55.6	87.5	42.9	50.0	42.9	50.0	25.0	100.0	72.7	33.3

Appendix

Table 15 : Fine-grained task accuracy (%) on RISEBench++ (EN). Part III: Spatial Structure, Logical reasoning, and Counterfactual reasoning.

Model	Multi- img	Temporal
		T01	T02	T03	T04	T05	T06	T07	T08	T09	T10	T11	T12	T13	T14	T15	T16	T17	T18	T19
Closed-source Models
GPT-Image-2.5 Sunburst [ 52 ]	✓	62.5	77.8	69.2	75.0	87.5	57.1	50.0	46.7	57.1	86.7	75.0	50.0	37.5	66.7	50.0	50.0	20.0	75.0	66.7
Nano Banana Pro [ 20 ]	✓	58.3	100.0	76.9	70.0	62.5	57.1	50.0	80.0	71.4	53.3	37.5	50.0	62.5	66.7	50.0	44.4	60.0	50.0	66.7
Nano Banana 2 [ 20 ]	✓	70.8	100.0	69.2	75.0	75.0	71.4	50.0	53.3	42.9	60.0	50.0	50.0	62.5	44.4	50.0	50.0	20.0	100.0	100.0
Luma Uni-1 [ 44 ]	✓	70.8	100.0	76.9	90.0	87.5	71.4	75.0	53.3	35.7	46.7	62.5	100.0	37.5	66.7	83.3	38.9	60.0	50.0	66.7

Appendix

Table 16 : Fine-grained task accuracy (%) on RISEBench++ (ZH). Part I: Temporal reasoning.

Model	Multi- img	Causal														Spatial (Relation)
		C01	C02	C03	C04	C05	C06	C07	C08	C09	C10	C11	C12	C13	C14	S01	S02	S03	S04	S05	S06	S07	S08	S09
Closed-source Models
GPT-Image-2.5 Sunburst [ 52 ]	✓	77.4	92.6	86.4	60.0	60.0	70.0	60.0	36.0	73.9	91.7	100.0	50.0	100.0	30.0	63.6	81.8	54.5	66.7	61.5	80.0	55.6	55.6	16.7
Nano Banana Pro [ 20 ]	✓	80.6	74.1	77.3	50.0	90.0	80.0	30.0	52.0	82.6	66.7	100.0	25.0	0.0	30.0	63.6	54.5	45.5	58.3	76.9	80.0	38.9	33.3	16.7
Nano Banana 2 [ 20 ]	✓	87.1	81.5	81.8	40.0	70.0	50.0	70.0	56.0	65.2	66.7	100.0	50.0	66.7	40.0	72.7	45.5	27.3	58.3	46.2	60.0	55.6	22.2	0.0
Luma Uni-1 [ 44 ]	✓	74.2	77.8	72.7	50.0	70.0	40.0	80.0	40.0	91.3	100.0	66.7	25.0	66.7	20.0	63.6	18.2	54.5	66.7	61.5	66.7	50.0	33.3	50.0

Appendix

Table 17 : Fine-grained task accuracy (%) on RISEBench++ (ZH). Part II: Causal reasoning and Spatial Relation.

Model	Multi- img	Spatial (Structure)					Logical							Counterfactual
		S10	S11	S12	S13	S14	L01	L02	L03	L04	L05	L06	L07	F01	F02	F03	F04	F05	F06	F07	F08	F09	F10	F11
Closed-source Models
GPT-Image-2.5 Sunburst [ 52 ]	✓	79.2	66.7	65.4	47.4	50.0	20.8	25.9	43.2	34.3	15.4	55.6	10.0	73.9	66.7	75.0	28.6	75.0	71.4	25.0	50.0	100.0	81.8	55.6
Nano Banana Pro [ 20 ]	✓	75.0	73.3	30.8	31.6	0.0	41.7	63.0	45.9	22.9	30.8	33.3	35.0	60.9	88.9	62.5	42.9	50.0	28.6	0.0	50.0	60.0	18.2	44.4
Nano Banana 2 [ 20 ]	✓	66.7	73.3	50.0	31.6	40.0	20.8	55.6	37.8	25.7	23.1	38.9	20.0	43.5	66.7	62.5	42.9	50.0	57.1	25.0	75.0	90.0	63.6	55.6
Luma Uni-1 [ 44 ]	✓	66.7	80.0	34.6	47.4	50.0	4.2	37.0	18.9	22.9	7.7	16.7	10.0	65.2	77.8	50.0	57.1	50.0	57.1	25.0	50.0	50.0	54.5	44.4

Appendix

Table 18 : Fine-grained task accuracy (%) on RISEBench++ (ZH). Part III: Spatial Structure, Logical reasoning, and Counterfactual reasoning.

Figure 8 : Score distributions on RISEBench++ (EN). Results for ten models across 1,000 cases. The three columns report Instruction Reasoning (IR), Appearance Consistency (AC), and Visual Plausibility (VP). Bars count cases at each score from 1 to 5.

Explore similar work

Is This Edit Correct? A Multi-Dimensional Benchmark for Reasoning-Aware Image Editing
Apr 16, 2026Yixuan Ding, Wei Huang, Ruijie Quan +2Image Editing Evaluation Text-Guided Image Editing
Zhejiang University · The University of Hong Kong
PaintBench: Deterministic Evaluation of Precise Visual Editing
May 29, 2026Kai Xu, Ellis Brown, Shrikar Madhu +3Image Editing Evaluation Text-Guided Image Editing
New York University
RS-RIE-Bench: Benchmarking Reasoning-Guided Remote Sensing Image Editing
Jul 22, 2026Zihan Qin, Boao Xu, Zhao Dong +4Image Editing Evaluation Remote Sensing Benchmarks
School of Artificial Intelligence, Optics and Electronics (iOPEN), Northwestern Polytechnical University, Xi’an 710072, China. · Xi’an Modern Control Technology Research Institute, Xi’an 710065, China. · Lanzhou Institute of Physics, Lanzhou 730000, China. +2

Apr 16, 2026, cs.HCJ/K move · Enter open · S save

Is This Edit Correct? A Multi-Dimensional Benchmark for Reasoning-Aware Image Editing

Yixuan Ding, Wei Huang, Ruijie Quan, Xiaojuan Qi+1

Zhejiang University · The University of Hong Kong

Image Editing Evaluation Text-Guided Image Editing Image Editing

Open paper AskPDF arXiv

Diffusion-based image editing has achieved strong visual fidelity under natural language instructions, yet most existing systems still operate at the level of surface instruction following, without reasoning about the implicit contextual constraints embedded in real user requests. This often leads to visually plausible but logically inconsistent edits. In this work, we introduce RE-Edit, a benchmark for REasoning-aware image Editing that evaluates image editing systems across five complementary reasoning dimensions: physical, environmental, cultural, causal, and referential. RE-Edit comprises 1,000 carefully curated samples, each designed such that visual plausibility alone is insufficient and correct editing requires satisfying implicit logical constraints. To support fine-grained analysis, we establish dimension-aligned evaluation criteria and conduct a comprehensive study of ten open-source and two commercial image editing models. Our results show that even advanced systems frequently struggle with implicit multi-dimensional reasoning despite producing high-quality visuals. We further present a lightweight reasoning-guided post-edit baseline as an initial exploration, illustrating how inserting explicit reasoning can help mitigate such failures in a model-agnostic manner.

Abstract

Explore similar work

Is This Edit Correct? A Multi-Dimensional Benchmark for Reasoning-Aware Image Editing

PaintBench: Deterministic Evaluation of Precise Visual Editing

RS-RIE-Bench: Benchmarking Reasoning-Guided Remote Sensing Image Editing