RCEdit-500K: Reference Completion for Image-Conditioned Image Editing

📄Abstract

Image-conditioned image editing (ICIE) guides edits with a reference image to convey visual attributes—such as style, tone, or object identity—that are difficult to specify through text alone. Despite growing practical demand, no large-scale unified ICIE dataset currently exists in the open-source ecosystem; existing efforts cover only a narrow subset of edit types at small scale, typically constructed via costly forward synthesis.

We reformulate ICIE data construction as a reference-completion problem: high-quality text-conditioned image editing (TCIE) datasets already supply the input image, instruction, and edited target, and can be augmented with aligned reference images through type-specific synthesis and lightweight instruction adaptation. Building on this insight, we propose a scalable pipeline equipped with weak-instruction augmentation and five-dimensional VLM-based post-filtering, and use it to construct RCEdit-500K—the first large-scale unified open ICIE dataset comprising 477K quadruplets across six edit categories (add, remove, replace, background, style, alter) with both concrete and abstract reference types.

Training on RCEdit-500K consistently improves reference-guided editing: LoRA adaptation on diffusion models yields up to +1.22 average gain, and an autoregressive model without native editing ability acquires competitive ICIE performance, demonstrating that data availability is the primary bottleneck for open-source ICIE.

⚙Pipeline

Our reference-completion pipeline converts any TCIE dataset into ICIE training data through four stages: GPT-4o analysis, type-specific reference generation, GPT Image processing (for style/alter), and 5-dimensional VLM-based post-filtering.

Overview of the RCEdit-500K data construction pipeline. Given an existing TCIE pair (input, instruction, target), our pipeline synthesizes a compatible reference image through type-specific generation strategies, then applies multi-dimensional quality filtering.

🖼Dataset Examples

RCEdit-500K covers six edit categories with both concrete references (add, replace, remove, background) and abstract references (style, alter). Click the tabs below to browse examples from each category.

Add

Replace

Remove

Background

Style

Alter

Add: Insert an object from the reference image into the input scene. The reference provides a concrete subject that appears in the edited target.

Replace: Substitute an existing object in the input with the one depicted in the reference image.

Remove: The reference image indicates which object to remove from the input scene.

Background: Change the background to match the environment/scene shown in the reference image.

Style: Transfer the artistic style from the reference to the input image. The reference provides abstract visual attributes.

Alter: Transfer visual attributes (color, texture, material) from the reference to the subject in the input image.

📈Quantitative Results

We evaluate on the MultiBanana two-reference benchmark using GPT-4o across five axes: Instruction Alignment (IA), Reference Consistency (RC), Background-Subject Match (BSM), Physical Realism (PR), and Visual Quality (VQ).

Model	IA ↑	RC ↑	BSM ↑	PR ↑	VQ ↑	Avg ↑
Closed-source models
Nano Banana	4.59	4.00	5.13	5.72	5.44	4.67
GPT-Image-1.5	5.87	4.94	5.40	5.89	5.85	5.51
Open-source models
Flux-Kontext	2.94	2.80	2.39	2.94	2.81	2.82
Omnigen2	3.62	3.28	4.50	5.17	5.05	3.93
Qwen-Image-Edit-2509	3.60	4.06	4.83	5.61	5.39	4.31
Dreamomni2	4.45	3.61	5.07	5.83	5.80	4.54
Qwen-Image-Edit-2511	4.15	4.18	4.95	5.68	5.59	4.58
Flux-Klein-4B	4.69	4.01	5.05	5.62	5.57	4.71
Flux-Klein-9B	4.78	4.09	5.01	5.61	5.50	4.75
+ RCEdit-500K (Ours)
Janus-Pro (baseline)	1.02	1.01	1.02	1.04	1.11	1.03
Janus-Pro + RCEdit-500K	4.65	3.81	4.52	5.04	4.91	4.43
Flux-Kontext (baseline)	2.94	2.80	2.39	2.94	2.81	2.82
Flux-Kontext + RCEdit-500K	3.52	3.42	4.76	5.50	5.27	4.04
Qwen-Edit-2509 (baseline)	3.60	4.06	4.83	5.61	5.39	4.31
Qwen-Edit-2509 + RCEdit-500K	4.08	4.03	5.11	5.86	5.72	4.56

Training on RCEdit-500K yields consistent gains: +1.22 avg for Flux-Kontext, +0.25 for Qwen-Edit-2509, and transforms Janus-Pro from 1.03 to 4.43.

🎨Qualitative Comparison

Visual comparison between base models and their RCEdit-500K fine-tuned counterparts across concrete-object edits (add, replace) and abstract-attribute edits (style, alter).

Qualitative comparison of different models. Models labeled "Ours" are trained with RCEdit-500K. After fine-tuning, instruction alignment improves substantially, copy-paste artifacts are reduced, and visual blending between input and reference is resolved.

🔬Ablation Study

We ablate two key designs: five-dimensional post-filtering and weak-instruction augmentation, and compare against existing ICIE datasets.

Variant	IA ↑	RC ↑	BSM ↑	PR ↑	VQ ↑	Avg ↑
w/o post-filtering	3.60	3.30	3.62	4.20	4.04	3.62
w/o weak instruction	3.67	3.35	3.84	4.50	4.27	3.74
RCEdit-500K (full)	3.52	3.42	4.76	5.50	5.27	4.04

Comparison with Existing ICIE Datasets

Training Data	IA ↑	RC ↑	BSM ↑	PR ↑	VQ ↑	Avg ↑
Baseline (no LoRA)	2.94	2.80	2.39	2.94	2.81	2.82
ImgEdit	3.05	2.83	2.58	3.10	2.92	2.92
AnyEdit (w/o structure)	3.02	2.65	2.10	2.57	2.44	2.68
AnyEdit	2.17	1.87	4.55	5.15	4.93	2.97
RCEdit-60K (Ours)	3.62	3.28	3.78	4.45	4.16	3.68
RCEdit-500K (Ours)	3.52	3.42	4.76	5.50	5.27	4.04

Even at matched scale (60K), RCEdit outperforms existing ICIE datasets by a large margin, and performance continues to scale with data size.

📖Citation

@inproceedings{zhang2026rcedit,
  title={RCEdit-500K: Reference Completion for Image-Conditioned Image Editing},
  author={Zhang, Jingxu and Kim, Daneul and Pan, Yueming and Chen, Dong and Qiu, Kai and Liu, Yang and Yang, Yifan and Dai, Qi and Sun, Xiaoyan and Luo, Chong},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}