Research map / Image generation / Editing and style transfer

FunEditor: Achieving Complex Image Edits via Function Aggregation with Diffusion Models

AAAI 2025 · Image generation › Editing and style transfer · approach: Inversion-based diffusion editing · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FunEditor, an InstructPix2Pix-based diffusion editor that learns atomic editing task tokens with cross-attention masking and aggregates them to perform complex edits like object movement and pasting in 4 LCM steps.

Problem

Complex multi-edit image editing (e.g., object movement, pasting) with diffusion models is slow due to sequential editing or inference-time energy optimization, and textual prompts often cause unintended changes outside the edit region.

Approach

Learn trainable task tokens for atomic localized edit functions (object removal, edge enhancement, harmonization), enforce locality with cross-attention masking, and aggregate tokens at inference so multiple functions run simultaneously in few steps, combined with pixel-space copy-paste.

Key results

• On COCOEE object movement, FunEditor reaches LPIPS 0.017 and PSNR 34.27 for object consistency vs. 0.034/28.59 for the best baseline DragonDiffusion.
• On ReS, FunEditor achieves object-consistency PSNR 35.61 and CLIP-I2I 0.942, outperforming all baselines on those metrics.
• With 4 steps (one V100 latency ~1s), FunEditor is reported as 5-24x faster than baselines, e.g., 23 s faster than DiffEditor's 50-step pipeline.

Measured effects reported in the paper

Surprising findings

• Paint-by-Example attains the highest TOPIQ (0.717) and LIQE (4.75) on ReS, slightly above FunEditor on those IQA metrics.

Limitations

• Only object movement and object pasting evaluated as complex tasks; limited set of atomic functions.
• Relies on predefined masks and pixel-space copy-paste components.
• Some IQA metrics (TOPIQ, LIQE on ReS) are matched or beaten by baselines such as Paint-by-Example.

Experimental setup

Datasets: InstructPix2Pix editing dataset (basis) atomic edit training datasets (details in appendix)

Benchmarks: COCOEE ReS

Metrics: TOPIQ MUSIQ LIQE LPIPS PSNR CLIP-I2I NFEs latency

Baselines: SelfGuidance DragonDiffusion DiffEditor Paint-by-Example AnyDoor

Builds on: InstructPix2Pix EmuEdit Latent Consistency Model Prompt-to-Prompt

Closest papers by meaning

• PAIR Diffusion: A Comprehensive Multimodal Object-Level Image Editor (CVPR 2024 · similarity 0.89)
• h-Edit: Effective and Flexible Diffusion-Based Editing via Doob's h-Transform (CVPR 2025 · similarity 0.89)
• Image Editing As Programs with Diffusion Models (NeurIPS 2025 · similarity 0.88)
• ImgEdit: A Unified Image Editing Dataset and Benchmark (NeurIPS 2025 · similarity 0.88)
• An Item Is Worth a Prompt: Versatile Image Editing with Disentangled Control (AAAI 2025 · similarity 0.88)
• High-Fidelity Diffusion-Based Image Editing (AAAI 2024 · similarity 0.87)