AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Omni-IML: Towards Unified Interpretable Image Manipulation Localization

Chenfan Qu, Yiwu Zhong, Fengjun Guo, Lianwen Jin

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs

Read the paperCheck your idea against 137,217 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 Omni-IML, a generalist model that unifies image manipulation localization across multiple IML tasks to avoid the performance degradation seen when jointly training task-specific IML methods.

Problem

Existing image manipulation localization (IML) methods rely on task-specific designs and degrade substantially when jointly trained across multiple IML tasks, limiting real-world use.

Approach

A single generalist IML model designed to generalize across diverse manipulation localization tasks, positioned as the first such unified model; the abstract is truncated so the specific mechanism is not stated.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
joint training existing IML methods on all four tasks versus a single task (ablation)TIFDM (document IML) · Document IMLpixel IoU: .498 -> .428Hurt
joint training Omni-IML on all four tasks versus document only (ablation)Omni-IML · Document IMLpixel IoU: .774 -> .766No clear effect
add Modal Gate (versus fused frequency-vision features without gate) (ablation)Omni-IML · All four tasks (average)pixel IoU: .629 -> .728Improved
add Modal Gate (versus pure vision modality without gate) (ablation)Omni-IML · All four tasks (average)pixel IoU: .659 -> .728Improved
add Dynamic Weight Decoder (ablation)Omni-IML · All four tasks (average)pixel IoU: .608 -> .728Improved
make decoder filter weights input-adaptive instead of fixed (ablation)Dynamic Weight Decoder · All four tasks (average)pixel IoU: .688 -> .728Improved
add Anomaly Enhancement with box supervision (ablation)Omni-IML · All four tasks (average)pixel IoU: .682 -> .728Improved
add all three proposed modules to the baseline (ablation)Baseline · All four tasks (average)pixel IoU: .575 -> .728Improved
feed a visual reference prompt (mask-highlighted image) to the MLLMQwen2.5-VL 7B · Omni-273k documents: tampered text recognitionOCR accuracy: .312 -> .653Improved
feed a visual reference prompt to the MLLMQwen2.5-VL 7B · Omni-273k interpretation (average)average score: .512 -> .603Improved
visual reference prompt on face and natural image absolute position/descriptionInternVL3 2B · Omni-273k natural image: tampered object recognitionaccuracy: .778 -> .770 (face Obj. Rec. .757 -> .770)No clear effect
Omni-IML versus best prior on DocTamper SCD and CASIA1Omni-IML · CASIA1 and DT-SCDpixel IoU: CASIA1 .800 vs .810; DT-SCD .676 vs .691Mixed

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Surprising findings

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy