AI Paper Atlas

Research map / Image generation / Autoregressive and latent generation

RiSE: Residual Subspace Expert for Generalizable Text-Centric Image Forgery Localization

Kahim Wong, Kemou Li, Yiming Chen, Haiwei Wu, Jiantao Zhou

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: Generalizable fake image detection · 0 citations (OpenAlex)

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 RiSE, a residual subspace expert approach for text-centric image forgery localization that aims to generalize to unseen forgery types better than full-parameter fine-tuning.

Problem

Text-centric image forgery localization (TFL) detectors rely on full-parameter fine-tuning, which causes low-rank feature collapse and overfitting to training artifacts, hurting generalization to unseen forgery types.

Approach

Uses a residual subspace expert formulation as an alternative to full-parameter fine-tuning to preserve feature rank and reduce overfitting to training artifacts; details beyond the abstract are not available.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace full-parameter fine-tuning with residual subspace adaptation (ablation)RiSE-9 backbone · Avg F1 (7 test sets)Avg F1: 0.353 -> 0.625Improved
increase number of residual experts from 1 to 9 (ablation)RiSE · Cross-domain avg F1AvgC F1: 0.542 -> 0.611Improved
increase queries from 1 to 10 or 20 (ablation)RiSE-9 · Avg F1 (7 test sets)Avg F1: 0.625 -> 0.418 (10 queries), 0.469 (20 queries)Hurt
make the mask module shared across experts (ablation)RiSE-9 · Avg F1 (7 test sets)Avg F1: 0.579 -> 0.625Improved
use offline expert assignment instead of learned MoE routing (ablation)RiSE-9 · Avg F1 (7 test sets)Avg F1: 0.603 -> 0.625 (standard MoE), 0.541 -> 0.625 (two-stage MoE)Improved
adapt both attention and MLP layers rather than one (ablation)RiSE-9 · Avg F1 (7 test sets)Avg F1: 0.590 (attention only), 0.612 (MLP only) -> 0.625Improved
use higher residual rank (4, 8, 16) instead of rank 1 (ablation)RiSE-9 · Avg F1 (7 test sets)Avg F1: 0.625 -> 0.624 / 0.616 / 0.621No clear effect
use LoRA-style adaptation instead of SVD residual adaptation (ablation)RiSE-9 · Avg F1 (7 test sets)Avg F1: 0.625 -> 0.617 to 0.623Hurt
always fuse DCT features during training (p_g=1) rather than random gating (ablation)RiSE-9 with DCT · Cross-domain avg F1AvgC F1: 0.596 -> 0.519Hurt
add random-gated DCT fusion (p_g=0.5) (ablation)RiSE-9 · In-domain avg F1AvgI F1: 0.663 -> 0.690Improved
add DCT features to RiSE-9RiSE-9 · Cross-domain avg F1AvgC F1: 0.611 -> 0.616No clear effect
use latent perturbation confidence (LPC) instead of SDC for expert selection (ablation)RiSE-9 · Cross-domain avg F1AvgC F1: 0.534 -> 0.611Improved

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.

Experimental setup

Baselines: full-parameter fine-tuning TFL detectors

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