AI Paper Atlas

Research map / Video / Video generation

Dual-IPO: Dual-Iterative Preference Optimization for Text-to-Video Generation

Xiaomeng Yang, Mengping Yang, GONG JIA, Luozheng Qin, Zhiyu Tan, Hao Li

ICLR 2026 · Video › Video generation · approach: Compositional text-to-video diffusion · 24 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
update the critic via SRPO during iterations (ablation)Dual-IPO · VBench (iteration 3 to 4)VBench score: 82.33 (no update) -> 82.91Improved
use SRPO loss instead of plain DPO loss for critic (ablation)SRPO · VBenchVBench score: 82.83 -> 82.91Improved
add preference certainty estimator (PCE) (ablation)SRPO · VBenchVBench score: 82.69 -> 82.91Improved
add real video data SFT to preference optimization (ablation)Dual-IPO · VBenchTotal score: 81.17 -> 81.69Improved
fine-tune on real video data only, without preference optimization (ablation)CogVideoX-2B · VBenchTotal score: 80.91 -> 80.63Hurt
use KTO instead of DPO (ablation)Dual-IPO · VBenchTotal score: 82.42 -> 82.47No clear effect
add more Dual-IPO iterations (1 to 3) (ablation)CogVideoX-2B · VBenchTotal score: 81.69 -> 82.42 -> 82.74Improved
one extra iteration beyond the reported ones (ablation)Wan-1.3B · VBenchTotal score: 88.32 -> 88.33No clear effect
use custom reward model instead of VideoScore/VideoAlign/VideoReward for DPO (ablation)CogVideoX-2B Diffusion-DPO · VBenchVBench score: 80.87 / 81.27 / 81.31 -> 81.54Improved
scale reward model from 13B to 40B (ablation)Dual-IPO reward model · VBenchTotal score: 82.42 -> 82.52Improved
apply Dual-IPO to CogVideoX-5BCogVideoX-5B · VBenchTotal score: 81.61 -> 84.63Improved
apply Dual-IPO to Wan-1.3BWan-1.3B · VBenchTotal score: 84.26 -> 86.28Improved

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.

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