AI Paper Atlas

Research map / Image generation / Text-to-image diffusion

The Path to Reconciling Quality and Safety Alignment in Text-to-Image Generation

Shouwei Ruan, Zhenyu Wu, Yao Huang, Ruochen Zhang, Yitong Sun, Caixin Kang, Shiji Zhao, Weijun Qin, Jingzhi Li, Xingxing Wei

ECCV 2026 · Image generation › Text-to-image diffusion · approach: T2I evaluation and controllable adapters · 0 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
raise safety cost weight lambda from 0.01 to 0.5 in the composite reward (ablation)T2I-SPO on SD-v1.5 · I2P and COCO-30KIP and CLIPScore: IP 0.263 -> 0.207; CLIPScore 26.786 -> 25.643Mixed
add Dynamic Focusing Mechanism (DFM) (ablation)T2I-SPO · SD-v1.5 trainingtraining loss convergenceImproved
use all 4 DFM augmentation components instead of 1 (ablation)T2I-SPO DFM · SD-v1.5 trainingloss at epoch 20: 0.3737 -> 0.3593Improved
use all 4 DFM augmentation components instead of 2 (ablation)T2I-SPO DFM · SD-v1.5 trainingloss at epoch 20: 0.3676 -> 0.3593Improved
use composite quality plus safety reward instead of safety-only reward (T2I-SPO vs AlignGuard)DPO-based alignment · I2P-SexualUAScore: 0.651 -> 0.757Improved
T2I-SPO versus AlignGuardSD-v1.5 · I2P-SexualIP (%): 10.53 -> 4.40Improved
T2I-SPO versus AlignGuardSD-v1.5 · I2P (7 categories)IP (%): 13.18 -> 17.07Hurt
T2I-SPO versus AlignGuardSD-v1.5 · I2P (7 categories)UAScore: 0.690 -> 0.784Improved
T2I-SPO versus RECESD-v1.5 · NSFW-56K-SexualIP (%): 13.70 -> 13.65No clear effect
T2I-SPO versus unaligned SD-v1.5SD-v1.5 · I2P-SexualNudeNet total exposed parts: 485 -> 43Improved
T2I-SPO versus UCE under SneakyPrompt jailbreakSD-v1.5 · SneakyPromptIP (%): 12.0 -> 5.5Improved
T2I-SPO versus ESD-u under MMA jailbreakSD-v1.5 · MMAIP (%): 33.8 -> 30.2Improved

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