ECCV 2026

SeGroS: Enhancing Alignment for Unified Multimodal Models via Semantically-Grounded Supervision

1Ewha Womans University, South Korea, 2Soongsil University, South Korea

Corresponding author

Teaser figure / intro video (placeholder)

Placeholder caption — one-line overview of what SeGroS does.

Abstract

Unified Multimodal Models (UMMs) have emerged as a promising paradigm that integrates multimodal understanding and generation within a unified modeling framework. However, current generative training paradigms suffer from inherent limitations. We present Semantically-Grounded Supervision (SeGroS), a fine-tuning framework designed to resolve the granularity mismatch and supervisory redundancy in UMMs. At its core, we propose a novel visual grounding map to construct two complementary supervision signals. First, we formulate semantic Visual Hints to compensate for the sparsity of text prompts. Second, we generate a semantically-grounded Corrupted Input to explicitly enhance the supervision of masking-based UMMs by restricting the reconstruction loss to core text-aligned regions. Extensive evaluations on GenEval, DPGBench, and CompBench demonstrate that SeGroS significantly improves generation fidelity and cross-modal alignment across various UMM architectures.

Why Semantically-Grounded Supervision?

Placeholder — describe the two problems you address (granularity mismatch and supervisory redundancy) and the intuition behind grounding supervision on text-aligned regions.

Motivation figure (placeholder)

Method

Framework / architecture figure (placeholder)

Placeholder — walk through the pipeline: the visual grounding map, the semantic Visual Hints, and the semantically-grounded Corrupted Input. Inline equations render via MathJax, e.g. $\mathcal{L} = \mathcal{L}_{\text{gen}} + \lambda\,\mathcal{L}_{\text{ground}}$.

Visual Hints
Visual Hints illustration (placeholder)
Semantically-Grounded Corrupted Input
Corrupted Input illustration (placeholder)

Results

Placeholder galleries — replace boxes with your qualitative results.

Text-to-Image Generation
Prompt: "placeholder text prompt goes here."
Generated image (placeholder)
Prompt: "another placeholder prompt."
Generated image (placeholder)
Comparison with Baselines

Baseline

Baseline output (placeholder)

SeGroS (Ours)

Our output (placeholder)
Quantitative Evaluation
MethodGenEval ↑DPGBench ↑CompBench ↑
Baseline UMM
+ SeGroS

Placeholder — fill in with your numbers.

BibTeX

@article{kim2026segros,
  title   = {Enhancing Alignment for Unified Multimodal Models via Semantically-Grounded Supervision},
  author  = {Kim, Jiyeong and So, Yerim and Choi, Hyesong and Hwang, Uiwon and Min, Dongbo},
  journal = {arXiv preprint arXiv:2603.19807},
  year    = {2026}
}