# Unlocking Compositional Control: Self-Supervision for LVLM-Based Image Generation

arXiv: https://arxiv.org/abs/2507.04151 · HTML: https://mlsanity.com/paper/2507.04151/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 23 of its 29 references that are in the map.
- First posted on arXiv 5 Jul 2025.

## Builds on (most cited first, 23 of 23)

- [Learning Transferable Visual Models From Natural Language Supervision](https://mlsanity.com/paper/2103.00020/) (2021)
- [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://mlsanity.com/paper/2010.11929/) (2020)
- [Microsoft COCO: Common Objects in Context](https://mlsanity.com/paper/1405.0312/) (2014)
- [Diffusion Models Beat GANs on Image Synthesis](https://mlsanity.com/paper/2105.05233/) (2021)
- [Hierarchical Text-Conditional Image Generation with CLIP Latents](https://mlsanity.com/paper/2204.06125/) (2022)
- [BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models](https://mlsanity.com/paper/2301.12597/) (2023)
- [Flamingo: a Visual Language Model for Few-Shot Learning](https://mlsanity.com/paper/2204.14198/) (2022)
- [BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation](https://mlsanity.com/paper/2201.12086/) (2022)
- [GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models](https://mlsanity.com/paper/2112.10741/) (2021)
- [Taming Transformers for High-Resolution Image Synthesis](https://mlsanity.com/paper/2012.09841/) (2020)
- [LAION-400M: Open Dataset of CLIP-Filtered 400 Million Image-Text Pairs](https://mlsanity.com/paper/2111.02114/) (2021)
- [CoCa: Contrastive Captioners are Image-Text Foundation Models](https://mlsanity.com/paper/2205.01917/) (2022)
- [Visual In-Context Learning for Large Vision-Language Models](https://mlsanity.com/paper/2402.11574/) (2024)
- [Language Models with Image Descriptors are Strong Few-Shot Video-Language Learners](https://mlsanity.com/paper/2205.10747/) (2022)
- [Enhancing Code LLMs with Reinforcement Learning in Code Generation: A Survey](https://mlsanity.com/paper/2412.20367/) (2024)
- [SCORE: Story Coherence and Retrieval Enhancement for AI Narratives](https://mlsanity.com/paper/2503.23512/) (2025)
- [MAM: Modular Multi-Agent Framework for Multi-Modal Medical Diagnosis via Role-Specialized Collaboration](https://mlsanity.com/paper/2506.19835/) (2025)
- [Draw ALL Your Imagine: A Holistic Benchmark and Agent Framework for Complex Instruction-based Image Generation](https://mlsanity.com/paper/2505.24787/) (2025)
- [Improving Compositional Text-to-image Generation with Large Vision-Language Models](https://mlsanity.com/paper/2310.06311/) (2023)
- [Multimodal Event Transformer for Image-guided Story Ending Generation](https://mlsanity.com/paper/2301.11357/) (2023)
- [Beyond Empathy: Integrating Diagnostic and Therapeutic Reasoning with Large Language Models for Mental Health Counseling](https://mlsanity.com/paper/2505.15715/) (2025)
- [Self-Rewarding Large Vision-Language Models for Optimizing Prompts in Text-to-Image Generation](https://mlsanity.com/paper/2505.16763/) (2025)
- [EvolveDirector: Approaching Advanced Text-to-Image Generation with Large Vision-Language Models](https://mlsanity.com/paper/2410.07133/) (2024)

## Abstract (arXiv)

This paper introduces Hierarchical Self-Supervised LVLM (Hi-SSLVLM), a novel generative model designed to significantly advance text-to-image synthesis, particularly for complex and compositionally challenging prompts. Traditional methods often grapple with the high cost of meticulously curated paired image-text datasets and struggle with precise control over fine-grained visual attributes and intricate spatial relationships. Our Hi-SSLVLM addresses these limitations through a unique two-stage self-supervised learning strategy. The first stage, Multi-Granularity Visual-Language Grounding, enables the Large Vision-Language Model (LVLM) backbone to autonomously generate and align hierarchical captions (global and local) to images, cultivating a deep internal semantic understanding without reliance on extensive human annotation. The second stage, Self-Refinement and Guided Image Generation, leverages this acquired knowledge by an Internal Compositional Planning (ICP) mechanism, where the LVLM first formulates detailed textual sub-prompts to guide the image generation process, complemented by a novel Semantic Consistency Loss for precise output alignment. Comprehensive experiments against leading baselines, including Janus-Pro-1B, Stable Diffusion XL 1.0, DeepFloyd IF v1.0, and ControlNet-XL, on multi-dimensional benchmarks such as Gemini-2.0-Flash and InternVL3-78B, demonstrate Hi-SSLVLM's superior performance across all fine-grained metrics. An in-depth ablation study confirms the critical role of each proposed component. Furthermore, human evaluations corroborate our quantitative findings, highlighting Hi-SSLVLM's enhanced fidelity to prompt, compositional accuracy, and overall aesthetic quality, marking a significant step towards more controllable and semantically consistent open-ended text-to-image generation.