MemoGen: Can Past Experience Improve Future Text-to-Image Generation?
MemoGen enhances image generation via experience memory, surpassing Nano Banana Pro and GPT-Image-1.
Key Findings
Methodology
MemoGen is a training-free framework that augments existing image generators with an agentic evolution layer. For each task, MemoGen infers visual requirements, retrieves external evidence when necessary, translates them into executable generation constraints, evaluates the generated result, and stores task understanding, reference choices, visual feedback, successful strategies, and failure lessons as reusable experience memory.
Key Results
- On the WISE benchmark, MemoGen achieved an overall score of 0.91, surpassing Nano Banana Pro's 0.87 and GPT-Image-1's 0.80.
- On Mind-Bench, MemoGen reached an overall score of 0.52, significantly exceeding Nano Banana Pro's 0.41 and Mind-Brush's 0.31.
- After two evolution rounds, MemoGen demonstrated superior performance on knowledge-intensive and reasoning-oriented benchmarks.
Significance
MemoGen addresses the lack of continual learning signals in existing text-to-image generation systems by introducing experience memory. It improves generation quality at test time without updating generator parameters, offering new research directions for knowledge-driven and reasoning-driven image generation.
Technical Contribution
MemoGen's technical contribution lies in its training-free continual learning framework, enhancing generation performance through experience memory and agentic evolution. Unlike existing methods, it improves without relying on model parameter updates, using experience accumulation and retrieval.
Novelty
MemoGen is the first to introduce experience memory into text-to-image generation, allowing systems to improve generation quality at test time through experience accumulation and retrieval without updating the generator. This fundamentally differs from existing single-task enhancement methods.
Limitations
- MemoGen may perform poorly on entirely novel tasks due to a lack of relevant historical experience.
- It relies on the quality and coverage of external knowledge bases, which may lead to inconsistent results.
Future Work
Future work could explore applying MemoGen to a wider range of tasks and datasets, further enhancing its adaptability and performance across different domains.
AI Executive Summary
Modern text-to-image generation models have made significant strides in visual synthesis but remain unreliable when dealing with implicit visual constraints, relational reasoning, or external knowledge. MemoGen addresses this by introducing experience memory and an agentic evolution layer, improving generation quality at test time without updating generator parameters. Experimental results show that MemoGen excels on knowledge-intensive and reasoning-oriented benchmarks, surpassing strong proprietary systems. MemoGen's success demonstrates the potential of explicit experience memory as a powerful continual learning signal for reliable text-to-image generation.
Deep Analysis
Background
Text-to-image generation has rapidly advanced in recent years, driven by diffusion and autoregressive models. However, existing systems still face challenges in handling complex visual tasks, especially those involving implicit constraints and external knowledge.
Core Problem
Existing text-to-image generation systems typically treat each generation as an isolated episode, failing to systematically preserve past successes or failures for future use. This limits the reliability and adaptability of systems when handling complex tasks.
Innovation
MemoGen introduces experience memory and an agentic evolution layer, achieving training-free continual learning. It improves generation quality at test time through experience accumulation and retrieval without updating generator parameters.
Methodology
- �� Infer visual requirements
- �� Retrieve external evidence
- �� Translate into generation constraints
- �� Evaluate generated results
- �� Store experience memory
Experiments
Experiments were conducted on knowledge-intensive and reasoning-oriented benchmarks, including WISE and Mind-Bench. Using the open-source Qwen-Image as a backbone, MemoGen's performance was compared against Nano Banana Pro and GPT-Image-1.
Results
MemoGen achieved an overall score of 0.91 on the WISE benchmark, surpassing Nano Banana Pro and GPT-Image-1. On Mind-Bench, MemoGen reached an overall score of 0.52, significantly outperforming competitors.
Applications
MemoGen can be applied to applications requiring complex visual reasoning, such as cultural scene generation and scientific phenomenon visualization. Its training-free nature makes it easy to integrate into existing systems.
Limitations & Outlook
MemoGen may perform poorly on entirely novel tasks and relies on the quality and coverage of external knowledge bases. Future improvements could address these issues by expanding knowledge bases and improving experience retrieval mechanisms.
Plain Language Accessible to non-experts
Imagine you're cooking a complex dish in the kitchen. After each attempt, you note down what worked and what didn't, like which spices paired well and which steps need improvement. Next time you cook this dish, you can refer to these notes to avoid past mistakes and create a tastier meal. MemoGen acts like such a kitchen assistant, helping image generation systems improve through accumulated experience.
ELI14 Explained like you're 14
Imagine you're playing a game, and after each level, you jot down which strategies worked and which areas need improvement. Next time you play, you use these notes to level up faster. MemoGen is like that game helper, assisting image generation systems to improve quality through accumulated experience.
Glossary
MemoGen
A training-free framework that enhances image generation quality through experience memory and an agentic evolution layer.
Used to enhance existing image generators' performance.
Experience Memory
Reusable memory storing task understanding, reference choices, visual feedback, successful strategies, and failure lessons.
Used to retrieve and improve generation conditions in future tasks.
Agentic Evolution Layer
An evolution mechanism that improves generation quality through experience accumulation and retrieval without updating generator parameters.
Enables test-time self-evolution.
Qwen-Image
An open-source image generation model serving as MemoGen's backbone.
Used for performance comparison against other proprietary systems.
WISE Benchmark
A benchmark for evaluating the performance of knowledge-intensive image generation systems.
Used to test MemoGen's performance in knowledge-driven tasks.
Open Questions Unanswered questions from this research
- 1 How to effectively apply MemoGen to entirely novel tasks remains to be explored.
- 2 Expanding knowledge bases to enhance MemoGen's adaptability and coverage.
Applications
Immediate Applications
Cultural Scene Generation
MemoGen can generate complex cultural scenes, aiding artists and designers in creation.
Long-term Vision
Scientific Phenomenon Visualization
Through experience memory, MemoGen can be used for complex phenomenon visualization in scientific education and research.
Abstract
Modern text-to-image models have achieved strong visual synthesis, yet remain unreliable when prompts require implicit visual constraints, relational reasoning, or external knowledge. Existing retrieval-augmented and agentic generation methods mitigate this issue by acquiring external knowledge, references, or refined prompts for the current request, yet they typically treat each generation as an isolated episode and do not systematically preserve past successes or failures for future use. In this work, we ask whether a text-to-image system can continually improve from its own generation experience without updating the underlying generator. We propose MemoGen, a training-free framework that augments existing image generators with an agentic evolution layer. For each task, MemoGen explicitly infers visual requirements, retrieves external evidence and references when necessary, translates them into executable generation constraints, evaluates the generated result, and stores task understanding, reference choices, visual feedback, successful strategies, and failure lessons as reusable experience memory. Across evolution rounds, the agent retrieves relevant experience to improve similar future generations, selectively repairing previously failed cases while preserving successful ones, thereby enabling test-time self-evolution without parameter updates. Extensive experiments on knowledge-intensive and reasoning-oriented benchmarks demonstrate the effectiveness of this paradigm: after only two evolution rounds, MemoGen built upon the open-source Qwen-Image backbone surpasses strong proprietary systems such as Nano Banana Pro and GPT-Image-1 on WISE and Mind-Bench, showing that explicit experience memory can serve as a powerful continual learning signal for reliable text-to-image generation.