Multi-modal Synthetic Data Training and Model Collapse: Insights from VLMs and Diffusion Models

TL;DR

The paper investigates model collapse in multimodal synthetic-data training for vision-language models and diffusion models within recursive generate-train loops. It extends prior unimodal analyses to multi-agent, cross-modal settings, examining three training configurations (single-model recursion, relabeling with frozen models, and joint recursive finetuning) and a MSCOCO-based dataset. Key findings show that collapse manifests differently across modalities (variance can drop in diffusion models but increase in VLM captions; alignment can improve initially) and that strategies like decoding budgets, model diversity, and relabeling with frozen grounding models effectively mitigate collapse, while joint unfrozen training accelerates degradation. The work offers practical guidelines for curating robust multimodal synthetic data and stabilizing self-improving AI systems in realistic, multi-agent environments.

Abstract

Recent research has highlighted the risk of generative model collapse, where performance progressively degrades when continually trained on self-generated data. However, existing exploration on model collapse is limited to single, unimodal models, limiting our understanding in more realistic scenarios, such as diverse multi-modal AI agents interacting autonomously through synthetic data and continually evolving. We expand the synthetic data training and model collapse study to multi-modal vision-language generative systems, such as vision-language models (VLMs) and text-to-image diffusion models, as well as recursive generate-train loops with multiple models. We find that model collapse, previously observed in single-modality generative models, exhibits distinct characteristics in the multi-modal context, such as improved vision-language alignment and increased variance in VLM image-captioning task. Additionally, we find that general approaches such as increased decoding budgets, greater model diversity, and relabeling with frozen models can effectively mitigate model collapse. Our findings provide initial insights and practical guidelines for reducing the risk of model collapse in self-improving multi-agent AI systems and curating robust multi-modal synthetic datasets.

Figures (9)

• Figure 1: Recursive finetuning configurations for investigating model collapse: (1) Single model recursive finetuning ("T2I" stands for text-to-image diffusion model), (2) Finetuning with VLM or diffusion relabeling, and (3) Joint recursive finetuning. (1) allows independent study of the effect of recursive training on VLMs and text-to-image diffusion models. (2) enables information integration from another generative paradigm, which simulates a multi-agent environment where a frozen grounding model is available and used to label synthetic data. (3) simulates a multi-agent environment where all models have full freedom to update their weights using available synthetic data.
• Figure 2: Recursive finetuning shifts generated image properties. Mean (solid lines) and standard deviation (shaded area) are calculated from 5 groups of 200 evaluation samples. Two baselines are shown: "Gen 0 Finetune" stands for finetuning only on the synthetic data generated by the generation 0 (pretrained) model. "Real Finetune" stands for finetuning on the human-authored data. The most prominent effect of recursive finetuning compared to the other two baselines is the shifted saturation (1), reduced CLIP embedding variance (7), increased modality gaps (5, 6), and increased CLIP Score (10).
• Figure 3: Gender bias shift: In this example, we observe a shift towards females in tandem with saturation upshift.
• Figure 4: Gender bias shift: we see a shift to the female side after 10 generations of recursive finetuning on the diffusion model. After manually downshifting the saturation after each generation, we see a reduced gender bias shift, indicating that saturation drives the gender bias shift in recursive finetuning of the Stalbe Diffusion 1.4 model.
• Figure 5: Recursive finetuning effect on the generated caption properties: Mean and standard deviation are calculated from 10 groups of 200 evaluation images randomly selected from the COCO evaluation set. Two baselines are shown: "Gen 0 Finetune" stands for finetuning on the synthetic data by the base (Generation 0) model. "Real Finetune" stands for finetuning on the MSCOCO data. The most prominent effect of recursive finetuning compared to the other two baselines is the vocabulary size and perplexity explosion (2, 3).