Expedited Training of Visual Conditioned Language Generation via Redundancy Reduction

TL;DR

This work tackles the heavy computational demands of vision-language pre-training by introducing EVL_Gen, a one-stage framework that replaces BLIP-2's Stage-1 with TomeFormer-based token merging to condense visual information into semantic soft prompts for a frozen LLM. It adds a Temporal Attentive Soft Token Contextualizing module to extend image-language models to video without realignment, enabling effective video-captioning even without large-scale video pre-training. Empirically, EVL_Gen delivers competitive or superior performance on image-text benchmarks while achieving roughly fivefold faster training and performing well with significantly smaller data budgets (as low as 1/10 of the data in some settings). The approach preserves the ability to explore different ViTs and scales to video tasks, expanding accessible research on efficient vision-language generation and reducing resource barriers for researchers. Overall, EVL_Gen demonstrates that end-to-end vision-language generation can be learned efficiently with a single objective and a token-merging connector, maintaining strong performance and enabling broader applicability to video understanding.

Abstract

In this paper, we introduce $\text{EVL}_{\text{Gen}}$, a streamlined framework designed for the pre-training of visually conditioned language generation models with high computational demands, utilizing frozen pre-trained large language models (LLMs). The conventional approach in vision-language pre-training (VLP) typically involves a two-stage optimization process: an initial resource-intensive phase dedicated to general-purpose vision-language representation learning, focused on extracting and consolidating relevant visual features. This is followed by a subsequent phase that emphasizes end-to-end alignment between visual and linguistic modalities. Our novel one-stage, single-loss framework bypasses the computationally demanding first training stage by gradually merging similar visual tokens during training, while avoiding model collapse caused by single-stage training of BLIP-2 type models. The gradual merging process effectively condenses visual information while preserving semantic richness, resulting in rapid convergence without compromising performance. Our experimental findings demonstrate that our approach accelerates the training of vision-language models by a factor of 5 without a noticeable impact on overall performance. Furthermore, we illustrate that our models significantly narrow the performance gap to current vision-language models using only 1/10 of the data. Finally, we showcase how our image-text models can seamlessly adapt to video-conditioned language generation tasks through novel soft attentive temporal token contextualizing modules. Code is available at \url{https://github.com/yiren-jian/EVLGen}.

Figures (5)

• Figure 1: Overview of our $\text{EVL}_{\text{Gen}}$. $\text{EVL}_{\text{Gen}}$ employs a streamlined, single-stage training mechanism with a unified loss. Here, visual tokens (in grey) are progressively aggregated based on their inherent similarities at each layer of the TomeFormer architecture. The final set of merged tokens (in orange) serves as semantically rich but computationally efficient soft prompts, guiding the LLM to generate a corresponding caption for the input image.
• Figure 2: Overview of $\text{EVL}_{\text{Gen}}$-Video: In addition to TomeFormer's spatial token merging capabilities, our design introduces Temporal Attentive Soft Token Contextualizing for nuanced temporal modeling. Each frame's output is calculated as a learnable weighted average of other frames in the video. This approach maintains the integration of pre-existing, well-trained image-text models. For instance, when the input consists of static videos with identical frames, $\text{EVL}_{\text{Gen}}$-Video operates as if it were an image-text model. Importantly, this architecture avoids the need for complex modality realignment, a requirement in alternative designs that insert a temporal Q-former between the visual encoder and the language model. It also significantly enriches the shared semantic information distributed among these frame tokens, laying the groundwork for more efficient token merging in future spatial merging steps.
• Figure 3: Trade-off between MSCOCO captioning scores (depicted in red) and GPU training time (depicted in blue) as a function of the number of tokens merged ($r$) in TomeFormer.
• Figure 4: Pre- and post-training visualization of merged tokens in $\text{EVL}_{\text{Gen}}$. The visual features compressed via token merging exhibit semantic informativeness even prior to training. This inherent characteristic facilitates $\text{EVL}_{\text{Gen}}$'s ability to converge quickly in an end-to-end training setup.
• Figure 5: Additional pre- and post-training visualization of merged tokens in $\text{EVL}_{\text{Gen}}$.