TULIP: Token-length Upgraded CLIP

We thank Reviewer XgPN for the constructive feedback and insightful comments. We hope to address the concerns of the Reviewer with the responses below.

Clarification of the primary contributions. We propose the first contrastive vision-language model with relative positional encoding for long caption understanding. Our proposed approach is a substantial improvement over prior work and goes beyond a straightforward application of RoPE and CoPE. While RoPE and CoPE are established techniques, our primary contribution lies in designing a framework that can leverage any type of relative positional encodings to overcome the fundamental limitations of absolute positional encoding. Unlike Long-CLIP (ECCV 2024), which relies on absolute positional encodings combined with position stretching (fixing the first 20 tokens and interpolating positions for the rest up to 77 tokens), we propose a framework that learns pairwise relationships between any number of tokens, enabling a more generalizable long caption understanding. To further highlight the primary contribution of using relative encodings, we performed a new experiment to train our model with absolute positional encodings. We report results on long-caption retrieval tasks (top R@1) and observe significant performance drops, as detailed in the table below.

We hope this clarifies the scope and significance of our primary contributions, which we will better stress in the introduction of the paper.

Comparison to T5-based models. We acknowledge the strengths of T5-based text encoders in handling long text descriptions. However, T5-based models are not well-suited for vision-language tasks. In particular, T5’s text encoder is not inherently aligned with the image modality and would require considerable additional training to establish the image-text alignment - that is central to TULIP. We also note the Reviewer’s suggestion to compare with Muse, a T5-based image generation model. Unfortunately, this model is not publicly available. Therefore, we compare TULIP-based image generation with similar open-source T5-based models like PIXART-Alpha [a] and ELLA [b]. These results, included in the Appendix of the revised paper (Figures 10 and 11), demonstrate that TULIP effectively generates images with nuanced details while maintaining its cross-modal alignment without any costly retraining.

[a] Chen et al., “PIXART-Alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis” (ICLR 2024)

[b] Hu et al., “ELLA: Equip Diffusion Models with LLM for Enhanced Semantic Alignment” (arXiv 2024)

Theoretical analysis. RoPE builds upon a mathematical foundation for long-context modeling by its computation of token relationships in a translation-invariant manner. This is useful for handling long contexts where absolute positional information loses relevance. In existing literature, several works attempt to provide a better theoretical understanding of RoPE mechanisms [c, d] in language models. For instance, some of them explore the frequency spectrum of RoPE, showing that lower frequencies primarily encode semantic information while higher frequencies focus on positional relationships [c]. However, a general theoretical framework has not been fully established yet. We agree with the Reviewer that a theoretical analysis is important and we will provide a discussion on the points we raised above in the revised version of the paper.

[c] Barbero et al., Round and Round We Go! What makes Rotary Positional Encodings useful? (arXiv 2024)

[d] Ruscio et al., Beyond position: how rotary embeddings shape representations and memory in autoregressive transformers. (arXiv 2024)

We hope this addresses your concerns. We appreciate your consideration and hope this motivates you to upgrade your score. Please let us know if you have any further questions or feedback.

Dear Reviewer XgPN,

Thank you once again for taking the time and effort to review our submission. As the deadline for the discussion phase is approaching, we wanted to kindly check if you have any additional questions or concerns that we could address in our rebuttal.

We appreciate your insights and look forward to your feedback.

Best regards,

The Authors

Dear Reviewer XgPN,

Thank you for your continued encouragement and guidance.

Following your suggestions to compare against T5-based models, we have expanded the comparison presented in Figures 10 and 11 of the Appendix. We included additional text prompts and corresponding generated images using publicly available T5-based image generation models (PIXART-Alpha and ELLA), and SDXL + TULIP model. We provide the qualitative results in Figures 15 - 20 in the Appendix, where we can observe that our TULIP approach demonstrates overall better capability in handling nuanced descriptions. We used green image frames to highlight images that align correctly with the prompt and red frames to indicate those with missing elements. Also, note that our approach is plug-and-play for image generation, whereas the counterpart models require specific fine-tuning to perform this task. To further honor your request we will update Figure 3 in the main paper, showcasing qualitative results comparing the T5-based models to our model.

In addition, given the time constraints, we performed a preliminary small-scale human evaluation to provide a quantitative comparison as well. We performed two experiments to compare to T5-based models, first to compare SDXL + TULIP to PIXART-Alpha and second to compare to ELLA. For each experiment, we gathered 5 annotators and presented 20 text prompts each with two images generated by the corresponding models. The experimental setup is shown in Figure 14 of the Appendix. The annotators are asked to assess the prompt-image alignment, by paying attention to the nuanced details of the image such as object attributes, numeracy, relationships, etc. We observe promising results as follows: compared to PIXART-Alpha we achieved a 64% win rate and compared to ELLA we achieved a 72% win rate. Currently, we are working on conducting a larger-scale human evaluation to further improve the comprehensiveness of our results. However, as the deadline for updating the paper has already passed, we will include all the details in the final version of the paper.

Thank you once again for your valuable feedback. If you have any remaining questions or suggestions, we are most willing to address them.

Sincerely,

The Authors

We thank Reviewer FB4m for the constructive feedback and insightful comments. With the responses below, we hope to address the Reviewer's concerns.

Contributions, w.r.t prior work. A key difference of our approach w.r.t prior work, such as Long-CLIP (Zhang et al., ECCV 2024), is the use of relative positional encodings. Long-CLIP is performing stretching of the fixed absolute encodings, which is an ad-hoc solution since it is not scalable to longer contexts and is bounded to 248 tokens. Unlike them, our proposed solution is generalizable since we lift the constraint on fixed encodings and leverage the pairwise distance between tokens despite the length. In this manner, our approach can accommodate variable token lengths, ranging from the original 77 tokens to 300 tokens and beyond, as shown in Figure 4.

Comparison to autoregressive models. Our work aims to explore why CLIP, a basis for most foundation models, lags behind autoregressive models in long-context understanding. One reason why autoregressive models excel in this area is due to their use of relative positional encodings. We adapt these principles to CLIP and will clarify this focus further in the Introduction section.

Evaluation on DOCCI. We have performed the suggested experiments to evaluate our TULIP model on DOCCI, as well as the prior works CLIP and Long-CLIP. We provide the top R@1 results below and will add them to the paper. We observe that TULIP obtains the best performance, especially for img2txt retrieval, demonstrating its efficiency and applicability across one more benchmark.

Quantitative evaluation of image generation. We designed a pilot human evaluation by manually crafting captions and generating images by using Stable Diffusion XL (SDXL) with CLIP and TULIP-based text encoders. The resulting images were randomized, and annotators were tasked with selecting the image that was best aligned with the prompt. They were instructed to evaluate alignment based on objects, their attributes, and their relationships. A total of 15 annotators reviewed 20 samples. The win ratios are 89% for the SDXL + TULIP model and 11% for the SDXL + CLIP model, reflecting the enhanced alignment capability that we observed in the qualitative comparisons as well. We will add a discussion on this in the Experiment section of the paper.

Amount of data for TULIP student model. To analyze the needed amount of data with long context, we conduct experiments to train the student model using 33%, 66%, and the entire ShareGPT4V dataset (1.2 million image-caption pairs). The results on the retrieval tasks (top R@1) shown in the table below, emphasize the importance of utilizing the full dataset, particularly for improving the performance on out-of-distribution data i.e. Long-DCI and Urban1k. These findings will be included in the Experiments section of the paper.

We hope this addresses your concerns. Please let us know if you have any further questions or feedback.

We thank the Reviewer for the engagement and suggestions. Below we provide our answers:

[Q1] For the human evaluation, the instructions given to the annotators are as follows:

“Instructions: You are provided with 20 samples, each containing a prompt and two images generated by two different Stable Diffusion models. Your task is to compare the two images given the input prompt and choose the one you find more aligned with the prompt. Note that the differences might be very nuanced in some cases. Pay attention to the object attributes, relationships, verbs, object counting, etc.”

We added this in a separate section to the Appendix, as well as two samples included in the human evaluation experiment. Regarding the win rates, indeed they are calculated across annotations instances. We agree that calculating win rates across samples (defined by majority agreement) would provide a valuable perspective on the performance of our model. While it may not be feasible due to limited time, we plan to conduct an additional human study incorporating your suggestions to improve the robustness and interpretability of our results.

[Q2] In the ablation for the amount of data, we indeed trained the models on 1 pass of the dataset to ensure a fair comparison.

We hope this addresses your questions and welcome further comments or suggestions.

Dear Reviewer FB4m,

Thank you once again for your encouragement and thoughtful feedback. As the discussion deadline approaches, we wanted to follow up on our recent response to ensure we have addressed all your concerns. If there is anything further that might convince you of the value of our work, we would be happy to provide additional clarifications.

Sincerely,

The Authors

We thank Reviewer 3TVK for the constructive feedback and acknowledgment. We hope to address all comments below and will add the new experiments to the Experiments section of the paper.

Retaining fine-grained relationships in shorter contexts. Based on prior works (e.g. Llama 3, Gemma) and our empirical results, we believe that there is no evidence that switching from absolute to relative positional encodings poses a challenge in dealing with shorter contexts. Table 2 of the paper shows that TULIP not only retained the original performance of CLIP but even improved it by 6.1% R@1 on image-to-text retrieval and 10.7% R@1 on text-to-image retrieval on COCO. Similarly, we observed 8.2% R@1 on image-to-text retrieval and 13.6% R@1 on text-to-image retrieval on the Flickr dataset.

Details about human annotators. The Long-DCI benchmark is derived from the dataset introduced in the paper “A Picture is Worth More Than 77 Text Tokens: Evaluating CLIP-Style Models on Dense Captions”, by Urbanek et al. (CVPR 2024). All human annotations were provided in this dataset, by using Mephisto (Urbanek and Ringshia, 2023) to host the task and pay crowdworkers to annotate the images, as detailed in Section 3.2 of their paper. Unlike their work, which uses only captions truncated to 77 tokens, we leverage their full annotations to construct a long-caption retrieval benchmark.

Qualitative comparison for short context. To compare our TULIP model to CLIP, we perform a new qualitative analysis for short caption retrieval and present the results in the Appendix (Figures 12 and 13). We observe that our TULIP can also capture fine-grained details in short captions, such as object numeracy, relationships, attributes, etc. This is attributed to using relative positional encodings, which model pairwise token relationships unlike absolute ones, improving its understanding of complex details even in shorter contexts. Note that in the figure, the misclassified concepts are marked in red and the correct ones are in green.

Discussion on perceptual understanding and compositionality. Indeed CLIP has challenges in perceptual understanding such as spatial understanding, numeracy, and compositionality of objects and attributes. To better evaluate TULIP's capabilities in these areas—specifically in handling compositionality in text inputs—we considered the two benchmarks suggested by the Reviewer (CompPrompt and ControlledImCaps). However, these were unfortunately not publicly available. Nevertheless, we turned to two well-established baselines for multimodal perceptual understanding and compositionality: ARO (Yuksekgonul et al. (ICLR 2023)) and VL-Checklist (Zhao et al. (EMNLP 2022)). The results, shown below, indicate that TULIP consistently outperforms CLIP and surpasses Long-CLIP in 4 out of 7 settings.

Improved performance in shorter contexts (Table 2). We attribute the improved performance on short-caption retrieval to using relative positional encodings. This allows TULIP to dynamically encode token relationships by modeling their dependencies without being constrained by fixed positions. This benefits both short and long contexts. To validate this, we perform an additional experiment to train TULIP with absolute encodings and compare it to its original version where we use relative ones. We present the top R@1 results below, where it can be observed that relative encodings indeed yield much better performance in short-caption retrieval for all four settings.

Results in Table 4 compared to Tables 1 and 3. Table 4 reflects the performance of the distilled models before the relative position expansion phase, whereas Tables 1 and 3 present results after full training is completed. This distinction was intentional, as Table 4 specifically aims to isolate and analyze the relative position distillation stage, allowing us to evaluate the effectiveness of different distillation losses. To address the potential confusion, we will revise the caption of Table 4 to clarify this point.

We hope this addresses your concerns. We appreciate your consideration and hope this motivates you to upgrade your score. Please let us know if you have any further questions.