Divide, Evaluate, and Refine: Evaluating and Improving Text-to-Image Alignment with Iterative VQA Feedback

The field of text-conditioned image generation has made unparalleled progress with the recent advent of latent diffusion models. While remarkable, as the complexity of given text input increases, the state-of-the-art diffusion models may still fail in generating images which accurately convey the semantics of the given prompt. Furthermore, it has been observed that such misalignments are often left undetected by pretrained multi-modal models such as CLIP. To address these problems, in this paper we explore a simple yet effective decompositional approach towards both evaluation and improvement of text-to-image alignment. In particular, we first introduce a Decompositional-Alignment-Score which given a complex prompt decomposes it into a set of disjoint assertions. The alignment of each assertion with generated images is then measured using a VQA model. Finally, alignment scores for different assertions are combined aposteriori to give the final text-to-image alignment score. Experimental analysis reveals that the proposed alignment metric shows significantly higher correlation with human ratings as opposed to traditional CLIP, BLIP scores. Furthermore, we also find that the assertion level alignment scores provide a useful feedback which can then be used in a simple iterative procedure to gradually increase the expression of different assertions in the final image outputs. Human user studies indicate that the proposed approach surpasses previous state-of-the-art by 8.7% in overall text-to-image alignment accuracy. Project page for our paper is available at https://1jsingh.github.io/divide-evaluate-and-refine

Paper

References (48)

Scroll for more · 36 remaining

Similar papers

Peer review

Reviewer 5iW37/10 · confidence 3/52023-07-04

Summary

This paper proposes a decompositional approach to evaluate and improve a pre-trained text-to-image diffusion model given complex input prompts describing multiple objects and novel combinations. To that end, a LLM decomposes a difficult caption into disjoint assertions that can be individually evaluated using a VQA model. The assertion scores can be used to a) evaluate fine-grained image-text alignment and b) provide feedback for iterative refinement of generated images. The proposed metric correlates stronger with human ratings and the iterative refinement procedure increases the fidelity of generated images on complex input prompts.

Strengths

- The paper is very well written and the method tackles an important problem - The proposed DA-score has clear motivation and correlates much better with human ratings - Using a LLM to produce a diverse pool of input prompts across varying complexity in terms of participating objects and realism is novel - Quantitative and visual results are strong. The resulting model produces images that align a lot better on unreal/creative input captions.

Weaknesses

I don't have any substantial weaknesses but including early approaches to combining text-to-image with VQA models would enhance the comprehensiveness of the related work section. - https://dl.acm.org/doi/abs/10.1145/3372278.3390684 - https://aclanthology.org/2020.lantern-1.2/

Questions

- TIFA is a recent, similar looking approach that uses VQA for evaluation. How does it differ?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

- Yes, authors have extensively discussed limitations of the proposed method in the supplementary material.

Reviewer bVCd7/10 · confidence 4/52023-07-04

Summary

This paper makes a twofold contribution: namely introducing a new metric for evaluating text-to-image alignment, as well as a method that builds upon this insight to improve the process of text-to-image generation. For the first contribution of evaluating text-to-image alignment, the paper employs the strategy of using a Large Language Model (LLM) to decompose the textual prompt into multiple components, and then using a Visual Question Answering (VQA) model to answer the individual components/assertions to obtain a final score for the alignment between the generated image and the textual prompt. To improve the image generation process, the paper builds upon the existing approach Attend-and-Excite, that uses the cross-attention maps to focus on different tokens in the textual prompt. Here, the weights applied to the sub-prompts, as well as the strength of the cross-attention are used as parameters to be optimized in an iterative refinement procedure (while optimizing the decompositional metric) to obtain more faithful generated images.

Strengths

1) The paper tackles an important problem (of both evaluating text-image alignment, and improving text-to-image generation), and proposed interesting ideas for the same. 2) The Prompt Decompositional Model intuitively makes a lot of sense, and quantitative results indicate that it seems to be significantly better than prior methods (i.e CLIPScore) 3) Using the improved evaluation metric to improve the generation is a good idea. While in terms of novelty, it is almost entirely based off Attend-and-Excite and prior work that uses prompt weighting, optimizing it during test-time is new, and seems to bring good improvements. 4) The experiments seems to be reasonably thorough. There are user studies that validate the key claims (of improved text-image alignment evaluation, and improved generation), as well as a new dataset of prompts.

Weaknesses

One major concern I have with respect to the evaluation metric proposed in the paper is its similarity to [a]. While [a] is only an Arxiv preprint, and can definitely be considered as contemporary work, I would think that it is a good idea to atleast acknowledge it, since the idea behind breaking down a prompt into multiple Question-Answer pairs and then using a VQA model to evaluate it has been first introduced there. In terms of the quantitative results, my worries are that the paper mostly relies on user studies/human evaluations to compare the proposed method against existing methods. While this is understandable (and quantitative results on the DAScore metric would be unfair, since the model is explicitly optimizing it), it might make sense to have quantitative results on other metrics (such as T2TScore) to demonstrate that the improvements are visible on automatic quantitative metrics. This would also allow benchmarking on other datasets beyond the Decomposable Prompts. [a] Hu et al. TIFA: Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering

Questions

The only question I have is that all the experiments in this work are done on the Decomposable Prompts dataset. While it's large enough (with 4k prompts), I wonder if it is sufficient in terms of variety, of if it makes sense to also benchmark on other datasets such as DrawBench.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Yes, the limitations section in the supplementary material provides a fairly comprehensive overview of the possible limitations of the proposed method in the paper.

Reviewer sTZE5/10 · confidence 4/52023-07-06

Summary

The paper proposes a method called Decompositional-Alignment-Score (DA-Score) to assess and enhance the alignment between text and generated images during training. The authors accomplish this by breaking down each prompt or text into a series of assertions using a large language model (LLM). They then calculate correlation scores between each assertion and the generated image, employing a pre-trained VQA model. The final score, obtained by combining all the assertion scores, serves as an evaluation metric for text-image alignment. It further guides the generative models to produce more aligned images iteratively.

Strengths

This paper seeks to address a significant challenge in the text-to-image generation task, i.e., text-image alignment. The proposed framework adopts LLM and pre-trained multi-modal models, which is reasonable.

Weaknesses

1. There is a concern regarding the performance of the VQA model, as it plays a crucial role in determining the accuracy of the proposed score. The effectiveness and reliability of the evaluation process heavily rely on the capabilities and performance of the VQA model. 2. The experiments are not convincing enough and there and there is a lack of some specific details regarding the design of the dataset. Please refer to more details in the following.

Questions

*Method 1. The performance of the proposed alignment score heavily relies on the accuracy of the VQA's answering results (also mentioned in the limitations by the authors). It raises the question of whether the used VQA model is sufficiently robust to handle all the given questions. If available, it would be helpful to provide quantitative results or an assessment of the accuracy of the VQA model in addressing the given questions on the dataset in this paper. 2. It seems there is a lack of definition for the meaning of the variable $l$ in Eq.(5). *Experiment and Dataset 1. How about the quality of the generated images? Based on the visual results presented in the paper (e.g., Figure 5), it appears that focusing solely on text-image alignment may lead to a reduction in image quality. It would be beneficial to provide additional results that consider image quality evaluation metrics such as FID and FID-CLIP. 2. It seems that the generated images of the baseline generative models (e.g., stable diffusion) are not satisfied enough. I think it would be more realistic if use a more recent version like stable diffusion 2.1. Besides, it raises the question of whether the proposed alignment method remains effective or necessary when applied to more powerful text-to-image generation models, such as Imagen [1] or Parti [2]. 3. The paper does not provide clarity on the number of annotators involved in rating the samples within the proposed dataset. If only a small number of annotators, the obtained human annotations may lack sufficient credibility and be susceptible to subjective biases. [1] Saharia, Chitwan, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35 (2022): 36479-36494. [2] Yu, Jiahui, et al. "Scaling autoregressive models for content-rich text-to-image generation." arXiv preprint arXiv:2206.10789 (2022).

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors discuss three limitations of the current version and provide the corresponding proposal for a solution.

Reviewer GK7B7/10 · confidence 4/52023-07-07

Summary

The paper proposes a new approach to evaluate text-image alignment of generated images and how to use that feedback to improve the image quality. For this, a LLM is first used to parse the text description into individual assertions and the alignment of those individual assertions with the generated image is then measured via a VQA model. The resulting metric is shown to better correlate with human judgement than other metrics such as CLIP/BLIP score. Furthermore, the feedback of the VQA model can be used to increase the impact of specific parts of the promp during the denoising process which improves the text-image alignment of the generated image without requiring additional model training.

Strengths

The paper is well written and presented. The novel evaluation process is intuitive and automated, and shows better alignment with human judgement than other metrics. The use of the metric to improve tet-image alignment at test time is also shown and leads to better results than other approaches. The evaluation includes several human user studies highlighting the merits of the new approach.

Weaknesses

The approach introduces the need for two additional networks (LLM and VQA model) into the pipeline which will incur additional cost and compute requirements. The overall quality and the resulting scores are likely also heavily dependent on the quality of the LLM and VQA model.

Questions

While the approach works I wonder how much slower it is (as an evaluation metric) compared to other approaches such as CLIP. CLIP only requires a forward pass through the text and image encoder, whereas the new approach requires interactions with an LLM (to get the assertions) and then also needs to query a VQA model for each assertion.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

Overall this is a good paper and the evaluation metric reminds me of an updated and improved version of the SOA score (Semantic object accuracy for generative text-to-image synthesis, TPAMI, 2020).

Reviewer bVCd2023-08-18

I thank the authors for their detailed response, addressing all the remaining concerns. I have raised my score to 7 on account of this.

Area Chair dMhB2023-08-21

Thank you for this comment and elaborating the required results in the rebuttal as well. We will take it into consideration while making the final decision regarding the paper.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC