Intelligent Grimm -- Open-ended Visual Storytelling via Latent Diffusion Models

Generative models have recently exhibited exceptional capabilities in text-to-image generation, but still struggle to generate image sequences coherently. In this work, we focus on a novel, yet challenging task of generating a coherent image sequence based on a given storyline, denoted as open-ended visual storytelling. We make the following three contributions: (i) to fulfill the task of visual storytelling, we propose a learning-based auto-regressive image generation model, termed as StoryGen, with a novel vision-language context module, that enables to generate the current frame by conditioning on the corresponding text prompt and preceding image-caption pairs; (ii) to address the data shortage of visual storytelling, we collect paired image-text sequences by sourcing from online videos and open-source E-books, establishing processing pipeline for constructing a large-scale dataset with diverse characters, storylines, and artistic styles, named StorySalon; (iii) Quantitative experiments and human evaluations have validated the superiority of our StoryGen, where we show StoryGen can generalize to unseen characters without any optimization, and generate image sequences with coherent content and consistent character. Code, dataset, and models are available at https://haoningwu3639.github.io/StoryGen_Webpage/

Paper

References (72)

Scroll for more · 38 remaining

Similar papers

Reviewer zaiF4/10 · confidence 4/52023-06-25

Summary

The authors aim to generate a series of coherent images given a series of text prompts resembling a visual storybook. To do so, the authors focus on two fronts: (1) leveraging the Stable Diffusion model to generate the series of images and (2) generating a diverse dataset used to train the model on a range of styles. For generating a set of coherent images, the authors condition the Stable Diffusion on both the text prompt and a set of previously generated frames, both encoded using a frozen CLIP encoder. The text conditioning is passed to the U-Net layers via the standard cross-attention module and LoRA. To insert the image conditioning, the authors introduce a Visual Context Model resembling the standard text conditioning module. The diffusion model is then partially trained to generate images that are both consistent with the text prompt and previously generated frames. To attain images ranging in style, the authors construct a new dataset, named StorySalon, consisting of Youtube videos and E-Books. These raw storybooks are filtered and re-captioned to better align with the visual content of the storybook. Qualitative results demonstrate the ability to generate new storybooks on prompts generated by ChatGPT while quantitative results demonstrate improvements over simple baselines.

Strengths

- The authors focus on an important task of generating a series of coherent images that follow a given text prompt. This could potentially be useful beyond a simple storybook creation, e.g., for video generation. - The Visual Context Module as a means for injecting image-level details to the denoising network is simple and intuitive and could be useful in other tasks. For example, in image editing where the desired edit cannot be easily described using language. - The visual results generated by StoryGen are impressive in comparison to the evaluated baselines.

Weaknesses

**Dataset Creation:** - I have some reservations regarding the use of the term Human Feedback in Section 3.2.3 and Section 4.2. While ChatGPT was fine-tuned to align with human preference, I believe that using ChatGPT for generating additional prompts should not be considered Human Feedback. While this is an intriguing approach, I believe replacing Human with LLM is more reflective of what is actually done here. - Regarding the ablation study performed by the authors, it seems that “Human Feedback” leads to a quite negligible decrease in FID and I am therefore uncertain if this really contributes to the curriculum learning scheme. The authors mention that more stories can be added, but I would have expected to see a bigger improvement if this stage is truly important. **Evaluation:** - There are numerous essential evaluations that are missing from the current submission. Among these, the most important is a thorough evaluation and comparison of StoryGAN, Story-DALL-E, and AR-LDM. All three have publicly available code so an evaluation is needed to understand the improvement realized by StoryGen. - I am not sure that FID is a particularly interesting metric here since all evaluated methods in Table 1 use Stable Diffusion to generate the images. Moreover, I do not believe that FID is a good metric when trying to measure how much the image captures a given style, as is the goal here. Maybe a CLIP-based metric using a prompt depicting a style would be more appropriate here? - There are numerous ablation studies that I believe are required to understand the contribution of both the proposed architecture and dataset. - Architecture: - An ablation study on the Visual Context Module and whether a simpler conditioning is possible (see my detailed question below). - Was an ablation study performed on the BERT-like masking during the multi-frame fine-tuning? - Dataset - The impact of the visual-language alignment stage in preparing the dataset. The authors state that directly fine-tuning on the story narrative may be detrimental, but do not validate this claim. - Some additional evaluations could help validate the effectiveness of the method. - First, the authors claim that the method can be used to generate stories of arbitrary lengths (Line 106). It would be great to quantify this by generating stories of varying lengths and validating whether there is a loss in quality after a certain length. - One particularly interesting component of the method is the Visual Context Module, so I would have liked to see far more evaluations performed on it. For example, the authors mention that multiple frames can be used for conditioning by concatenating their CLIP feature. Some interesting questions that could help strengthen the importance of the component include: - How much was this evaluated? - How much does conditioning on more frames assist in temporal consistency? - How many previous frames can be concatenated without hindering performance? - Stating that the model achieves a significant improvement in the alternative models seems like a strong over-claiming when approximately 30 participants were used for the user study. A substantially larger pool of participants would be needed to truly quantify this improvement, especially since this is the only relevant metric used to evaluate the methods. Why are the reported FID metrics between Table 1 and Table 2 different? Is a different dataset used? I would expect only “without HF” to be different if both use the same dataset.

Questions

I hope the authors can help clarify several questions that I have regarding both the method and the evaluations. **General Comments:** - A reference to Rombach et al. is missing when discussing Stable Diffusion (e.g., in Line 38 where it is first used). - First stated contribution is that the authors propose the task of open-ended visual storytelling, but this was a previously studied problem (as also mentioned in Line 44). - When describing the style transfer module and visual context module, the authors reuse the same notation for the weight matrices \(W^K\) and \(W^V\). To make the writing a bit clearer, the authors should use a different notation for each case since these are separate modules if I understood correctly. **Method:** - What happens if the image conditioning is simply done using the standard image-to-image technique where the noised latent is conditioned on the previous image? Adding an ablation study would help validate this design choice. If performing a thorough ablation study is difficult, providing intuition on the motivation for the Visual Context Module would help highlight the contribution to readers. - Where is the style transfer model illustrated in Figure 2? Is this the blocked labeled LoRA? The authors mention that this is a LoRA-like architecture. Could the authors kindly clarify what exactly the difference between the original LoRA design is and why this modification was made? - Regarding the Multi-Frame Conditioning: - According to Figure 2, it seems like StoryGen receives several previous frames as conditioning while Line 136 indicates that only the previous frame is used for conditioning. Could the authors kindly clarify this? - Similarly, in the general setting, it seems that multiple frames are used for conditioning. Therefore, would it be more accurate to revise the equation in Line 150 to indicate any number of \ell frames as conditioning? - In Section 5.1, the authors mention that the multiple-frame fine-tuning stage is done using a single conditioning image (Lines 256-257). Where do the multiple frames come into play? Is this really a multi-frame conditioning? **Small comments:** - Line 70: ou -> our - Line 142: generate -> generates, align -> aligns - Line 276: Propmt -> Prompt

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good 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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

The authors include a discussion on current limitations and potential societal impacts in the supplementary.

Reviewer WrUp6/10 · confidence 5/52023-07-02

Summary

This work proposes the model StoryGen for the task of visual storytelling. Visual storytelling is a task to generate a sequence of consistent images given a story (several sentences). StoryGen is a diffusion model taking in both image and text as conditions, and outputs an image consistent with the conditions. The training process includes pre-training for single image, finetuning for multiple image and finetuning with human feedback. On top of the StoryGen model, this work also provides a dataset, called StorySalon, which consists of 2k story books (30k well aligned text-image pairs). The overall structure of StoryGen model is simple. It is built upon existing well trained diffusion models and image/text encoders. To generate cartoon-like image, LoRA is adopted into the text conditioning module in a diffusion model. The author calls it the style-transfoer module. The parameters in LoRA are updated at this pre-training stage to give single cartoon image. Next is the multipe image fune-tuning. StoryGen conditions on both text and image, which is implemented by using two cross attention layers: One is noise input + text and the other is noise input + encoded previous generated image. After the second step, StoryGen if further finetuned on 100 high-quality stories. The author also spends some efforts to collect the StorySalon dataset. To begin with, the author downloads a huge number of videos and subtitles from online web resources with potential stories. Then give story-level description and visual level description for each story. The story level description is obtained by using dynamic time warping algorithm using subtitles. The visual level description is derived from ChatCaptioner. Finally, OCR method is applied to get potential videos captions. The experiment section shows that StoryGen model can give consistent and story-like output images, while other methods fail.

Strengths

The paper is clearly written. The collected StorySalon dataset could benefit the research community.

Weaknesses

There is not much technical novelty, and the experimental results are limited.

Questions

- Line 192 mentions the 100 high quality books are added into the training set. For the third stage training, is it still using the datasets from previous steps, but with 100 more samples? Would it make more sense to only use the 100 high quality books for further fine-tuning? Also, the name (fine-tuning with human feedback) is confusing since there is neither human feedback nor reward model. Is there any human involved to give preferences of different generated outputs from the StoryGen model? - The human feedback fine-tuning stage doesn’t seem to help a lot in the quantitative scores. It would be better to include visual examples without human feedback fine-tuning. If the difference is too small, it would be better to re-structure the paper. - For the visual examples given in Figure 4, I noticed the boy with yellow hair shows in both stories. Does this boy appear a lot in the training set? Also, it would be better if the author could provide more than one examples for each story using StoryGen. Otherwise, it feels like the model is overfitting to the training data.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

4 excellent

Contribution

2 fair

Limitations

- Since there is no human labeler involved in collecting StorySalon dataset and the total number of stories in StorySalon is only 2k, I’m concerned about its quality. - It is also unclear from the examples provided if the results is derived by just overfitting the training set.

Reviewer tcbg3/10 · confidence 4/52023-07-03

Summary

The work focuses on the application of image generation based on a given story. Specifically, the proposed model is conditioned on the current sentence and prior generated images to ensure the story is engaging and coherent. A progressive training strategy is proposed to achieve a good model. To improve the proposed method, a new dataset is collected, while a set of human-verified generative samples are also utilized to improve the generated images. -------------------------- I acknowledge the author's effort in the rebuttal and have made changes to the review accordingly.

Strengths

+ This work demonstrates the possibility of generating visual storytelling images conditioned on the given stories. + A three-stage curriculum training strategy is proposed to train the proposed model. However, it would be great to demonstrate the limitation of training the model with multiple-frame (i.e., without single-frame pre-training) + The authors collected a large-scale dataset to enable model training for storytelling purposes.

Weaknesses

- the technical contribution of this work is limited. Most of the components are not novel and the key contributions are the way it is combined to generate a plausible output. It is unclear what the insights generated from this work that is not previously obvious to the community. - The description of the new StorySalon dataset is limited. Specifically, it is unclear if the collected dataset has obtained legal consensus and properly handled copyright issues. - The work lacks a comparison with existing work, such as those introduced in line 44 and line 93-102. The two baselines in Table 1 are too naive as both are inherently limited to generate a fair comparison with the proposed method. Minor: - Fig 2 should clearly state that the Image encoder only considers a single previous frame to generate the next frame.

Questions

- Please justify what it means by "open-ended". Was the dataset divided in a manner that the training and validation/test dataset has distinct distribution or content? - The final stage of curriculum training strategy fine-tune the model with human feedback. Here, the stories are generated with ChatGPT and the images are synthesized with the same model. This creates a problem of bootstrapping machine learning and makes it unclear whether such an approach can benefit the model. A recent work [1] discussed the curse of recursion with generated data. I want the invite the author to discuss the problem of this. In addition, the results in Table 2 show that with or without HF has a small difference. Can the author also explain what does 0.19 translate to the differences in the qualitative results? [1] Shumailov et al. THE CURSE OF RECURSION: TRAINING ON GENERATED DATA MAKES MODELS FORGET. https://arxiv.org/pdf/2305.17493.pdf - In line 141, why is a small number of newly-added trainable parameters required and how does it impact the model performance/learnability? - In Table 1 and Two, why is the FID score in both table inconsistent?

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed 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

2 fair

Limitations

The paper (in supplementary) discusses that data bias is an issue that needs to address in this domain. Collecting a larger dataset for training is the solution discussed. This may be valid considering this work is still in the early stage of the research. I want to point out that the discussed approach is limited as (1) it is resource-consuming, and (2) it will face the problem of copyright in order to obtain a good dataset for training. The data ownership issue may be a major hindrance.

Reviewer SemY3/10 · confidence 4/52023-07-05

Summary

This paper presents StoryGen, an auto-regressive image generator that leverages text and image conditioning. StoryGen incorporates a style transfer module integrated into the text-conditioning module, along with a visual context module. The authors also constructed a substantial dataset called StorySalon, comprising 2K storybooks and 30K text-image pairs.

Strengths

1. This paper constructs a new dataset StorySalon contains 2K storybooks and more than 30K well-aligned text-image pairs. The authors have invested significant effort into filtering the data, making it a valuable resource for advancing the field of story visualization. 2. The paper is well-written and easy to follow.

Weaknesses

1. The illustration does not align with the description provided. In line 135, the authors state that "StoryGen generates the current frame $\mathcal{I}_k$ by conditioning on both the current text description $\mathcal{T}_k$ and the previous frame $\mathcal{I}_{k-1}$, as illustrated in Figure 2." However, the left figure of Figure 2 shows the image conditioned on more than one previous image, which contradicts the mentioned conditioning approach. 2. The improvement of Human feedback appears to be trivial, as indicated in Table 2. The 0.19 FID score gap could potentially be attributed to different training seeds, which raises doubts about the significance of the reported improvement. (I do not agree with the statement that 200 stories are too small since the model is trained using 2k stories overall. It appears to be sufficient for human alignment and does not require an extensive amount of data.) 3. The FID score lacks precision in the test set, particularly with only 100 storylines. It is recommended that the authors expand the test set by including more stories to provide a more accurate evaluation. 4. The baselines SDM and Prompt-SDM are too weak. It is suggested that the authors compare StoryGen with finetuned or LoRA-finetuned SDM models using the same training settings to establish a more robust baseline for comparison. 5. The auto-regressive generation approach employed by StoryGen has already been proposed by AR-LDM. Consequently, the architecture design itself lacks novelty. 6. StoryGen is only conditioned on one previous image and does not utilize the corresponding caption of the previous image. In the depicted cases of Figure 1 and Figure 4, there is only one main recurring character. If multiple characters were present, StoryGen may struggle to ground the characters in the previous images. Furthermore, if a character does not exist in the previous image, StoryGen may face difficulties in maintaining consistency between frames. 7. The language understanding capacity of StoryGen appears to be weak. For instance, in the second case of Figure 4, the rabbit appears small in the fourth frame, whereas it should be as big as it is in the fifth frame. Additionally, in the sixth frame, the boy's hair does not become brighter as described in the caption. Moreover, in the seventh frame, multiple other boys are depicted with the same yellow hair, which contradicts the previous story setting. This limitation may stem from StoryGen solely relying on the only one previous frame $\mathcal{I}_{k-1}$ and not incorporating previous captions into its generation process.

Questions

1. StoryGen generates the current frame $\mathcal{I}_k$ by conditioning on current text description $\mathcal{T}_k$ and the previous frame $\mathcal{I}_{k-1}$. I am wondering why the authors do not use all historical frames $\mathcal{I}_0, \cdots, \mathcal{I}_{k-1}$ in this process. 2. StoryGen introduces a style transfer module $\phi_{text}$ and a single-frame pre-training stage. However, since the training set is constructed from multiple sources (as mentioned by the authors in Line 244), the style is inconsistent. Moreover, many LoRA style transfer plugins not only tune the cross-attn module but also the self-attn modules. Why not follow this setting and add LoRA layers in both self- and cross-attn layers? This way, StoryGen can also benefit from other style transfer LoRA developed by the community. 3. In Line 186, the authors mention randomly dropping some words in the text with a certain probability following BERT. I have doubts about this technique because there is no masked language modeling task in StoryGen, and such a technique may not be helpful. 4. I question whether the authors truly need to replace descriptive captions with story narrative text. The story in the storybook and the story generated by LLMs are both in the form of story narrative text, rather than descriptive captions. Using story narrative text to train the model may lead to implicit alignment with human performance.

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed 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

2 fair

Presentation

3 good

Contribution

3 good

Limitations

1. The author should provide examples of the constructed dataset showcasing different visual styles and character appearances. 2. The StorySalon dataset consists of 2K storybooks and over 30K well-aligned text-image pairs, which is smaller compared to datasets such as FlintstonesSV (24K stories and 123K image-caption pairs), PororoSV (14K stories and 74K image-caption pairs), and VIST (27K stories and 136K image-caption pairs). Despite the authors' claim that StoryGen can perform open-ended story generation, it remains unclear whether StoryGen can generate stories involving more complex scenarios with unusual entities. 3. Apart from human feedback, the authors have not conducted any other ablation studies to evaluate the effectiveness of their proposed techniques, such as word dropout and curriculum learning. 4. There are concerns regarding the legality of using web-crawled e-books. The authors should provide additional information about the sources of the e-books and clarify whether proper copyright guidelines were followed.

Reviewer 4bmh5/10 · confidence 4/52023-07-07

Summary

The paper propose an approach for fine-tuning diffusion models for the task of story generation, where a model must generate frames for sentences in a story. To do so, they propose adding adaptors conditioned on both images and text into a pre-trained stable diffusion UNet. The authors also introduce a scraped dataset of 2k stories with 30k image-text pairs, which serves as the data foundation for their fine-tuning.

Strengths

S1. The dataset of story text and images is a significant contribution, which can be useful for future work on visual storytelling. S2. The paper is generally well framed and motivated. Writing and presentation is in general strong and polished. S3. Simplicity of the method. The method uses off-the-shelf components and algorithms (LoRA, cross-attention, etc.) to enable new capabilities. I see this simplicity as a strength not a weakness. S4. Inclusion of human evaluation is a strength.

Weaknesses

W1. Presentation. Figure 2 suggests that StoryGen is conditioned on all past frames. However, in reality StoryGen is conditioned on the most recent frame. W2. Results. It appears that the model is not able to preserve style and content as well as say DreamBooth. For example the stripes on the shirt in Figure 4 are not preserved. W3. Evaluation. It seems that the StoryGen model without human feedback model was not evaluated in the human evaluation (Table 1). Given that the ablation in Table 2 shows similar FID scores for StoryGen with and without human feedback, it is not clear if this step is really necessary. W4. Lack of baselines. StoryGen is specifically fine-tuned for the desired task, while stable diffusion is not. Given this, it is perhaps not surprising hat StoryGen greatly outperforms stable diffusion. Can some other baselines, perhaps based on the DreamBooth (with the subject in the first image representing the special [V] token) be used for a stronger baseline? This is just one idea; however, comparing against some prior work may help to elucidate the strength of the method. W5. Evaluation. Fine-tuning can sometimes hurt the generality of a model. How well are individual frames generated relative to the base SD model? Is it possible to do a human evaluation here or compute CLIP scores? Ideally, the story would be cohesive (frame-to-frame consistency) without degradation in quality of each frame.

Questions

Q1. Can the authors provide further discussion of why StoryGen without human feedback was not evaluated in the human evaluation (W3)? Is it possible to add this evaluation? I believe this would help contextualize the importance of the human feedback in the proposed method. Q2. Can the other concerns related to evaluation be addressed (W2, W4, W5)?

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 paper does not directly address limitations in the main paper, which is an additional weakness. I suggest discussing failure cases or otherwise conducting a failure analysis.

Reviewer 4bmh2023-08-14

Thanks for the responses to my comments! However, I do still feel that an additional baseline that has some sort of task-specific training is necessary to contextualize performance. DreamBooth was one such idea, but other ideas could include LoRA fine-tuning a StableDiffusion model on the proposed StorySalon dataset. I am electing to keep my initial score of 5.

Authorsrebuttal2023-08-15

- Thanks for your comments. But the results of additional baselines that you required have been provided in our supplementary and rebuttal. Please check **Q6** of the **Author Rebuttal** again. - The results of StableDiffusion with LoRA fine-tuning on StorySalon have been provided in **Table 1** of the **supplementary**, termed as **StoryGen-Single**. Besides, we have also included the StableDiffusion model with cross-attn layers finetuned on our proposed StorySalon dataset as another strong baseline, namely **Fine-tuned SDM**, in **Table 2** of the **rebuttal PDF appendix**. We compared our model with these stronger baselines on FID and CLIP-based scores, and the quantitative results have demonstrated that our full StoryGen can also significantly outperform these stronger baselines.

Reviewer tcbg2023-08-16

Thank you for the details respond to the review and additional results in the provided PDF. Can the author provide additional comments about the video data of the proposed dataset. While "we will release in the form of YouTube URLs" is a plausible practice, it has been evidenced that video could be removed from the platform and unaccessible by other researcher. This could potentially affect the reproducibility and availability to other researchers. About the bootstrapping problem, I hope the authors can provide more discussion based on the findings in [1]. Do also analyse the generated feedback and training size. The amount of data in ChatGPT is on different scale when compared to the dataset presented in this submission. It is hard to determine if the bootstrapping effect is positive or negative for this work.

Authorsrebuttal2023-08-18

- Thanks for your comments. - Releasing datasets in the form of URLs is a common practise, for example, Youtube-8M [2], WebVid-10M [3] and VideoCC [4]. However, we agree with the reviewer that the videos can be removed from the website, thus we are currently expanding our StorySalon dataset, and will replace all YouTube video data with open-source ebook data registered under **CC BY 4.0 license** in order to completely eliminate copyright concerns. - Thanks for pointing out the paper, we have read through it in detail. As mentioned in the response, in our case, we do observe positive gains by doing human feedback, as shown in **Table 1 (supplementary)**. This might be due to the fact that paper [1] has only investigated MNIST, which is a significantly simpler dataset than the ones we are using. - In addition, we do observe that these gains are correlated with the volume of feedback data, our current human feedback data for fine-tuning only constitutes a mere **10%** of the StorySalon dataset, which is also significantly smaller in comparison to the LAION-5B employed in SDM pre-training. Thus we are actively developing a user platform to gather more user feedback, aiming to explore the positive improvement upper bound of human feedback on our StoryGen model and analyze whether a negative effect would emerge as human feedback expands to a certain scale. - [1] Shumailov et al. THE CURSE OF RECURSION: TRAINING ON GENERATED DATA MAKES MODELS FORGET - [2] Sami Abu-El-Haija et al. YOUTUBE-8M: A LARGE-SCALE VIDEO CLASSIFICATION BENCHMARK - [3] Max Bain et al. FROZEN IN TIME: A JOINT VIDEO AND IMAGE ENCODER FOR END-TO-END RETRIEVAL - [4] Arsha Nagrani et al. LEARNING AUDIO VIDEO MODALITIES FROM IMAGE CAPTIONS

Reviewer SemY2023-08-19

Thanks for the author response

Thanks a lot for the response. I would like to keep my score as it is.

Reviewer WrUp2023-08-21

Thanks for the author's response. This is an interesting work. I'll keep my score as weak accept.

© 2026 NYSGPT2525 LLC