Pith. sign in

REVIEW 4 major objections 6 minor 44 references

RecipeGen: A Step-Aligned Multimodal Benchmark for Real-World Recipe Generation

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read RecipeGen introduces the first large-scale, step-aligned multimodal benchmark for recipe generation, pairing 26,453 recipes with 196,724 step images and 4,491 videos, and adds metrics for ingredient fidelity, interaction faithfulness, and…

desk verdict Useful dataset contribution, but the video subset's step alignment is unvalidated and several fixable inconsistencies need attention; worth peer review with heavy revision. read the letter →

arxiv 2506.06733 v3 pith:C4OVUXBT submitted 2025-06-07 cs.CV

Ruoxuan Zhang , Jidong Gao , Bin Wen , Hongxia Xie , Chenming Zhang , Hong-Han Shuai , Wen-Huang Cheng This is my paper
classification cs.CV
keywords recipegenerationmultimodalbenchmarkstep-aligneddatatext-to-imageimage-to-videotext-to-videofoodcomputingingredientfidelity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

RecipeGen proposes to fill a gap in food computing: existing recipe datasets tie a final dish photo to a recipe, but they do not provide an image, and for some recipes a video, for each individual step. The paper presents a benchmark of 26,453 real-world recipes, 196,724 step-aligned images, and 4,491 step-aligned videos, assembled from web platforms and existing video datasets and cleaned automatically with large-model tools. It also introduces three evaluation metrics aimed at procedural correctness: Ingredient Accuracy, Interaction Faithfulness, and Cross-Step Consistency. A sympathetic reader would care because this gives text-to-image, image-to-video, and text-to-video generation a shared, step-level testbed for a task with clear real-world use in cooking education and recipe assistants.

What carries the argument

The machine that carries the argument is the step-alignment pipeline: for each step, the frame with the highest CLIP similarity to the step text becomes the keyframe; videos dominated by people are discarded or have person frames removed and reassembled; and an LLM cleans and merges step text to remove noise. On the evaluation side, the load-bearing objects are three metrics: Ingredient Accuracy, which uses a fine-tuned vision-language model to check for expected ingredients; Interaction Faithfulness, which uses an LLM-generated caption of visible ingredient interactions scored by CLIP; and Cross-Step Consistency, which compares DINOv2 features and step counts across consecutive steps.

What would settle it

Sample RecipeGen's step-image pairs at random and have annotators, blind to the captions, judge whether the CLIP-selected keyframe shows the action in that step's text; the step-alignment claim is falsified if agreement approaches chance or if the reassembled video clips show systematic motion jumps at the splice points where person frames were removed.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's terms, is that no existing dataset provides fine-grained, step-aligned visual supervision for recipe generation, and RecipeGen fills that gap at scale: 26,453 recipes, 196,724 images, and 4,491 videos, with each step's text matched to a keyframe and, in the video subset, to a segment. The paper reports that its quality-control pipeline achieves higher Goal Faithfulness and Step Faithfulness than VGSI-Recipe and YouCook2, and its three new metrics are introduced as capturing ingredient fidelity, interaction fidelity, and cross-step coherence. The experiments support two findings: current text-to-image models either lose inter-step memory (independent generation) or misassign text to spatial regions (joint generation), and none of them reliably depicts ingredient interactions; for video, text-only generation outperforms keyframe-conditioned generation because the keyframe does not fully represent the action in the step text.

Load-bearing premise

The benchmark's value rests on the assumption that the fully automated pipeline--CLIP-based keyframe selection, person-frame removal and clip reassembly, and LLM step cleanup--produces step-image-video triples that genuinely show the cooking action each step describes, rather than frames that merely resemble the caption.

Editorial extensions

If this is right

  • Text-to-image, image-to-video, and text-to-video models can now be trained and evaluated at the granularity of individual recipe steps, not just the finished dish.
  • The three proposed metrics give the field a common protocol for comparing models on procedural correctness: ingredient presence, interaction type, and cross-step visual consistency.
  • The paper's findings indicate that step-by-step generation models need an inter-step memory mechanism, while joint-generation models need better binding of each step description to its spatial region.
  • The observed I2V lag behind T2V suggests that keyframe conditioning should be action-aware rather than purely CLIP-similar, or video generation should be trained to infer the action from the keyframe.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because keyframe selection is driven by CLIP similarity, the benchmark probably inherits CLIP's blind spots: steps whose cooking action is visually nondistinct (e.g., 'let rest') may receive near-arbitrary keyframes, so the claimed value of step alignment should be evaluated separately for visually distinctive and nondistinct steps.
  • The five-way interaction taxonomy (mix, blend, put on, cover, none) could transfer to other instructional domains such as assembly or cosmetics as a generic check that generated visuals match action-relationship semantics.
  • The Ingredient Accuracy VQA model could be repurposed as a reward signal for training recipe-generation models rather than only an evaluation tool.
  • The video pipeline's excision of person frames creates splice points whose temporal continuity is not measured by any current metric; a motion-discontinuity score at known splice boundaries would test whether the curated videos retain natural procedural flow.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces RecipeGen, a benchmark for recipe-based text-to-image, image-to-video, and text-to-video generation, claimed to contain 26,453 recipes, 196,724 images, and 4,491 videos with step-level alignment between text and visuals. The dataset is assembled from web-crawled sources (Douguo, YouTube) and the YouCook2 dataset, with a GPT-4o-based cleaning pipeline, a sampled human review, and three proposed evaluation metrics: Cross-Step Consistency, Ingredient Accuracy, and Interaction Faithfulness. The paper benchmarks several T2I, I2V, and T2V models and reports scores on these metrics. The central claim is that RecipeGen is the first large-scale, real-world, step-aligned multimodal benchmark for recipe generation.

Significance. If the step-alignment claim holds, RecipeGen fills a clear gap in food computing: existing recipe datasets mostly provide final dish images or unaligned video captions, while RecipeGen offers per-step image and video supervision. The scale (tens of thousands of recipes, hundreds of thousands of images) is genuinely useful for training and evaluating generative models. The paper's curation pipeline is detailed and combines automatic filtering with human checks, and the benchmark includes a broad set of T2I, I2V, and T2V baselines, which is a useful reference point for the community. The proposed metrics address real limitations of generic CLIP-based evaluation. However, the validity of the video-subset step alignment is not established, the dataset counts are internally inconsistent, and two of the three proposed metrics have potential circularity or unvalidated components. These issues are load-bearing for the paper's central claims.

major comments (4)
  1. [§3.2 (Quality Control for YouCook2)] After deleting person-containing frames and reassembling clips, the original YouCook2 step timestamp intervals no longer correspond to segments in the reassembled video. The paper does not describe re-estimating step boundaries or tracking which original frames fall into each reassembled segment. Consequently, the CLIP keyframe selection 'within the corresponding video segment' can pick frames that originally belonged to an adjacent step or miss the step's action entirely. This undermines the step-aligned claim for the 924 YouCook2-based video recipes and the keyframe conditioning used in the I2V evaluation in Table 4; note that the human review reported in §3.1 covers only 2,000 sampled image-text recipes, not the reassembled video subset. Please either re-annotate step boundaries after reassembly or validate that each selected keyframe originates from the original step interval, and report the resulting alignment accuracy.
  2. [§1 and §3 (Dataset statistics)] The stated total of 26,453 recipes does not match the component counts given in the paper. Section 3.1 reports 21,944 image-text recipes after quality control; Section 3.2 reports 924 YouCook2 videos and 3,567 web-crawled videos, totaling 4,491 videos. If the video recipes are disjoint from the image-text recipes, the total is 21,944 + 4,491 = 26,435, not 26,453. The image count is also unexplained: the image-text subset has 139,872 images, which leaves 56,852 images unaccounted for in the 196,724 total. Please clarify whether video recipes overlap with image-text recipes and provide a per-subset breakdown of recipes, images, and videos.
  3. [§3.3 (Interaction Faithfulness)] The Interaction Faithfulness metric as described is circular. It uses GPT-4o to generate a one-sentence caption from the generated image itself, then computes the CLIP similarity between the generated image and that caption. This measures self-consistency of the image–caption pair, not whether the depicted interaction matches the step's ground-truth interaction type (Mix up, Blend, Put on, Cover, No relationship). To validate this metric, the authors should compare the interaction labels predicted by the metric against human judgments on the five interaction categories, or ground the caption in the step text rather than in the image being evaluated.
  4. [§3.3 (Ingredient Accuracy)] The Ingredient Accuracy metric depends on a fine-tuned Qwen2.5-VL ingredient-detection model, but the paper provides no details on the training data, train/test split, or detection accuracy of this model. A three-epoch fine-tune with unspecified data could have systematic errors that bias the IA scores in Tables 3 and 4. Please report the detector's precision and recall on a held-out set, and ideally show that IA correlates with human judgments of ingredient presence, before the metric is used to compare generative models.
minor comments (6)
  1. [§2 (Related Work)] The text refers to 'FoodEarth [42]', but reference [42] is titled 'FoodSky: A Food-oriented Large Language Model...'; please correct the name and citation.
  2. [§4.1 (Text to Image Generation)] Table 3 includes SD3.5 in the results, but the model list in §4.1 does not mention Stable Diffusion 3.5; please add it to the method description or remove it from the table.
  3. [§4.2 (Recipe Video Generation)] The statement 'We randomly select 384 videos for test' does not specify the source subset (YouCook2-based, web-crawled, or both) or whether the test split is publicly released; please clarify for reproducibility.
  4. [§3.1 (Metrics Computation and Human Check)] The human review is described as covering 2,000 sampled recipes, but Figure 1's overview shows a 'manual quality control pipeline' without noting that the manual check is sampled rather than exhaustive; please make this explicit in the figure caption and main text.
  5. [§3.3 (Metrics)] The phrase 'we design metrics along two primary dimensions, , as summarized' contains a stray comma and missing text; please fix the typo.
  6. [References] Reference [17] cites 'Flux.1 AI' with a URL to flux1ai.com; if an official model card or technical report exists, please cite it instead of the possibly unofficial URL.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark construction and model evaluations are self-contained, and the CLIP keyframe-selection loop is a curation step rather than a fitted result used as validation.

full rationale

RecipeGen is a benchmark and dataset paper rather than a derivation chain, and none of the load-bearing claims reduce to their inputs by construction. The dataset images are sourced externally (Douguo, YouTube, YouCook2), and the T2I, I2V, and T2V experiments in Tables 3 and 4 compare external generative models against the benchmark using fixed metrics; these evaluations are not fitted to the dataset's own outputs. The closest candidate for circularity is the YouCook2 preprocessing in Section 3.2, where keyframes are chosen by maximum CLIP similarity to step text while Section 3.3 defines Step Faithfulness as CLIP similarity between an image and its step caption. However, the GF/SF values reported in Table 1 are introduced in Section 3.1 immediately after the 21,944-recipe image-text quality-control stage and before the video subset is constructed, so the reported faithfulness scores are computed on independently sourced step images rather than on the CLIP-argmax keyframes. The video keyframe selection remains a correctness risk (reassembling clips after deleting person frames can break YouCook2 step intervals), but that is an external-validity concern, not a definitional circularity. There are no load-bearing self-citations or imported uniqueness theorems; GPT-4o, CLIP, YOLOv8, and Qwen2.5-VL are external systems. Accordingly, the paper receives a circularity score of 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper does not fit physical constants; its central claims rest on the reliability of source annotations and of large-model judgments. Cleaning thresholds, person-detection cutoffs, and a three-epoch fine-tuning budget are hand-chosen. The dataset and metric definitions therefore carry domain assumptions that are not independently verified.

free parameters (3)
  • Person-presence cutoff for YouCook2 filtering = 50% of video duration
    Hand-chosen threshold in Section 3.2; videos with over 50% person-visible duration are discarded, which reshapes the video subset.
  • YOLOv8 person confidence threshold = 0.5
    Used in Section 3.2 to detect and remove person-containing frames; no sensitivity analysis is reported.
  • Qwen2.5-VL fine-tuning epochs = 3 epochs
    Ingredient Accuracy (Section 3.3) fine-tunes the VQA model for three epochs; the model's own ingredient-detection accuracy is not reported, so the metric's reliability is unknown.
assumptions (5)
  • domain assumption YouCook2 provides reliable step-level temporal annotations, and the curated subset retains alignment after person-frame removal.
    Paper uses YouCook2 train and validation annotations as ground truth without independent verification; Section 3.2.
  • domain assumption CLIP similarity is a valid measure of semantic alignment between recipe images and step text.
    Used for Goal Faithfulness, Step Faithfulness, and keyframe selection in Sections 3.1 and 3.3; no calibration or human agreement is reported.
  • domain assumption GPT-4o text cleaning and interaction-caption generation produce trustworthy content consistent with visible cooking actions.
    GPT-4o is used both to refine recipe text (Section 3.1) and to generate captions for Interaction Faithfulness (Section 3.3), with no systematic error analysis.
  • domain assumption DINOv2 embeddings and l2 distance measure visual cross-step consistency.
    Cross-Step Consistency metric in Section 3.3 relies on this without validation.
  • domain assumption Human review of 2,000 sampled recipes is representative of the whole benchmark of 26,453 recipes.
    Section 3.1; only about 7.6% of recipes are checked, with 12 mislabeled dishes and 17 missing-step-image cases found, yet those error types are treated as negligible.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RecipeGen: A Step-Aligned Multimodal Benchmark for Real-World Recipe Generation." pith.science (2026). https://pith.science/paper/C4OVUXBT

@misc{pith2026250606733,
  author       = {Pith},
  title        = {Pith review of: RecipeGen: A Step-Aligned Multimodal Benchmark for Real-World Recipe Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C4OVUXBT}},
  note         = {Machine review of arXiv:2506.06733}
}
read the original abstract

Creating recipe images is a key challenge in food computing, with applications in culinary education and multimodal recipe assistants. However, existing datasets lack fine-grained alignment between recipe goals, step-wise instructions, and visual content. We present RecipeGen, the first large-scale, real-world benchmark for recipe-based Text-to-Image (T2I), Image-to-Video (I2V), and Text-to-Video (T2V) generation. RecipeGen contains 26,453 recipes, 196,724 images, and 4,491 videos, covering diverse ingredients, cooking procedures, styles, and dish types. We further propose domain-specific evaluation metrics to assess ingredient fidelity and interaction modeling, benchmark representative T2I, I2V, and T2V models, and provide insights for future recipe generation models. Project page is available now.

Figures

Figures reproduced from arXiv: 2506.06733 by the authors.

Figure 1
Figure 1. RecipeGen Benchmark Overview. A comprehensive real-world benchmark for multimodal recipe generation, sup [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of Our RecipeGen Benchmark. Dataset #Recipes Images Videos Modalities Tasks Source Data Faithfulness Goal (GF) Step (SF) VGSI-Recipe [36] 1,157 6,417 0 T, I T2I wikiHow 27.35 29.03 YouCook2 [41] 2,000 0 2,000 T, V VU YouTube 27.77 29.93 RecipeGen 26,453 196,724 4,491 T, I, V T2I, I2V, T2V Douguo, YouTube, YouCook2 28.94 30.31 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of T2I models on Recipe Image Generation task. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 28 canonical work pages

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report.arXiv preprint arXiv:2303.08774 (2023)

    arXiv 2023

  2. [2]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. 2025. Qwen2.5-VL Technical Rep...

    arXiv 2025

  3. [3]

    Devansh Batra, Nirav Diwan, Utkarsh Upadhyay, Jushaan Singh Kalra, Tript Sharma, Aman Kumar Sharma, Dheeraj Khanna, Jaspreet Singh Marwah, Sri- lakshmi Kalathil, Navjot Singh, et al. 2020. Recipedb: a resource for exploring recipes.Database2020 (2020), baaa077

    work page 2020

  4. [4]

    Michał Bień, Michał Gilski, Martyna Maciejewska, Wojciech Taisner, Dawid Wis- niewski, and Agnieszka Lawrynowicz. 2020. RecipeNLG: A Cooking Recipes Dataset for Semi-Structured Text Generation. InProceedings of the 13th Interna- tional Conference on Natural Language Generation, Brian Davis, Yvette Graham, John Kelleher, and Yaji Sripada (Eds.). Associatio...

    work page doi:10.18653/v1/2020.inlg-1.4 2020

  5. [5]

    Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. 2014. Food-101–mining discriminative components with random forests. InProceedings of the European Conference on Computer Vision. Springer, 446–461

    work page 2014

  6. [6]

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. 2021. Emerging properties in self-supervised vision transformers. InProceedings of the IEEE/CVF international conference on computer vision. 9650–9660

    2021

  7. [7]

    Anthony Chen, Jianjin Xu, Wenzhao Zheng, Gaole Dai, Yida Wang, Renrui Zhang, Haofan Wang, and Shanghang Zhang. 2024. Training-free regional prompting for diffusion transformers.arXiv preprint arXiv:2411.02395(2024)

    arXiv 2024

  8. [8]

    Jingjing Chen and Chong-Wah Ngo. 2016. Deep-based ingredient recognition for cooking recipe retrieval. InProceedings of the 24th ACM international conference on Multimedia. 32–41

    work page 2016

Show all 44 references
  1. [9]

    Prateek Chhikara, Dhiraj Chaurasia, Yifan Jiang, Omkar Masur, and Filip Ilievski

  2. [10]

    Alexey Dosovitskiy. 2020. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929(2020)

    2020 arXiv

  3. [11]

    Lee, Ke Shu, Palakorn Achananuparp, Philips Kokoh Prasetyo, Yue Liu, Ee-Peng Lim, and Lav R Varshney

    Helena H. Lee, Ke Shu, Palakorn Achananuparp, Philips Kokoh Prasetyo, Yue Liu, Ee-Peng Lim, and Lav R Varshney. 2020. RecipeGPT: Generative pre-training based cooking recipe generation and evaluation system. InCompanion Proceedings of the Web Conference 2020. 181–184

    2020

  4. [12]

    Fangda Han, Ricardo Guerrero, and Vladimir Pavlovic. 2020. CookGAN: Meal im- age synthesis from ingredients. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 1450–1458

    2020

  5. [13]

    Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. 2024. In-context lora for diffusion transformers.arXiv preprint arXiv:2410.23775(2024)

    2024 arXiv

  6. [14]

    Shuqiang Jiang, Weiqing Min, Linhu Liu, and Zhengdong Luo. 2020. Multi-Scale Multi-View Deep Feature Aggregation for Food Recognition.IEEE Transactions on Image Processing29 (2020), 265–276. doi:10.1109/TIP.2019.2929447

    2020

  7. [15]

    2023.Ultralytics YOLO

    Glenn Jocher, Jing Qiu, and Ayush Chaurasia. 2023.Ultralytics YOLO. https: //github.com/ultralytics/ultralytics

    2023

  8. [16]

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. 2024. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603 (2024)

    2024 arXiv

  9. [17]

    Black Forest Labs. 2024. Flux.1 AI. https://flux1ai.com/. Accessed: 2025-04-04

    2024

  10. [18]

    Xing Lan, Jiayi Lyu, Hanyu Jiang, Kun Dong, Zehai Niu, Yi Zhang, and Jian Xue

  11. [19]

    Zhiming Liu, Kai Niu, and Zhiqiang He. 2023. ML-CookGAN: Multi-label gen- erative adversarial network for food image generation.ACM Transactions on Multimedia Computing, Communications and Applications19, 2s (2023), 1–21

    2023

  12. [20]

    Javier Marın, Aritro Biswas, Ferda Ofli, Nicholas Hynes, Amaia Salvador, Yusuf Aytar, Ingmar Weber, and Antonio Torralba. 2021. Recipe1m+: A dataset for learning cross-modal embeddings for cooking recipes and food images.IEEE Transactions on Pattern Analysis and Machine Intell...

    2021

  13. [21]

    Sachit Menon, Ishan Misra, and Rohit Girdhar. 2024. Generating Illustrated Instructions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6274–6284

    2024

  14. [22]

    Weiqing Min, Shuqiang Jiang, and Ramesh Jain. 2019. Food recommendation: Framework, existing solutions, and challenges.IEEE Transactions on Multimedia 22, 10 (2019), 2659–2671

    2019

  15. [23]

    Weiqing Min, Shuqiang Jiang, Linhu Liu, Yong Rui, and Ramesh Jain. 2019. A survey on food computing.ACM Computing Surveys (CSUR)52, 5 (2019), 1–36

    2019

  16. [24]

    Weiqing Min, Zhiling Wang, Yuxin Liu, Mengjiang Luo, Liping Kang, Xiaoming Wei, Xiaolin Wei, and Shuqiang Jiang. 2023. Large scale visual food recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence45, 8 (2023), 9932– 9949

    2023

  17. [25]

    Siyuan Pan, Ling Dai, Xuhong Hou, Huating Li, and Bin Sheng. 2020. ChefGAN: Food image generation from recipes. InProceedings of the 28th ACM International Conference on Multimedia. 4244–4252

    2020

  18. [26]

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. 2023. Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952 (2023)

    2023 arXiv

  19. [27]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. InInternational conference on machine learnin...

    2021

  20. [28]

    Markus Rokicki, Christoph Trattner, and Eelco Herder. 2018. The impact of recipe features, social cues and demographics on estimating the healthiness of online recipes. InProceedings of the international AAAI conference on web and social media, Vol. 12

    2018

  21. [29]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10684–10695

    2022

  22. [30]

    Stability AI. 2022. Stable Diffusion v2.1 Model Card. https://huggingface.co/ stabilityai/stable-diffusion-2-1. Accessed: 2025-04-11

    2022

  23. [31]

    Stability AI. 2024. Stable Diffusion 3.5 Large Model Card. https://huggingface. co/stabilityai/stable-diffusion-3.5-large. Accessed: 2025-04-11

    2024

  24. [32]

    Hao Wang, Guosheng Lin, Steven CH Hoi, and Chunyan Miao. 2020. Structure- aware generation network for recipe generation from images. InProceedings of the European Conference on Computer Vision. Springer, 359–374

    2020

  25. [33]

    Hao Wang, Guosheng Lin, Steven CH Hoi, and Chunyan Miao. 2022. Learn- ing structural representations for recipe generation and food retrieval.IEEE Transactions on Pattern Analysis and Machine Intelligence45, 3 (2022), 3363–3377

    2022

  26. [34]

    Wenjie Wang, Ling-Yu Duan, Hao Jiang, Peiguang Jing, Xuemeng Song, and Liqiang Nie. 2021. Market2Dish: health-aware food recommendation.ACM Trans- actions on Multimedia Computing, Communications, and Applications (TOMM)17, 1 (2021), 1–19

    2021

  27. [35]

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing13, 4 (2004), 600–612

    2004

  28. [36]

    Yue Yang, Artemis Panagopoulou, Qing Lyu, Li Zhang, Mark Yatskar, and Chris Callison-Burch. 2021. Visual Goal-Step Inference using wikiHow.arXiv preprint arXiv:2104.05845(2021)

    2021 arXiv

  29. [37]

    Yuehao Yin, Huiyan Qi, Bin Zhu, Jingjing Chen, Yu-Gang Jiang, and Chong-Wah Ngo. 2023. Foodlmm: A versatile food assistant using large multi-modal model. arXiv preprint arXiv:2312.14991(2023)

    2023 arXiv

  30. [38]

    Yudong Zhang, Lijia Deng, Hengde Zhu, Wei Wang, Zeyu Ren, Qinghua Zhou, Siyuan Lu, Shiting Sun, Ziquan Zhu, Juan Manuel Gorriz, et al . 2023. Deep learning in food category recognition.Information Fusion98 (2023), 101859

    2023

  31. [39]

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. 2024. Open-sora: Democratizing efficient video production for all.arXiv preprint arXiv:2412.20404(2024)

    2024 arXiv

  32. [40]

    Luowei Zhou, Chenliang Xu, and Jason Corso. 2018. Towards automatic learn- ing of procedures from web instructional videos. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 32

    2018

  33. [41]

    Luowei Zhou, Chenliang Xu, and Jason J Corso. 2017. Towards Automatic Learn- ing of Procedures from Web Instructional Videos.arXiv preprint arXiv:1703.09788 (2017)

    2017 arXiv

  34. [42]

    Pengfei Zhou, Weiqing Min, Chaoran Fu, Ying Jin, Mingyu Huang, Xiangyang Li, Shuhuan Mei, and Shuqiang Jiang. 2024. FoodSky: A Food-oriented Large Language Model that Passes the Chef and Dietetic Examination.arXiv preprint arXiv:2406.10261(2024)

    2024 arXiv

  35. [2023]

    Foodsam: Any food segmentation.IEEE Transactions on Multimedia(2023)

    2023

  36. [2024]

    InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Fire: Food image to recipe generation. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 8184–8194

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.