Pith. sign in

REVIEW 5 major objections 5 minor 22 references

The paper claims that for text-to-video models, training-data distribution and caption correctness dominate behavior: a balanced mix of simple and complex clips matches or beats specialists, precision of captions matters more than completen

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 14:17 UTC pith:6CZ6GZAI

load-bearing objection A well-constructed synthetic testbed with plausible qualitative findings; treat the precision-cliff numbers as provisional until seeds are reported. the 5 major comments →

arxiv 2607.18789 v1 pith:6CZ6GZAI submitted 2026-07-21 cs.CV

Moving Alphabet: A Controlled Study of Training Data for Text-to-Video Generation

Amber Yijia Zheng , Lu Liu , Raymond A. Yeh , Xi Yin This is my paper
classification cs.CV
keywords text-to-video generationtraining data curationcaption qualityprecision and recalldata mixingsynthetic testbedMoving Alphabetclassifier-free guidance
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper isolates how training data—the distribution of video content and the quality of paired captions—shapes text-to-video models, using a synthetic testbed called Moving Alphabet that renders letters with known attributes so every caption error is measurable. It establishes three things: equal mixing of simple and complex scenes (and of short and long clips) generalizes better than specialist training; caption precision (correctness) is a more damaging axis than recall (completeness), with a sharp quality cliff when even a few percent of caption claims go wrong; and classifier-free guidance or clean-data fine-tuning only partially recovers lost quality, leaving severe corruption permanent. A sympathetic reader should care because these are concrete, actionable guidelines for curating the large real-world datasets behind commercial video generators.

Core claim

The central discovery is that text-to-video model quality is bounded primarily by caption correctness and by a balanced, diverse training distribution, rather than by model capacity alone. Using Moving Alphabet, where every attribute is known and captions can be corrupted along two independent axes, the paper shows that an equal mix of 1-, 2-, and 3-letter scenes (and of 2s/4s/8s clips) matches or exceeds specialists on their own distributions; that reducing caption precision from 1.0 to 0.95 already costs 5.8 of the 8.1 dB lost across the full corruption range, while recall loss is more gradual; and that corrupted-caption pretraining requires 2–4× more compute to reach the same validation l

What carries the argument

The Moving Alphabet testbed is the load-bearing object: a procedural renderer that places one-to-three letters with independently sampled fonts, colors, sizes, rotations, positions, and velocities, paired with templated captions derived from exact ground-truth metadata. Because every attribute is known, the authors can corrupt captions along separable axes (precision α: fraction of stated attributes correct; recall β: fraction of true attributes mentioned), compose arbitrary data mixtures, and evaluate with exact per-video metrics (foreground PSNR and per-attribute classifiers calibrated on ground-truth videos). The same testbed supports the compute-efficiency measurement (validation loss wi

Load-bearing premise

That Moving Alphabet—a synthetic world of letters, templated captions, and independently applied random caption corruption—captures the structure of real-world video data curation, so that the measured quantitative effects (the precision cliff, the compute multiplier, the recovery caps) transfer to natural videos and VLM-generated captions.

What would settle it

Train a video model on a real, naturally captioned dataset (e.g., web clips with automatic captions) and measure generation quality as a function of independently measured caption precision, then check whether the sharp cliff between α=0.99 and α=0.95 replicates; or test whether corrupting real captions by swapping attributes (rather than templated metadata) reproduces the >2× compute gap. If real caption errors are content-correlated or the cliff is absent, the quantitative recommendations fail to transfer.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Data curators should aim for a balanced mix of simple and complex clips, and of short and long clips, rather than filtering toward either extreme; equal mixing matches or beats specialists at no extra data cost.
  • Under a fixed captioning budget, invest in verifying the correctness of stated attributes before maximizing coverage, because precision is the more damaging failure mode and also the one that collapses per-attribute controllability.
  • Caption quality is a compute multiplier: models trained on corrupted captions need 2–4× more compute to reach the same validation loss, so improving captions can be cheaper than extending training.
  • Post-hoc fixes are partial: CFG recovers attribute accuracy for moderate corruption but not pixel fidelity, and fine-tuning recovers up to 55% of lost FG PSNR only for mild corruption; severe preprocessing damage is permanent.
  • Systematically omitted attributes do not stay diverse: unless a redundant cue (like start and end positions) pins them down, the model collapses to seed-dependent defaults, making those attributes uncontrollable.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The quantitative thresholds (precision cliff near α=0.95, 55% recovery cap, 2–4× compute multiplier) are measured on synthetic letters; the directional principles—precision-first, balanced mixing, early caption investment—likely transfer to real T2V data, but exact numbers probably shift with captioner error structure and scene distribution.
  • The balanced-mixing result suggests a candidate data-mixing law for video: for any content axis (resolution, scene type, motion complexity), equal mixing may dominate specialist training; this is directly testable on a real natural-video generator by comparing equal versus dominated mixes at fixed compute.
  • The compute-multiplier framing implies an economic rule for dataset builders: spending on caption quality can substitute for training compute roughly 2–4×, giving a concrete cost-benefit target for captioning pipelines.
  • The attribute-omission bias finding predicts that automated captioners that under-describe attributes (e.g., object size) will produce models with arbitrary per-seed defaults; a testable extension is to measure whether real caption pipelines generate controllable attributes in proportion to how often those attributes are explicitly stated.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper introduces Moving Alphabet, a procedural testbed of letters with known ground-truth metadata, and uses it to study how training data composition and caption quality affect an 800M-parameter text-to-video diffusion transformer. Three empirical findings are reported: (i) balanced mixtures over content complexity and video duration match or exceed specialist models on their own distributions; (ii) caption precision is more damaging than recall, with a sharp cliff between alpha=0.99 and alpha=0.95, and corrupted captions cost 2-4x more compute to reach a fixed validation loss; (iii) classifier-free guidance and finetuning on clean data provide only partial recovery for moderate corruption and essentially none for severe corruption. The authors draw practical recommendations for real-world T2V data curation.

Significance. If the findings hold, Moving Alphabet is a valuable controlled instrument for studying data curation axes that are entangled in real-world datasets. The evaluation design is unusually careful: attribute classifiers are calibrated on ground-truth videos, a VAE reconstruction ceiling is established, and data budgets are mostly held fixed while token counts are reported. These strengths make the internal comparisons meaningful in principle. However, the headline quantitative claims currently rest on single training runs per condition and on a validation-loss threshold selected after inspecting convergence behavior, so the practical recommendations are not yet established at the precision the paper claims. The external-validity question—whether synthetic letters and independent random caption corruption transfer to real T2V curation—also remains untested.

major comments (5)
  1. [§5.1, Fig. 10a] The precision cliff is the backbone of the precision-first recommendation, but the supporting grid is not reproducible as reported. At beta=1.0, FG PSNR is 12.3 dB at alpha=1.0, 12.4 dB at alpha=0.99, 6.59 dB at alpha=0.95, and 6.34 dB at alpha=0.9; at beta=0.9 the grid is non-monotonic in alpha, with 6.93 dB at alpha=0.99 versus 9.70 dB at alpha=0.95. These inconsistencies (0.1-2.8 dB) are typical of run-to-run noise, and no seeds or error bars are reported for the 35-condition grid; Appendix Fig. 11 explicitly labels direction and speed as single-run. The 5.8 dB cliff and the positive precision advantage need seed-level replication or uncertainty quantification before the paper's central recommendation can rest on them.
  2. [§5.2, Fig. 3] The compute-efficiency claim ('2-4x more compute') rests on a validation-loss threshold of 0.035, described as 'the lowest validation loss that all 35 conditions can achieve.' Yet the caption of Fig. 3 and the text state that (alpha=0.3, beta=0.3) never reaches the threshold within the 200-epoch budget, so the threshold is not achievable by all conditions. Because the threshold was chosen after sweeping 0.030/0.035, the normalized compute ratios are a selected outcome. Please report a sensitivity analysis over thresholds and state the convergence fraction per condition.
  3. [§5.1 vs §5.2] There is a direct inconsistency in the training protocol. §5.1 says the uniform 200-epoch budget is 'large enough for even the slowest condition to converge, so all conditions are compared at a fully-trained point'; §5.2 reports a condition that never converges to the threshold, and Fig. 10a shows substantial quality differences at the final checkpoint. If different checkpoints or budgets are used across the two analyses, that must be stated; as written, the comparison of final quality and the compute-to-threshold analysis are mutually incompatible.
  4. [§6.2, Figs. 6-8] The finetuning recovery percentages (55%, 43%, etc.) and the claim that mild recall corruption is 'fully recoverable' are based on a single finetuning run per condition and a single evaluation checkpoint. Given the non-monotonicity in the pretraining grid, these recovery numbers need the same seed-level replication before the asymmetry between precision and recall recovery is established. In addition, the stated 55% recovery at alpha=0.9, 1L appears to be (9.6-7.2)/(11.7-7.2)=0.53, not 0.55.
  5. [§3 and §7] External validity: the paper's practical recommendations (balanced mixes, precision-first captioning) are framed for real T2V data curation, but the evidence comes from a synthetic testbed where caption corruption is applied independently and uniformly to templated attributes. Real caption errors are often content-dependent and structured, and compositional letter scenes may not capture natural-scene failure modes. The assumption that Moving Alphabet is an adequate proxy is untested. Please either soften the generalization claims or provide a small transfer check, e.g., a real video dataset with known metadata or a natural-image caption-corruption study, to show that the qualitative findings survive outside the testbed.
minor comments (5)
  1. [§5.1, Fig. 2c] The main text says the precision advantage is positive for all four attributes, but Appendix Fig. 11 shows that precision wins only 6/10 pairs for direction and speed; the text should reflect this noisiness rather than presenting all attributes as equally decisive.
  2. [§4.2, Tab. 3] The text says Mix Equal 'stays within 0.5dB of the best pure model at each duration,' but at 4s and 8s Mix Equal exceeds the pure specialists by 1.2dB and 1.1dB, respectively. The sentence should be reworded.
  3. [§3.2] The baseline section states that 'all metrics plateau by epoch 100' but Fig. 9 shows only the Mix Equal condition. Since different corruption conditions may converge at different rates, the checkpoint justification should be stated per experiment or supported by additional convergence plots.
  4. [Appendix Fig. 11] The axis labels in Fig. 11 render as 'Recall-favored ( =lo, =hi)' and 'Precision-favored ( =hi, =lo)' with missing variables; this should be fixed for readability.
  5. [General] No code or dataset release is mentioned. For a testbed whose main contribution is controlled data generation and evaluation, releasing the generator, corruption code, and evaluation harness would substantially strengthen reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the paper's claims are empirical measurements against independent procedural ground-truth metadata; no fitted quantity is relabeled as a prediction and no self-citation is load-bearing.

full rationale

This paper makes no formal derivation claims; its results are controlled training experiments. The key quantities—FG PSNR, FG MSE, and per-attribute accuracy—are computed by comparing generated videos against procedurally generated ground-truth metadata (§3.3), an external reference not derived from the trained models. The precision (α) and recall (β) corruption grid (§5.1) manipulates input captions, and the reported effects are measured outcomes, not fitted parameters renamed as predictions. The 'precision advantage' (Fig. 2c) is an empirical difference between two measured cells, not an identity. The metric classifiers are calibrated on ground-truth videos (Tab. 1a) and the VAE ceiling is measured independently (Tab. 1b); this is instrument validation, not circularity. The only author-overlapping citation (Make-A-Video, Singer et al. 2022, which includes co-author Xi Yin) appears in related work and is not load-bearing for any conclusion. The appendix's admission that 'direction and speed are noisier (single run, CFG=1.0)' (Fig. 11 caption) and the inconsistency about whether the α=0.3, β=0.3 condition reaches the compute-efficiency threshold (§5.2 says 'lowest validation loss that all 35 conditions can achieve' yet reports '>3.91×' for that condition) are reproducibility/correctness concerns, not circularity. No step in the paper reduces, by definition or self-citation, to its own inputs, so the circularity score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The paper introduces the Moving Alphabet testbed as a constructed instrument rather than postulating an unexplained natural entity. The load-bearing free parameters are the validation-loss threshold and checkpoint selection, plus hand-tuned metric thresholds. The main empirical conclusions rest on domain assumptions about testbed-to-real transfer and the realism of random caption corruption.

free parameters (3)
  • Validation loss threshold for compute comparison = 0.035
    Chosen by sweeping in increments of 0.005 so that most corruption conditions converge within budget; directly determines the reported 2–4× compute multiples (§5.2).
  • Evaluation checkpoint epoch = 100
    Selected as the point where metrics 'plateau' (Appendix Fig. 9); affects all reported results and recovery percentages.
  • Attribute classifier thresholds = FG mask >40/255; color >200; size bins 44/68 px
    Hand-specified thresholds for computing foreground PSNR, color, and size accuracy (§3.3, Appendix C). Reasonable but not fitted independently; could influence reported accuracies.
axioms (3)
  • domain assumption Synthetic moving letters on a black background are a faithful proxy for real-world video content complexity and temporal dynamics.
    Invoked in §3 and used to justify generalizing recommendations to real T2V systems; if false, the quantitative findings may not transfer.
  • domain assumption Random independent corruption of caption attributes (replacing with random values, dropping) models real caption errors such as hallucination and omission.
    Used throughout §5 to draw conclusions about precision vs. recall; real caption errors are likely correlated with content and model confidence, so this is a significant simplification.
  • domain assumption Automated attribute classifiers using ground-truth ROIs and foreground masks measure generation quality validly.
    The paper calibrates classifiers on ground truth (§3.3, Table 1), but evaluation relies on trajectory ROIs and simple pixel statistics that may be easier than natural-scene evaluation.

pith-pipeline@v1.3.0-alltime-deepseek · 220 in / 6526 out tokens · 105917 ms · 2026-08-01T14:17:20.215250+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Moving Alphabet: A Controlled Study of Training Data for Text-to-Video Generation." pith.science (2026). https://pith.science/paper/6CZ6GZAI

@misc{pith2026260718789,
  author       = {Pith},
  title        = {Pith review of: Moving Alphabet: A Controlled Study of Training Data for Text-to-Video Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6CZ6GZAI}},
  note         = {Machine review of arXiv:2607.18789}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Text-to-video generation has advanced significantly over the past five years through scaling of model size, data, and compute. Unlike model architecture, training data is often underexplored. Real-world data curation is complex and non-trivial, involving clip selection from raw videos and captioning to create video-text pairs for learning text-to-video mappings. We study how data distribution and caption quality impact text-to-video models. To enable controlled experiments, we introduce Moving Alphabet, a procedural testbed that renders letters with varying fonts, colors, sizes, and positions, moving in different directions and speeds against a black background. This design allows precise control over data distribution and caption quality by corrupting ground-truth metadata. Our experiments yield three findings: a) a diverse and balanced distribution of video content and duration is critical for generalization; b) caption quality significantly affects both model performance and training efficiency, suggesting that text-to-video models are bounded by video understanding capabilities; and c) classifier-free guidance and fine-tuning on high-quality data provide partial recovery from models trained on corrupted captions, but cannot fully compensate for poor pre-training data. We believe these insights can inform the development of large-scale text-to-video models, and we advocate for greater attention to the science of pre-training data.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 20 linked inside Pith

  1. [3]

    VideoCrafter2: Overcoming data limitations for high-quality video diffusion models.arXiv preprint arXiv:2401.09047, 2024a

    Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models.arXiv preprint arXiv:2401.09047, 2024a. Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, Li Yuan, Yu Qiao, Dahua L...

    Pith/arXiv arXiv

  2. [5]

    The Pile: An 800GB dataset of diverse text for language modeling.arXiv preprint arXiv:2101.00027,

    Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. The Pile: An 800GB dataset of diverse text for language modeling.arXiv preprint arXiv:2101.00027,

    Pith/arXiv arXiv

  3. [7]

    Imagen video: High definition video generation with diffusion models.arXiv preprint arXiv:2210.02303,

    Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models.arXiv preprint arXiv:2210.02303,

    Pith/arXiv arXiv

  4. [11]

    Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao

    Blog post. Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation.arXiv preprint arXiv:2401.03048,

    Pith/arXiv arXiv

  5. [12]

    Niklas Muennighoff, Alexander M Rush, Boaz Barak, Teven Le Scao, Aleksandra Piktus, Nouamane Tazi, Sampo Pyysalo, Thomas Wolf, and Colin Raffel

    https://github.com/deepmind/dsprites-dataset/. Niklas Muennighoff, Alexander M Rush, Boaz Barak, Teven Le Scao, Aleksandra Piktus, Nouamane Tazi, Sampo Pyysalo, Thomas Wolf, and Colin Raffel. Scaling data-constrained language models.arXiv preprint arXiv:2305.16264,

    Pith/arXiv arXiv

  6. [13]

    Movie Gen: A cast of media foundation models.arXiv preprint arXiv:2410.13720,

    Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih- Yao Ma, Ching-Yao Chuang, et al. Movie Gen: A cast of media foundation models.arXiv preprint arXiv:2410.13720,

    Pith/arXiv arXiv

  7. [14]

    The layout bet, June 2026.https://reve.com

    Reve Team. The layout bet, June 2026.https://reve.com. Accessed: 2026-06-03. Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. InProc. CVPR,

    2026

  8. [15]

    Make-A-Video: Text-to-video generation without text-video data.arXiv preprint arXiv:2209.14792,

    Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-A-Video: Text-to-video generation without text-video data.arXiv preprint arXiv:2209.14792,

    Pith/arXiv arXiv

  9. [16]

    Kimi-VL technical report.arXiv preprint arXiv:2504.07491,

    Kimi Team, Angang Du, Bohong Yin, Bowei Xing, Bowen Qu, Bowen Wang, Cheng Chen, Chenlin Zhang, Chenzhuang Du, Chu Wei, et al. Kimi-VL technical report.arXiv preprint arXiv:2504.07491,

    Pith/arXiv arXiv

  10. [17]

    InternVid: A large-scale video-text dataset for multimodal understanding and generation.arXiv preprint arXiv:2307.06942,

    Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. InternVid: A large-scale video-text dataset for multimodal understanding and generation.arXiv preprint arXiv:2307.06942,

    Pith/arXiv arXiv

  11. [18]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    Pith/arXiv arXiv

  12. [19]

    CogVideoX: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072,

    16 Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. CogVideoX: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072,

    Pith/arXiv arXiv

  13. [20]

    Data mixing laws: Optimizing data mixtures by predicting language modeling performance.arXiv preprint arXiv:2403.16952,

    Jiasheng Ye, Peiju Liu, Tianxiang Sun, Yunhua Zhou, Jun Zhan, and Xipeng Qiu. Data mixing laws: Optimizing data mixtures by predicting language modeling performance.arXiv preprint arXiv:2403.16952,

    Pith/arXiv arXiv

  14. [21]

    Open-Sora: Democratizing efficient video production for all.arXiv preprint arXiv:2412.20404,

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhao, Yining Li, and Yang You. Open-Sora: Democratizing efficient video production for all.arXiv preprint arXiv:2412.20404,

    Pith/arXiv arXiv

  15. [22]

    4 lists all attributes used in the Moving Alphabet dataset and their possible values

    17 Appendix A Attribute Specifications Tab. 4 lists all attributes used in the Moving Alphabet dataset and their possible values. Table 4Moving Alphabet attributes and their value sets. All attributes are sampled independently and uniformly from the listed values. Category Attribute Values Appearance Letter identity 52 classes (26 uppercase A–Z, 26 lowerc...

    2022

  16. [2017]

    Scaling laws for neural language models.arXiv preprint arXiv:2001.08361,

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.arXiv preprint arXiv:2001.08361,

    Pith/arXiv arXiv 2001

  17. [2018]

    HunyuanVideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603,

    15 Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xu, Bohao Pang, Hao Liu, et al. HunyuanVideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603,

    Pith/arXiv arXiv

  18. [2020]

    Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598,

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598,

    Pith/arXiv arXiv

  19. [2022]

    T2I-CompBench: A comprehensive benchmark for open-world compositional text-to-image generation.arXiv preprint arXiv:2307.06350,

    Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2I-CompBench: A comprehensive benchmark for open-world compositional text-to-image generation.arXiv preprint arXiv:2307.06350,

    Pith/arXiv arXiv

  20. [2023]

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorber, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al

    OpenAI technical report. Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorber, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127,

    Pith/arXiv arXiv

  21. [2024]

    Lumiere: A space-time diffusion model for video generation.arXiv preprint arXiv:2401.12945,

    Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, Oliver Wang, Deqing Sun, Tali Dekel, and Inbar Mosseri. Lumiere: A space-time diffusion model for video generation.arXiv preprint arXiv:2401.12945,

    Pith/arXiv arXiv

  22. [2025]

    TinyStories: How small can language models be and still speak coherent English?arXiv preprint arXiv:2305.07759,

    Ronen Eldan and Yuanzhi Li. TinyStories: How small can language models be and still speak coherent English?arXiv preprint arXiv:2305.07759,

    Pith/arXiv arXiv