Pith. sign in

REVIEW 4 major objections 6 minor 30 references

Exploring In-Image Machine Translation with Real-World Background

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 DebackX handles complex-scenario in-image translation by decomposing the source image into a background and a text-image, translating the text-image, and fusing the two back together.

desk verdict A useful new dataset and a visually better model, but the translation-quality claim is confounded by a pre-training advantage and sits inside OCR noise. read the letter →

arxiv 2505.15282 v1 pith:S3MP54P2 submitted 2025-05-21 cs.CL cs.CV

Yanzhi Tian , Zeming Liu , Zhengyang Liu , Yuhang Guo This is my paper · ORCID
classification cs.CLcs.CV
keywords in-imagemachinetranslationcomplexbackgroundtext-imageseparationimagefusionvectorquantizationsubtitlefontconsistencybenchmarkdataset
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

In-image machine translation (IIMT) has mostly been tested on text with black fonts on white or single-color backgrounds, a setting the paper argues is far from real use. The paper introduces complex-scenario IIMT, where captions sit on top of real-world images, and builds IIMT30k, a dataset of German-English subtitle images derived from Multi30k backgrounds. It then proposes DebackX, which separates the source image into a background and a text-image, translates only the text-image through a vector-quantized image translation model, and fuses the translated text-image with the preserved background. The paper reports that DebackX attains better translation quality and visual effect than previous IIMT models and OCR-NMT-render pipelines on this dataset, and that it keeps rendered fonts consistent with the source. The reason a reader should care is that this recasts IIMT from a full-image generation problem into a decomposition-plus-translation problem, making the task tractable on realistic backgrounds.

What carries the argument

The load-bearing machinery is the three-module DebackX architecture. The Text-Image Background Separation module uses two ViT encoders and two ViT decoders to output a background image and a text-image. The Image Translation module encodes the text-image with a ViT, quantizes each feature into one of 8,192 learned codebook vectors, translates the resulting source code sequence into a target code sequence with a code encoder-decoder, and uses a Pivot Decoder trained on a text-image translation (TIT) auxiliary task to bring semantic text information into the code sequence. The Text-Image Background Fusion module encodes the background and the target text-image separately and sums their features before a ViT decoder produces the final image. Pre-training the Image Translation module on rendered text-image pairs from parallel corpora is what lets the model leverage large amounts of text data despite the small IIMT30k training set.

What would settle it

Measure DebackX and the best OCR-NMT-render baseline on a held-out set of real subtitle screenshots or video frames with varied perspective, lighting, and font sizes, scoring with the same OCR-BLEU and FID protocol; if the quality gap narrows or reverses, the synthetic IIMT30k benchmark is not representative.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that complex-scenario IIMT is best solved by never generating the whole image at once. DebackX first decomposes the source into a background image and a text-image, translates the text-image from German to English (or the reverse) as a sequence of discrete codebook tokens, and then reconstructs the target image by fusing the untouched background with the translated text-image. This separation avoids the two failure modes of earlier systems: OCR-to-NMT error propagation and text-region removal that damages visual quality. The paper supports the claim with ablations showing that, in the no-pretraining setting, removing the separation/fusion modules drops De-En valid BLEU from 7.4 to 1.2, and removing the pivot decoder's TIT auxiliary task drops it to 1.5.

Load-bearing premise

The load-bearing premise is that IIMT30k, built by resizing Multi30k images, rendering captions with Pillow, and cropping to 48x512 strips, faithfully represents real-world complex-background IIMT; if actual subtitle conditions such as perspective, lighting, layout, and longer text behave differently, the reported gains may not transfer.

Editorial extensions

If this is right

  • Complex-scenario IIMT becomes a well-defined benchmark task: IIMT30k gives later systems a fixed test set with real-world-like backgrounds and three font conditions.
  • Because only the text-image needs translation, large parallel text corpora can be rendered into text-image pairs for pre-training, so model scale no longer depends on the amount of complex-background image data available.
  • Separating background and text preserves the background essentially intact, eliminating the visual artifacts of OCR-NMT-render approaches that blank out text regions.
  • Font consistency becomes measurable: the paper reports 97.3% consistency when training on two fonts and 96.5% on three, suggesting the text-image representation carries style information through translation.
  • The ablation results imply that direct whole-image generation is not a viable route for complex-scenario IIMT and that decomposition into text and background is required for the reported quality.

Reading between the lines

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

  • If IIMT30k's rendering recipe transfers to real video subtitles, the same decompose-translate-fuse design could be applied to subtitle translation in video with temporal smoothing added; the paper does not test video, moving text, or perspective distortion.
  • The evaluation reads output images with EasyOCR and scores the recognized text with BLEU/COMET, so reported translation quality is capped by OCR accuracy; on golden output images OCR BLEU is only 64-81, meaning the model's true translation quality could be higher than the headline numbers.
  • DebackX's background-versus-text separation may generalize to other tasks such as OCR front ends and scene-text editing, since the paper itself observes that OCR errors are lower on text-images than on full images.
  • Because the paper's stated limitations are that only basic ViT and vector-quantization modules were used and training is multi-stage, upgrading those components or collapsing the stages into one training pass are open directions rather than tested claims.
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 a complex-background version of In-Image Machine Translation (IIMT), where the text to be translated is overlaid on real-world photographic backgrounds rather than on plain white or single-color backgrounds. It constructs a synthetic benchmark, IIMT30k, by resizing Multi30k images to 512x512, rendering German and English captions with Pillow, and cropping to 48x512 strips in three fonts. The proposed model, DebackX, separates the source image into a background image and a text-image, translates the text-image directly via a codebook-based image-to-image translation model with an auxiliary text-image translation (TIT) task and a Pivot Decoder, and fuses the translated text-image with the background to produce the output image. The paper reports BLEU, COMET, and FID results against image-generation baselines, TIT-Render pipelines, and Translatotron-V, and claims improvements in both translation quality and visual effect. It also includes pre-training studies, ablations, a multi-font adaptation study, and qualitative comparisons with TIT-Render and GPT-4o.

Significance. If the claims were fully supported, the paper would be a useful step toward making IIMT applicable to realistic subtitle scenarios. The separation-then-fusion design is a reasonable way to preserve the background while changing only the text, and the reported FID improvements are large (e.g., De-En test FID 9.0 for DebackX vs. 10.1 for Translatotron-V and 21.3 for VQGAN). The IIMT30k dataset, with three fonts and explicit training/validation/test splits, plus the released code and data, are concrete community assets. The paper is also transparent in several places: it reports an ablation without pre-training, provides an OCR error analysis in Appendix D, and gives detailed hyperparameters. However, the translation-quality claim is only partially supported: the main comparison is confounded by the use of 100K additional text-image pre-training pairs for DebackX, the COMET scores are not consistently better, and the OCR-based evaluation is noisy enough that the reported BLEU gaps may not be significant.

major comments (4)
  1. [§6.3, Table 1; §7.1, Table 2] The translation-quality comparison in Table 1 is confounded by pre-training data. The DebackX row is the '+IWSLT PT' configuration from Table 2, which is pre-trained on 100K additional German-English text-image pairs constructed from IWSLT before fine-tuning on IIMT30k-TNR. Table 2 shows that without this pre-training, De-En valid BLEU drops from 14.9 to 7.4, which is below the 14.2 reported for McTIT-Render in Table 1. No baseline in Table 1 receives an equivalent data increment. Please report a data-matched comparison (for example, pre-training the TIT-Render or McTIT pipelines on the same constructed text-image pairs, or evaluating DebackX without pre-training against the full baseline set) or explicitly limit the claim to the pre-trained recipe.
  2. [§6.3, Table 1] The claim of 'better translation quality' is not consistent across the two metrics. While DebackX has higher BLEU than McTIT-Render on every column, its COMET scores are lower on De-En valid (51.2 vs. 53.5), En-De valid (42.2 vs. 43.3), and En-De test (40.0 vs. 42.8). Please either explain this discrepancy or temper the conclusion to 'higher BLEU' rather than 'better translation quality' as a general statement.
  3. [§6.1; Appendix D] The OCR-based evaluation is too noisy to support differences of the size reported. Appendix D states that even on golden output images, OCR WER is 0.23–0.32 on real images and BLEU is no higher than 81.0; the BLEU gaps in Table 1 are as small as 0.7–1.1 points. Without confidence intervals, significance tests, or a human evaluation of translation adequacy, the observed differences may lie within OCR noise. Please add error bars or bootstrap resampling, use a second OCR engine as a sanity check, or include a human evaluation on a sample.
  4. [§3] The synthetic IIMT30k construction is a proxy for real-world subtitles, but the paper asserts without supporting evidence that it is 'comparable with real-world video subtitle images'. The 48x512 crops with Pillow-rendered single-line captions omit perspective distortion, lighting variation, multi-line text, and the varied layouts of real subtitles, so the title-level claim of 'real-world background' should be scoped. Please either validate on a small real-world subtitle test set or explicitly restrict the conclusions to the synthetic rendering conditions.
minor comments (6)
  1. [§6.1] The word 'calaulted' should be 'calculated'.
  2. [Table 1] Table 1 contains formatting errors such as missing spaces in cells like '46.310.2 /43.9137.4' and '14.2 /53.511.7'; please reformat all cells consistently for readability.
  3. [Appendix A; Table 6] Appendix A states that all ViT encoders and decoders have d_ff=2,048, while Table 6 lists d_ff values of 512, 1,024, 2,048, and 4,096 for the Code Encoder, Pivot Decoder, and Code Decoder; please reconcile the descriptions.
  4. [§4.2] The notation PivotDecoder(E_TIT, H_E_code, H_E_code) should be defined explicitly in terms of Query, Key, and Value to avoid confusion about which representation is used in cross-attention.
  5. [§7.4] The GPT-4o comparison is based on only two cases and should be described as anecdotal rather than as evidence that GPT-4o generally fails at IIMT.
  6. [Limitations] The Limitations section lists module choice and computational cost as limitations but not the synthetic nature of the benchmark or the OCR-evaluation noise; adding these would make the limitations more complete.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: DebackX is an empirical architecture paper whose claims rest on held-out evaluations, not on definitions or self-citations.

full rationale

The paper's derivation chain is empirical rather than definitional. IIMT30k is constructed independently from Multi30k images and captions rendered with Pillow, and DebackX is trained and evaluated on held-out splits of that dataset (Section 3; Section 6.2). Translation quality is measured by OCR-then-BLEU/COMET against reference captions (Section 6.1); the model is not trained to optimize OCR output or the evaluation metric, so the metric is not equivalent to the training objective. The main comparison in Table 1 uses DebackX with an extra IWSLT text-image pre-training stage, and the paper transparently reports the without-pre-training BLEU (7.4 vs. 14.9 in Table 2) and the ablation in Table 3. This is a training-data confound that weakens the fairness of the headline comparison, but it is not circularity: no fitted parameter is renamed as a prediction, and the pre-training advantage is stated as an explicit data advantage rather than derived from the model. The only self-citation (Tian et al., 2023) supports the general claim that OCR-NMT cascades suffer error propagation; it is corroborated by an independent citation (Lan et al., 2023) and is not load-bearing for the architecture's identity. The dataset's synthetic nature and OCR evaluation noise are acknowledged in Section 3 and Appendix D as limitations on external validity, not as circular reasoning. No equation in the paper reduces to its own input by construction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The load-bearing ingredients are the synthetic benchmark, the OCR-based evaluation, and the learned separation and fusion modules, all of which rest on domain assumptions rather than external ground truth.

free parameters (4)
  • perceptual loss weight lambda_p = 0.1
    Hand-set in Eq. 3 and Eq. 4; weights L2 vs LPIPS losses in separation and VQ training. Not fitted to data, but reported results depend on this choice.
  • codebook size and dimension = V=8192, dim=32
    Chosen ad hoc following vector-quantize-pytorch defaults; affects tokenization granularity of text-images.
  • model hyperparameters (d_model, layers, heads) = multiple (Table 6)
    Set to match parameter counts across systems, e.g., DebackX main uses d_model 256/1024, layers 3/6. Not derived from theory.
  • pre-training corpus sizes = 100K IWSLT; 1M WMT14
    Used in Table 2 to show scaling; the choice is arbitrary and directly affects final BLEU scores.
assumptions (4)
  • domain assumption EasyOCR provides sufficiently accurate recognition for BLEU/COMET evaluation of generated images.
    All translation metrics pass through OCR; Appendix D shows golden images achieve only 64.6 to 81.0 BLEU and WER up to 0.32, so OCR error materially depresses scores.
  • ad hoc to paper The IIMT30k synthetic rendering is representative of real-world complex-background subtitle images.
    Section 3 states this comparability without validation; there is no real-world test set to confirm transfer.
  • domain assumption The Text-Image Background Separation and Fusion models can decompose and reconstruct images without losing background or font information.
    Sections 4.1 and 4.3 assume this decomposition is invertible enough for downstream translation; separation quality is not measured directly.
  • standard math FID computed with pytorch-fid is a valid proxy for visual effect of IIMT outputs.
    Widely used in image generation, but FID does not directly measure text legibility or font consistency; the paper supplements with manual font consistency checks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring In-Image Machine Translation with Real-World Background." pith.science (2026). https://pith.science/paper/S3MP54P2

@misc{pith2026250515282,
  author       = {Pith},
  title        = {Pith review of: Exploring In-Image Machine Translation with Real-World Background},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S3MP54P2}},
  note         = {Machine review of arXiv:2505.15282}
}
read the original abstract

In-Image Machine Translation (IIMT) aims to translate texts within images from one language to another. Previous research on IIMT was primarily conducted on simplified scenarios such as images of one-line text with black font in white backgrounds, which is far from reality and impractical for applications in the real world. To make IIMT research practically valuable, it is essential to consider a complex scenario where the text backgrounds are derived from real-world images. To facilitate research of complex scenario IIMT, we design an IIMT dataset that includes subtitle text with real-world background. However previous IIMT models perform inadequately in complex scenarios. To address the issue, we propose the DebackX model, which separates the background and text-image from the source image, performs translation on text-image directly, and fuses the translated text-image with the background, to generate the target image. Experimental results show that our model achieves improvements in both translation quality and visual effect.

Figures

Figures reproduced from arXiv: 2505.15282 by the authors.

Figure 1
Figure 1. Illustration of simplified and complex scenario [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Architecture of our proposed DebackX. 4.1 Text-Image Background Separation The Text-Image Background Separation model de￾composes the source image into a background im￾age and a source text-image. The input source image x is encoded with two ViT (Dosovitskiy et al., 2021) encoders, Edeback and Edetext, obtain￾ing corresponding features separately. Then two ViT decoders Gback and Gtext take the features as [PITH_FUL… view at source ↗
Figure 3
Figure 3. Samples from the IIMT30k dataset with multiple fonts (Times New Roman, Arial, and Calibri). [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Detailed GPT-4o output of Case #1 [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Detailed GPT-4o output of Case #2 [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 5 canonical work pages

  1. [1]

    Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. 2022. https://arxiv.org/abs/2202.04200 Maskgit: Masked generative image transformer . Preprint, arXiv:2202.04200

    arXiv 2022

  2. [2]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. https://arxiv.org/abs/2010.11929 An image is worth 16x16 words: Transformers for image recognition at scale . Preprint, arXiv:2010.11929

    arXiv 2021

  3. [3]

    Desmond Elliott, Stella Frank, Khalil Sima'an, and Lucia Specia. 2016. https://doi.org/10.18653/v1/W16-3210 Multi30k: Multilingual english-german image descriptions . In Proceedings of the 5th Workshop on Vision and Language, pages 70--74. Association for Computational Linguistics

    doi:10.18653/v1/w16-3210 2016

  4. [4]

    Patrick Esser, Robin Rombach, and Bjorn Ommer. 2021. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873--12883

    2021

  5. [5]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2018. https://arxiv.org/abs/1706.08500 Gans trained by a two time-scale update rule converge to a local nash equilibrium . Preprint, arXiv:1706.08500

    arXiv 2018

  6. [6]

    Hirofumi Inaguma, Sravya Popuri, Ilia Kulikov, Peng-Jen Chen, Changhan Wang, Yu-An Chung, Yun Tang, Ann Lee, Shinji Watanabe, and Juan Pino. 2023. https://doi.org/10.18653/v1/2023.acl-long.872 U nit Y : Two-pass direct speech-to-speech translation with discrete units . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguist...

    doi:10.18653/v1/2023.acl-long.872 2023

  7. [7]

    Zhibin Lan, Liqiang Niu, Fandong Meng, Jie Zhou, Min Zhang, and Jinsong Su. 2024. https://arxiv.org/abs/2407.02894 Translatotron-v(ison): An end-to-end model for in-image machine translation . Preprint, arXiv:2407.02894

    arXiv 2024

  8. [8]

    Zhibin Lan, Jiawei Yu, Xiang Li, Wen Zhang, Jian Luan, Bin Wang, Degen Huang, and Jinsong Su. 2023. https://arxiv.org/abs/2305.17415 Exploring better text image translation with multimodal codebook . Preprint, arXiv:2305.17415

    arXiv 2023

Show all 30 references
  1. [9]

    Ilya Loshchilov and Frank Hutter. 2019. https://arxiv.org/abs/1711.05101 Decoupled weight decay regularization . Preprint, arXiv:1711.05101

    2019 arXiv

  2. [10]

    Mario Lucic, Karol Kurach, Marcin Michalski, Sylvain Gelly, and Olivier Bousquet. 2018. https://arxiv.org/abs/1711.10337 Are gans created equal? a large-scale study . Preprint, arXiv:1711.10337

    2018 arXiv

  3. [11]

    Cong Ma, Yaping Zhang, Mei Tu, Xu Han, Linghui Wu, Yang Zhao, and Yu Zhou. 2022. https://arxiv.org/abs/2210.03887 Improving end-to-end text image translation from the auxiliary text translation task . Preprint, arXiv:2210.03887

    2022 arXiv

  4. [12]

    Cong Ma, Yaping Zhang, Mei Tu, Yang Zhao, Yu Zhou, and Chengqing Zong. 2023 a . https://doi.org/10.18653/v1/2023.findings-emnlp.330 CCIM : Cross-modal cross-lingual interactive image translation . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages ...

    2023 doi

  5. [13]

    Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. 2023 b . https://arxiv.org/abs/2303.17870 Glyphdraw: Seamlessly rendering text with intricate spatial structures in text-to-image generation . Preprint, arXiv:2303.17870

    2023 arXiv

  6. [14]

    Elman Mansimov, Mitchell Stern, Mia Chen, Orhan Firat, Jakob Uszkoreit, and Puneet Jain. 2020. https://doi.org/10.18653/v1/2020.nlpbt-1.8 Towards end-to-end in-image neural machine translation . In Proceedings of the First International Workshop on Natural Language Processing ...

    2020 doi

  7. [15]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311--318

    2002

  8. [16]

    Wong, Xiaoshuai Sun, and Rongrong Ji

    Zhipeng Qian, Pei Zhang, Baosong Yang, Kai Fan, Yiwei Ma, Derek F. Wong, Xiaoshuai Sun, and Rongrong Ji. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.137 A ny T rans: Translate A ny T ext in the image with large scale models . In Findings of the Association for Comput...

    2024 doi

  9. [17]

    Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.213 COMET : A neural framework for MT evaluation . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2685--2702, ...

    2020 doi

  10. [18]

    Rodr \' guez, David Vazquez, Issam Laradji, Marco Pedersoli, and Pau Rodriguez

    Juan A. Rodr \' guez, David Vazquez, Issam Laradji, Marco Pedersoli, and Pau Rodriguez. 2023. Ocr-vqgan: Taming text-within-image generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3689--3698

    2023

  11. [19]

    Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. https://doi.org/10.18653/v1/P16-1162 Neural machine translation of rare words with subword units . In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages ...

    2016 doi

  12. [20]

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. 2024. https://arxiv.org/abs/2404.02905 Visual autoregressive modeling: Scalable image generation via next-scale prediction . Preprint, arXiv:2404.02905

    2024 arXiv

  13. [21]

    Yanzhi Tian, Xiang Li, Zeming Liu, Yuhang Guo, and Bin Wang. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.1004 In-image neural machine translation with segmented pixel sequence-to-sequence model . In Findings of the Association for Computational Linguistics: EMNLP 202...

    2023 doi

  14. [22]

    Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. 2024. https://arxiv.org/abs/2311.03054 Anytext: Multilingual visual text generation and editing . Preprint, arXiv:2311.03054

    2024 arXiv

  15. [23]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2023. https://arxiv.org/abs/1706.03762 Attention is all you need . Preprint, arXiv:1706.03762

    2023 arXiv

  16. [24]

    Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. 2022. https://openreview.net/forum?id=pfNyExj7z2 Vector-quantized image modeling with improved VQGAN . In International Conference on Learning Repre...

    2022

  17. [25]

    Lingjun Zhang, Xinyuan Chen, Yaohui Wang, Yue Lu, and Yu Qiao. 2023. https://arxiv.org/abs/2312.12232 Brush your text: Synthesize any scene text on images via diffusion model . Preprint, arXiv:2312.12232

    2023 arXiv

  18. [26]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. 2018. https://arxiv.org/abs/1801.03924 The unreasonable effectiveness of deep features as a perceptual metric . Preprint, arXiv:1801.03924

    2018 arXiv

  19. [27]

    Shaolin Zhu, Shangjie Li, Yikun Lei, and Deyi Xiong. 2023. https://doi.org/10.18653/v1/2023.acl-long.751 PEIT : Bridging the modality gap with pre-trained models for end-to-end image translation . In Proceedings of the 61st Annual Meeting of the Association for Computational L...

    2023 doi

  20. [28]

    Adrian Łańcucki, Jan Chorowski, Guillaume Sanchez, Ricard Marxer, Nanxin Chen, Hans J. G. A. Dolfing, Sameer Khurana, Tanel Alumäe, and Antoine Laurent. 2020. https://arxiv.org/abs/2005.08520 Robust training of vector quantized bottleneck models . Preprint, arXiv:2005.08520

    2020 arXiv

  21. [29]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  22. [30]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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