Pith. sign in

REVIEW 5 major objections 7 minor 44 references

OracleFusion: Assisting the Decipherment of Oracle Bone Script with Structurally Constrained Semantic Typography

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read OracleFusion claims that undeciphered oracle bone glyphs can be turned into semantically clear vector images whose structure stays faithful to the original, giving epigraphers visual clues they did not have before.

desk verdict Solid engineering application with a real method and dataset, but the decipherment claim is not supported by the evidence. read the letter →

arxiv 2506.21101 v1 pith:JNC4FOS5 submitted 2025-06-26 cs.CV

Caoshuo Li , Zengmao Ding , Xiaobin Hu , Bang Li , Donghao Luo , AndyPian Wu , Chaoyang Wang , Chengjie Wang
show 5 more authors
Taisong Jin SevenShu Yunsheng Wu Yongge Liu Rongrong Ji
This is my paper
classification cs.CV
keywords oraclebonescriptdeciphermentsemantictypographyvectorgraphicsgenerationscoredistillationsamplingmultimodallargelanguagemodelspatialawarenessreasoningglyphstructuralconstraintskeleton-basedmaintenance
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

OracleFusion tries to make the decipherment of oracle bone script more tractable by turning each glyph into a vector illustration that combines the original character's shape with a picture of what it means. The paper claims that a two-stage pipeline—first a multimodal large language model that names the glyph's components and where they sit, then a structurally constrained diffusion optimization that redraws the SVG—keeps the ancient form intact while making its semantics visible. If the claim holds, scholars gain a concrete aid for the roughly 2,900 undeciphered characters: instead of staring at abstract marks, they can inspect generated images that render plausible component-level scenes. The paper reports that this approach beats the existing baselines ClipDraw and Word-As-Image on semantic relevance, visual appeal, and glyph preservation, and that on unseen characters its outputs offer interpretation clues that resemble expert reasoning.

What carries the argument

The load-bearing mechanism is the pair of loss constraints inside SOVF. The Glyph Structural Constraint converts the MLLM's component layouts into binary masks applied to the cross-attention maps of a pretrained diffusion model, forcing each semantic token to attend inside its assigned region and suppressing attention outside it; this is optimized through the GSDS loss. Glyph maintenance adds the SKST loss, which compares angle-preserving vectors from skeleton points to contour points of the original glyph and the generated one, so strokes stay aligned while semantics are added. Around these sit OBSUG's Spatial Awareness Reasoning (SAR), which makes the MLLM output component positions as a directed acyclic graph, and the Oracle Glyph Vectorization (OGV) method, which skeletonizes glyph images to obtain the contour and skeleton points that SKST needs.

What would settle it

Take a random sample of the RMOBS glyph images and have two independent specialists in oracle bone script annotate the same components and bounding boxes without seeing the dataset's labels; if their agreement with the dataset is low on a meaningful fraction of components, the grounding results and the structural constraints are validated against an unreliable reference, and the claimed decipherment insights would not be expected to transfer to expert use.

Watch

Extended reading notes

Core claim

The paper's central claim is that oracle bone script can be decoded—not into modern Chinese characters directly, but into structurally faithful semantic pictograms—by combining a multimodal large language model (MLLM) that reasons about glyph composition with a vector-optimization stage that obeys the glyph's own layout. The first stage (OBSUG) identifies key components, expresses their relative positions as a directed graph, and grounds each component in a bounding-box layout. The second stage (SOVF) keeps those boxes as constraints on the cross-attention maps of a pretrained text-to-image diffusion model, uses a skeleton-derived loss to stop the redrawn strokes from drifting away from the original glyph, and optimizes the SVG with score distillation. The result is a black-and-white vector font that carries the semantic image inside the original strokes. On both deciphered and undeciphered characters, the paper argues, these renderings preserve the glyph's objective structure while suggesting concrete readings, and the framework outperforms ClipDraw and Word-As-Image in semantic relevance, visual appeal, and glyph maintenance.

Load-bearing premise

The method assumes that the hand-written labels of each character's component parts, their meanings, and their positions in the dataset are correct; these labels come from one web dictionary and were not checked by an independent oracle-script expert, so any systematic error in them would flow through the whole system.

Editorial extensions

If this is right

  • Undeciphered characters get a concrete visual hypothesis: for each unseen glyph the system outputs component locations, a semantic description, and a redrawn glyph, giving epigraphers a structured starting point.
  • The generated black-and-white SVGs can be fed into post-processing pipelines such as depth-to-image and scribble-to-image generation to add color and texture, so the same structural interpretation can be rendered in richer styles.
  • Semantic typography here is not limited to characters that already have vector fonts: OGV vectorizes arbitrary glyph images, so the pipeline applies to the wider corpus beyond the 4K-font library.
  • Because the layout constraints act on attention maps rather than requiring additional training, the generation stage is training-free and can be rerun on any new glyph once OBSUG has produced a layout.
  • The comparison results indicate that users judge OracleFusion higher on semantics, visual appeal, and glyph maintenance than the two baselines, making it the strongest current candidate among this class of decipherment aids.

Reading between the lines

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

  • Editorial extension: the layout-constrained score distillation plus skeleton-loss recipe is not specific to oracle bone script; the same two constraints should transfer to other pictographic or ideographic scripts, such as Egyptian hieroglyphs or cuneiform signs, where component inventories and approximate layouts are known.
  • Editorial extension: the paper's 'expert-like insights' are demonstrated only qualitatively on a handful of unseen characters; converting that into a quantitative claim would require a blind study in which independent epigraphers rate the generated interpretations against their own readings before and after seeing them.
  • Editorial extension: because the dataset labels come from a single web dictionary, the system currently demonstrates that one particular set of component-to-meaning assignments can be rendered faithfully, not that those assignments are historically correct; testing on characters with disputed readings would separate rendering fidelity from decipherment accuracy.
  • Editorial extension: a falsifiable use test would be to give epigraphers the generated typography for a set of unresolved glyphs and measure whether their candidate readings converge or change compared with a control condition without the images.
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

5 major / 7 minor

Summary. The paper presents OracleFusion, a two-stage system for generating semantically enriched vector fonts from oracle bone script (OBS) glyphs. In the first stage, a fine-tuned MLLM (Qwen-VL with LoRA) analyzes the glyph to identify key components, their spatial relations (via Spatial Awareness Reasoning, SAR), and a semantic description, and then grounds components in a bounding-box layout. In the second stage, the proposed Structural Oracle Vector Fusion (SOVF) optimizes SVG parameters using a combination of latent Score Distillation Sampling (SDS), a glyph structural constraint (GSDS loss) that shapes cross-attention maps to match the predicted layout, a skeleton-based structure loss (SKST loss), and a tone loss. The authors introduce a new dataset, RMOBS, with over 20K annotated samples across 900 characters. Experiments compare OracleFusion against ClipDraw and Word-As-Image on CLIPScore, a custom Distance metric, and a user study, reporting consistent improvements. The paper further claims that the method provides 'expert-like insights' that assist the decipherment of unseen OBS characters, illustrated mainly through qualitative examples in Section 5.2.

Significance. If the technical contributions hold, OracleFusion represents a novel and worthwhile application of semantic typography to a culturally significant domain. The paper makes several concrete contributions: a new annotated OBS dataset (RMOBS) with radical-level structure and semantics; a two-stage framework that combines MLLM-based structural analysis with layout-conditioned vector generation; and a skeleton-aware loss that helps preserve glyph identity during semantic morphing. The authors also provide a project page and qualitative results for undeciphered characters, which could be valuable for philologists. However, the central claim of assisting decipherment is currently supported only by anecdotal examples and has not been validated by domain experts. The reliability of the underlying annotations is also not independently verified. These are not fatal flaws in the generation pipeline itself, but they are load-bearing for the paper's stated significance and must be addressed before the claims can be accepted.

major comments (5)
  1. [§5.2, Abstract, Conclusion] The paper's headline claim is that OracleFusion 'provides expert-like insights' that 'assist experts in deciphering' OBS (Abstract; Conclusion). However, Section 5.2 only shows qualitative results for two undeciphered characters and two deciphered ones, with no evaluation by any OBS expert or linguist. The user study in §5.4 uses 70 participants 'familiar with the principles of Chinese character formation' but rates semantics, visual appeal, and glyph maintenance on deciphered characters, not whether the visualizations improve decipherment or are linguistically plausible. As the claimed utility beyond prior semantic typography is the decipherment assistance, please add a domain-expert evaluation (e.g., blinded philologists assessing whether the generated scenes for undeciphered characters are plausible or helpful) or substantially soften the claim to 'potential visual aid' without asserting expert-level insight.
  2. [§4.1 and Introduction] RMOBS annotations are described as 'manually annotated' from a single web dictionary (Vividict) with no independent expert validation, yet the Introduction states the dataset contains 'expert validation.' This is load-bearing because the MLLM is fine-tuned on these annotations and the layout constraints in SOVF are derived directly from them; systematic annotation errors would propagate through the whole pipeline. Please clarify who performed the annotation, whether any domain experts were involved, and report any inter-annotator agreement or independent review process. Without this, the reliability of the structural and semantic ground truth is unverified.
  3. [§4.5, Eq. (11)] The GSDS gradient formula is not a valid chain rule for the defined loss. Eq. (11) writes ∇P L_GSDS = E[ L_GS · ∂z_t/∂z · ∂z/∂x_aug · ∂x_aug/∂P ], where L_GS is a scalar loss (L_IR + L_OR). For a scalar loss, the gradient should involve the derivative of L_GS with respect to z_t, not the loss value itself multiplied by the Jacobians; the expression is dimensionally inconsistent and does not follow from the definition of L_GS as a function of cross-attention maps. Please provide the correct derivation (e.g., via ∂L_GS/∂A_t · ∂A_t/∂z_t) or clarify the notation if the intent is a weighted update based on L_GS rather than a true gradient.
  4. [§5.4, Table 1] The user study is reported without error bars or significance tests: 70 self-selected participants rate 28 characters on a 1–5 scale, and the table shows only means. Given the small margins (e.g., CLIPScore 28.30 vs 27.78), it is unclear whether the improvements are statistically reliable. Additionally, the automatic metrics are partially aligned with the training objectives: CLIPScore uses the same CLIP/Stable Diffusion prior family as the SDS loss, and the custom 'Distance' metric is not described in the paper. Please report variance and significance tests, and provide a precise definition of Distance (including the matching algorithm and hyperparameters) so that the glyph-maintenance claim is reproducible.
  5. [§5.2, Figure 7] The decipherment capability is evaluated only through a few hand-picked examples. There is no quantitative comparison to OBSD [9] or other decipherment-oriented baselines, and no metric that measures whether a generated scene corresponds to the historical meaning of an undeciphered character. If 'assisting decipherment' is a central contribution, the evaluation should include a systematic protocol (e.g., a set of undeciphered characters with known eventual decipherments, assessed by experts, or at least a larger corpus with machine-readable ground truth). As it stands, the claim is supported by anecdotal figures.
minor comments (7)
  1. [Throughout, e.g., Abstract and §4.5] The acronym for the method is inconsistent: the Abstract and contributions use 'OSVF' (Oracle Structural Vector Fusion) while the method section uses 'SOVF' (Structural Oracle Vector Fusion). Please unify.
  2. [§4.5, Eq. (14)] Eq. (14) includes a 'learnable weight parameter' w for the GSDS term, but no learning rule or initial value is given. In the ablation (Figure 8) the mechanism is shown as 'w/o GSDS' but the effect of varying w is not studied. Please specify how w is learned or fixed.
  3. [§4.5 and Contributions] The method is called 'training-free' in §4.5 but described as 'test-time training' in the contribution list. This wording is confusing; please clarify whether the optimization at inference is considered training.
  4. [§4.3] The OGV algorithm has hyperparameters k (normal smoothing window) and stroke width w that are not specified. Please report the values used in the experiments.
  5. [Figure 6 caption] The caption says results are ordered Word-As-Image, ClipDraw, and OracleFusion from left to right, but the figure order is ClipDraw, Word-As-Image, OracleFusion. Please fix the mismatch.
  6. [Figure 9] The reference '[12]' in the figure label 'w/o [12]' is unclear; it seems to refer to the ACAP loss from [16] but the citation is incorrect. Please clarify.
  7. [§4.1] Please provide statistics on the number of unique key components/radicals in RMOBS, and clarify whether the 'over 20K samples' correspond to unique glyph images or augmented instances.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the generation pipeline is trained and evaluated against external objectives and baselines, and the self-citations are not load-bearing.

full rationale

OracleFusion's derivation chain is self-contained relative to its inputs. The vector generation is optimized with SDS loss from DreamFusion/VectorFusion, a new GSDS layout constraint built on cross-attention maps, and a new SKST skeleton loss; none of these objectives is defined in terms of the evaluation metrics or vice versa. Semantic relevance is measured with CLIPScore, which is not the optimization objective (SDS uses Stable Diffusion noise prediction), and the comparison is made against external baselines ClipDraw and Word-As-Image plus a 70-participant user study. Glyph maintenance is evaluated with an author-designed "Distance" metric; although this metric could in principle coincide with the SKST training loss, the paper gives no equation for Distance, so the specific reduction required by the circularity standard cannot be exhibited. The RMOBS annotations are manually produced by the authors from the Vividict web dictionary and fine-tune the MLLM, which is a legitimate data-provenance and external-validity concern for the decipherment claim, but it is not a circular step: the MLLM is tested on unseen characters, and the generation quality is judged by independent raters and external metrics. Self-citations appear only in related-work and implementation contexts and are not load-bearing for the central claim. The paper's assertion that OracleFusion provides "expert-like insights" for decipherment is not validated by any expert study, but absence of evidence is a correctness risk, not circularity. Therefore no step in the claimed derivation reduces by construction to its own inputs.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on domain assumptions about annotation correctness, cross-attention semantics, structural preservation, and the utility of illustrations for decipherment. The system depends on several manually chosen or unreported hyperparameters (beta, top-k, OGV stroke width, smoothing window), so the contribution is not fully characterized. No new physical or conceptual entities are postulated; the modules (OBSUG, SOVF, OGV, SAR) and the dataset (RMOBS) are software artifacts, not invented entities of the kind this ledger tracks.

free parameters (6)
  • beta (SKST loss weight) = 0.5
    Set by visual inspection in Fig. 9 to balance semantic expression and glyph preservation; not tuned on a validation set.
  • w (GSDS loss weight) = learnable (value not reported)
    Eq. 14 defines w as learnable, but the final value or training schedule is not given.
  • gamma_t (tone loss weight) = step-dependent, inherited from [16]
    Taken from Word-As-Image without re-tuning; its exact schedule is not restated.
  • P (number of top-k attention responses) = not reported
    Used in Eq. 8-9 to constrain cross-attention; the value of P is never specified.
  • k (normal smoothing window size in OGV) = not reported
    Sliding window size in Section 4.3, chosen without stated criterion.
  • Stroke width w in OGV = not reported
    Used to offset contour points in Section 4.3; its value is not given.
assumptions (4)
  • domain assumption RMOBS annotations (component identities, semantic concepts, bounding boxes) are accurate ground truth for OBS.
    Section 4.1: data is collected from Vividict and manually annotated by the authors; no detailed expert validation protocol is reported.
  • domain assumption Cross-attention maps in Stable Diffusion indicate where objects are generated in the image.
    Section 4.5, Eq. 7-10: the GSDS loss assumes high attention inside the box and low attention outside corresponds to correct object placement.
  • domain assumption Skeleton-based angular consistency (SKST loss) preserves the perceived glyph structure.
    Section 4.5, Eq. 12-13: no evidence beyond the paper's own ablations that angle preservation is the right structural invariant.
  • domain assumption Semantically illustrated vector fonts assist human experts in deciphering OBS.
    The application claim in Section 5.2 and 6 is not tested with expert decipherment tasks or measured improvement in decipherment accuracy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OracleFusion: Assisting the Decipherment of Oracle Bone Script with Structurally Constrained Semantic Typography." pith.science (2026). https://pith.science/paper/JNC4FOS5

@misc{pith2026250621101,
  author       = {Pith},
  title        = {Pith review of: OracleFusion: Assisting the Decipherment of Oracle Bone Script with Structurally Constrained Semantic Typography},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JNC4FOS5}},
  note         = {Machine review of arXiv:2506.21101}
}
read the original abstract

As one of the earliest ancient languages, Oracle Bone Script (OBS) encapsulates the cultural records and intellectual expressions of ancient civilizations. Despite the discovery of approximately 4,500 OBS characters, only about 1,600 have been deciphered. The remaining undeciphered ones, with their complex structure and abstract imagery, pose significant challenges for interpretation. To address these challenges, this paper proposes a novel two-stage semantic typography framework, named OracleFusion. In the first stage, this approach leverages the Multimodal Large Language Model (MLLM) with enhanced Spatial Awareness Reasoning (SAR) to analyze the glyph structure of the OBS character and perform visual localization of key components. In the second stage, we introduce Oracle Structural Vector Fusion (OSVF), incorporating glyph structure constraints and glyph maintenance constraints to ensure the accurate generation of semantically enriched vector fonts. This approach preserves the objective integrity of the glyph structure, offering visually enhanced representations that assist experts in deciphering OBS. Extensive qualitative and quantitative experiments demonstrate that OracleFusion outperforms state-of-the-art baseline models in terms of semantics, visual appeal, and glyph maintenance, significantly enhancing both readability and aesthetic quality. Furthermore, OracleFusion provides expert-like insights on unseen oracle characters, making it a valuable tool for advancing the decipherment of OBS.

Figures

Figures reproduced from arXiv: 2506.21101 by the authors.

Figure 1
Figure 1. Illustration of OBS decipherment process. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Visualization of RMOBS annotations, including oracle [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Delaunay triangulation of the initial and resulting points [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Overview of the proposed OracleFusion framework. This framework first employs OGV to transform oracle images into [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Cross-attentions between target tokens e.g., bird, moun￾tain) and latent features of the UNet in Stable Diffusion Model. applies targeted constraints on At . This constraint guides the gradient of latent features zt and propagates through the SVG parameters via backpro…
Figure 6
Figure 6. Figure 6: Qualitative comparison of our OracleFusion with other popular methods. On the left are the initial glyph as input. The results [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Results of unseen undeciphered and deciphered OBS. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 9
Figure 9. Figure 9: Effect of the SKST loss and comparison of [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Examples of utilizing D2I in SD2 [31] and Scribbles￾to-Image with ControlNet [41] as a post-processing step. 6. Conclusion In this paper, we propose OracleFusion, a two-stage frame￾work for generating semantically rich vectorized fonts to support in the interpretation…
Figure 11
Figure 11. Figure 11: Examples of utilizing Depth-to-Image in Stable Diffusion 2 [ [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: Examples of utilizing Scribbles-to-Image with ControlNet-SDXL 1.0 [ [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: More grounding and semantic typography results for undeciphered oracle bone scripts. Note that the different colors of [PITH_FULL_IMAGE:figures/full_fig_p014_13.png]
Figure 14
Figure 14. Figure 14: The description of our user study [PITH_FULL_IMAGE:figures/full_fig_p015_14.png]
Figure 15
Figure 15. Figure 15: The description of our user study [PITH_FULL_IMAGE:figures/full_fig_p015_15.png]
Figure 16
Figure 16. Figure 16: More qualitative comparison results between ClipDraw, Word-As-Image, and our OracleFusion in our user study. [PITH_FULL_IMAGE:figures/full_fig_p016_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 42 canonical work pages

  1. [9]

    De- ciphering oracle bone language with diffusion models

    Haisu Guan, Huanxin Yang, Xinyu Wang, Shengwei Han, Yongge Liu, Lianwen Jin, Xiang Bai, and Yuliang Liu. De- ciphering oracle bone language with diffusion models. In Proceedings of the 62nd Annual Meeting of the Associa- tion for Computational Linguistics (Volume 1: Long Papers), pages 15554–15567, Bangkok, Thailand, 2024. Association for Computational Li...

    work page 2024

  2. [1]

    Freetype.https://freetype.org/. 2009. 4

    work page 2009

  3. [2]

    Qwen-vl: A versatile vision-language model for un- derstanding, localization, text reading, and beyond.arXiv preprint arXiv:2308.12966, 2023

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for un- derstanding, localization, text reading, and beyond.arXiv preprint arXiv:2308.12966, 2023. 4, 6

    arXiv 2023

  4. [3]

    Attend-and-excite: Attention-based se- mantic guidance for text-to-image diffusion models, 2023

    Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based se- mantic guidance for text-to-image diffusion models, 2023. 5

    work page 2023

  5. [4]

    Sur la sphere vide.Izv

    Boris Delaunay et al. Sur la sphere vide.Izv. Akad. Nauk SSSR, Otdelenie Matematicheskii i Estestvennyka Nauk, 7 (793-800):1–2, 1934. 6

    work page 1934

  6. [5]

    The pascal visual object classes (voc) challenge.International journal of computer vision, 88:303–338, 2010

    Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge.International journal of computer vision, 88:303–338, 2010. 6

    2010

  7. [6]

    Layoutgpt: Compositional visual plan- ning and generation with large language models

    Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Ar- jun Akula, Xuehai He, Sugato Basu, XinEric Wang, and WilliamYang Wang. Layoutgpt: Compositional visual plan- ning and generation with large language models. 2023. 3

    work page 2023

  8. [7]

    Soros, and Olaf Witkowski

    Kevin Frans, L.B. Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders.Cornell University - arXiv,Cornell University - arXiv, 2021. 7, 1

    work page 2021

Show all 44 references
  1. [8]

    Image translation for oracle bone character interpretation

    Feng Gao, Jingping Zhang, Yong ge Liu, and Yahong Han. Image translation for oracle bone character interpretation. Symmetry, 14:743, 2022. 2

    2022

  2. [10]

    CLIPScore: a reference-free evaluation met- ric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: a reference-free evaluation met- ric for image captioning. InEMNLP, 2021. 6, 8

    2021

  3. [11]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. InIn- ternational Conference on Learning Representations, 2022. 6

    2022

  4. [12]

    Face super-resolution guided by 3d facial priors

    Xiaobin Hu, Wenqi Ren, John LaMaster, Xiaochun Cao, Xi- aoming Li, Zechao Li, Bjoern Menze, and Wei Liu. Face super-resolution guided by 3d facial priors. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IV 16, pages 763...

    2020

  5. [13]

    Autogan-synthesizer: neural architecture search for cross-modality mri synthesis

    Xiaobin Hu, Ruolin Shen, Donghao Luo, Ying Tai, Chengjie Wang, and Bjoern H Menze. Autogan-synthesizer: neural architecture search for cross-modality mri synthesis. InIn- ternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 397–409. Springer,

  6. [14]

    High-resolution it- erative feedback network for camouflaged object detection

    Xiaobin Hu, Shuo Wang, Xuebin Qin, Hang Dai, Wenqi Ren, Donghao Luo, Ying Tai, and Ling Shao. High-resolution it- erative feedback network for camouflaged object detection. InProceedings of the AAAI Conference on Artificial Intelli- gence, pages 881–889, 2023. 3

    2023

  7. [15]

    Diffumatting: Synthe- sizing arbitrary objects with matting-level annotation

    Xiaobin Hu, Xu Peng, Donghao Luo, Xiaozhong Ji, Jin- long Peng, Zhengkai Jiang, Jiangning Zhang, Taisong Jin, Chengjie Wang, and Rongrong Ji. Diffumatting: Synthe- sizing arbitrary objects with matting-level annotation. In European Conference on Computer Vision, pages 396–413....

    2024

  8. [16]

    Word-as-image for semantic typography.ACM Trans

    Shir Iluz, Yael Vinker, Amir Hertz, Daniel Berio, Daniel Cohen-Or, and Ariel Shamir. Word-as-image for semantic typography.ACM Trans. Graph., 42(4), 2023. 3, 4, 6, 7, 8, 1

    2023

  9. [17]

    Vectorfusion: Text-to-svg by abstracting pixel-based diffusion models

    Ajay Jain, Amber Xie, and Pieter Abbeel. Vectorfusion: Text-to-svg by abstracting pixel-based diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 1911–1920, 2023. 3, 4

    1911

  10. [18]

    Sonic: Shifting focus to global au- dio perception in portrait animation

    Xiaozhong Ji, Xiaobin Hu, Zhihong Xu, Junwei Zhu, Chum- ing Lin, Qingdong He, Jiangning Zhang, Donghao Luo, Yi Chen, Qin Lin, et al. Sonic: Shifting focus to global au- dio perception in portrait animation. InProceedings of the Computer Vision and Pattern Recognition Conferenc...

    2025

  11. [19]

    Oraclepoints: A hybrid neural rep- resentation for oracle character

    Runhua Jiang, Yongge Liu, Boyuan Zhang, Xu Chen, Deng Li, and Yahong Han. Oraclepoints: A hybrid neural rep- resentation for oracle character. InProceedings of the 31st ACM International Conference on Multimedia, page 7901–7911, New York, NY , USA, 2023. Association for Comput...

    2023

  12. [20]

    Wein- berger, Jitendra Malik, Trevor Darrell, and Dan Klein

    Boyi Li, Rodolfo Corona, Karttikeya Mangalam, Cather- ine Chen, Daniel Flaherty, Serge Belongie, KilianQ. Wein- berger, Jitendra Malik, Trevor Darrell, and Dan Klein. Does unsupervised grammar induction need pixels? 2022. 3

    2022

  13. [21]

    Mix-up augmentation for oracle character recognition with imbalanced data distribution

    Jing Li, Qiu-Feng Wang, Rui Zhang, and Kaizhu Huang. Mix-up augmentation for oracle character recognition with imbalanced data distribution. InDocument Analysis and Recognition – ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–10, 2021, Proceeding...

    2021

  14. [22]

    Differentiable vector graphics rasterization for editing and learning.ACM Transactions on Graphics, page 1–15, 2020

    Tzu-Mao Li, Michal Luk ´aˇc, Micha ¨el Gharbi, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for editing and learning.ACM Transactions on Graphics, page 1–15, 2020. 4

    2020

  15. [23]

    Gligen: Open-set grounded text-to-image generation

    Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jian- wei Yang, Jianfeng Gao, Chunyuan Li, and YongJae Lee. Gligen: Open-set grounded text-to-image generation. 2023. 3

    2023

  16. [24]

    Llm-grounded diffusion: Enhanc- ing prompt understanding of text-to-image diffusion models with large language models

    Long Lian, Boyi Li, Adam Yala, Trevor Darrell, UC Berke- ley, and Ucsf Ucsf. Llm-grounded diffusion: Enhanc- ing prompt understanding of text-to-image diffusion models with large language models. 3

  17. [25]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. 3

  18. [26]

    Bleu: a method for automatic evaluation of machine translation

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. InProceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311– 318, Philadelphia, Pennsylvania, USA, 2002. Associ...

    2002

  19. [27]

    Dreamfusion: Text-to- 3d using 2d diffusion

    Ben Poole, Ajay Jain, JonathanT Barron, Ben Mildenhall, Google Research, and UC Berkeley. Dreamfusion: Text-to- 3d using 2d diffusion. . 4, 6

  20. [28]

    Dreamfusion: Text-to- 3d using 2d diffusion

    Ben Poole, Ajay Jain, JonathanT Barron, Ben Mildenhall, Google Research, and UC Berkeley. Dreamfusion: Text-to- 3d using 2d diffusion. . 3

  21. [29]

    Making visual sense of oracle bones for you and me.CVPR,

    Runqi Qiao, Lan Yang, Kaiyue Pang, and Honggang Zhang. Making visual sense of oracle bones for you and me.CVPR,

  22. [30]

    Hierarchical text-conditional image gener- ation with clip latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener- ation with clip latents. 3

  23. [31]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 4, 5, 8, 1, 2

    2022

  24. [32]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox.U- Net: Convolutional Networks for Biomedical Image Segmen- tation, page 234–241. 2015. 4

    2015

  25. [33]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar, Seyed Ghasemipour, Burcu Karagol, SSara Mahdavi, RaphaGon- tijo Lopes, Tim Salimans, Jonathan Ho, DavidJ Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models wi...

  26. [34]

    Align and surpass human camouflaged perception: Visual refocus re- inforcement fine-tuning.arXiv preprint arXiv:2505.19611,

    Ruolin Shen, Xiaozhong Ji, Kai Wu, Jiangning Zhang, Yijun He, HaiHua Yang, Xiaobin Hu, and Xiaoyu Sun. Align and surpass human camouflaged perception: Visual refocus re- inforcement fine-tuning.arXiv preprint arXiv:2505.19611,

    arXiv

  27. [35]

    Ds-fusion: Artistic typography via discriminated and stylized diffusion

    Maham Tanveer, Yizhi Wang, Ali Mahdavi-Amiri, and Hao Zhang. Ds-fusion: Artistic typography via discriminated and stylized diffusion. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV), 2023. 3

    2023

  28. [36]

    Chinese pictograph dictionary.https://www

    Vividict. Chinese pictograph dictionary.https://www. vividict.com/. Accessed: November, 2024. 3

    2024

  29. [37]

    Unsuper- vised structure-texture separation network for oracle charac- ter recognition.IEEE Transactions on Image Processing, 31: 3137–3150, 2022

    Mei Wang, Weihong Deng, and Cheng-Lin Liu. Unsuper- vised structure-texture separation network for oracle charac- ter recognition.IEEE Transactions on Image Processing, 31: 3137–3150, 2022. 2

    2022

  30. [38]

    Visorgpt: Learning visual prior via generative pre- training

    Jinheng Xie, Kai Ye, Yudong Li, Yuexiang Li, Kevin- Qinghong Lin, Yefeng Zheng, Linlin Shen, and MikeZheng Shou. Visorgpt: Learning visual prior via generative pre- training. 3

  31. [39]

    Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion

    Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wen- tian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7452–7461,

  32. [40]

    Diffsketcher: Text guided vector sketch synthesis through latent diffusion models

    XiMing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. InThirty- seventh Conference on Neural Information Processing Sys- tems, 2023. 3

    2023

  33. [41]

    Adding conditional control to text-to-image diffusion models, 2023

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 8, 1, 3

    2023

  34. [42]

    Controllable text-to-image generation with gpt-4

    Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, Xin Wang, UC Berkeley, and Microsoft Research. Controllable text-to-image generation with gpt-4. 3

  35. [43]

    T. Y . Zhang and C. Y . Suen. A fast parallel algorithm for thinning digital patterns.Commun. ACM, 27(3):236–239,

  36. [44]

    Swift:a scal- able lightweight infrastructure for fine-tuning, 2024

    Yuze Zhao, Jintao Huang, Jinghan Hu, Xingjun Wang, Yun- lin Mao, Daoze Zhang, Zeyinzi Jiang, Zhikai Wu, Baole Ai, Ang Wang, Wenmeng Zhou, and Yingda Chen. Swift:a scal- able lightweight infrastructure for fine-tuning, 2024. 6 OracleFusion: Assisting the Decipherment of Oracle ...

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.