OrienText: Surface Oriented Textual Image Generation

Arushi Jain; Lovekesh Vig; Monika Sharma; Shubham Singh Paliwal; Vikram Jamwal

REVIEW 3 major objections 4 minor 5 references

OrienText: Surface Oriented Textual Image Generation

T0 review · 3 major / 4 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 The paper claims that feeding region-specific surface normals to a ControlNet-augmented text-to-image diffusion model, together with a surface-projected character mask, makes generated text correctly follow the orientation of angled…

desk verdict Promising idea, broken write-up: the mask alignment equations don't actually orient text, and the metric is self-contradictory. read the letter →

arxiv 2505.20958 v1 pith:NPWY7U2F submitted 2025-05-27 cs.CV

Shubham Singh Paliwal , Arushi Jain , Monika Sharma , Vikram Jamwal , Lovekesh Vig This is my paper

classification cs.CV

keywords text-to-imagegenerationdiffusionmodelssurfacenormalestimationtextrenderingperspectivealignmenteditingControlNetconditioningscene

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

The paper proposes OrienText, a method that makes text-to-image diffusion models render text on angled surfaces (billboards, building walls, banners, packaging) so that the letters follow the orientation of the underlying plane rather than appearing as flat frontal overlays. The central claim is that feeding region-specific surface normals into a ControlNet-augmented text-generation diffusion model, together with a character mask whose bounding boxes are projected onto the surface plane, is enough to produce correctly oriented and harmonized text. If true, this removes a manual perspective-warping step from text editing pipelines and improves the realism of AI-generated signage and product imagery. The paper supports the claim with a surface-normal consistency metric and a human study, reporting lower mean angular error than TextDiffuser, TextDiffuser-2, and AnyText, and higher human ratings on harmonization, text rendering, and perspective blending.

What carries the argument

The central object is the surface-aligned character mask $C^a_{\text{mask}}$, produced by projecting the original character mask onto the plane perpendicular to the estimated surface normal $\mathbf{N}=(n_x,n_y,n_z)$. Surface normals are estimated per-pixel from the single input image by learning the rotation between neighboring pixels; then, for each character bounding box, the center $\mathbf{C}$ is translated by a unit depth $d=1$ along the normal and projected onto the plane through the origin, and the four corners are placed at $\pm w/2$ and $\pm h/2$ around the projected center. This transformed mask tells the diffusion model where and how to paint characters. The second mechanism is the ControlNet-augmented text-to-image model, which takes the surface normal map $\mathbf{N}$ as an additional control input alongside the source image, the ROI mask, and $C^a_{\text{mask}}$, conditioning the base text-generation model on the surface geometry.

What would settle it

Render a checkerboard texture on a plane at a known grazing angle, estimate its surface normal, apply the Eqs. 1-5 projection to a text bounding box, and check whether the projected corners coincide with the perspective-mapped ground-truth corners; if the projected box stays axis-aligned at large slant angles, the projection step is not performing the orientation work attributed to it.

Watch

Extended reading notes

Core claim

The paper's core claim is that region-specific surface normals are a sufficient conditioning signal for a text-to-image diffusion model to orient generated text correctly on non-frontal surfaces. Given a source image, a region of interest, and a character mask, OrienText estimates the surface normal map of the masked region, projects each character bounding box onto the plane perpendicular to the normal (assuming unit depth because no depth map is available), and feeds both the normal map and the projected mask into a ControlNet-augmented version of a base text-generation diffusion model. The projection equations (1)-(5) translate each box's center onto the plane through the origin with normal $\mathbf{N}$ and redraw the box as an axis-aligned rectangle of the same width and height around the projected center. The authors report that this mask alignment lowers the surface-normal consistency error for every method they compare, with OrienText achieving the lowest mean angular error (MAE-N $=1.7411$) and the highest human ratings on harmonization, text rendering, and perspective blending.

Load-bearing premise

The load-bearing premise is that the estimated surface normal, combined with the unit-depth plane projection in Eqs. 1-5, faithfully describes how a text rectangle should lie on the photographed surface; if the normal estimate is off or the projection is too crude to rotate characters, the aligned mask will mislead the diffusion model rather than help it.

Editorial extensions

If this is right

Text painted onto angled surfaces no longer needs a manual perspective-warping step: the aligned-mask pipeline handles billboards, walls, banners, and packaging in one pass.
The proposed MAE-N metric gives an automatic, surface-normal-based way to score orientation quality, which can be applied to any text generation method.
Mask alignment improves surface-normal consistency for all compared baselines, not just OrienText, so the projection step is a transferable improvement.
OrienText works as a post-processing stage for diffusion-generated content, producing product-advertisement images with multi-line text.
Very small fonts and 3D engraved text remain failure cases, as the authors state in the conclusion.

Reading between the lines

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

Because the projection assumes a single unit-depth plane, strongly curved surfaces (soda cans, car bodies) will likely need a depth-aware or per-pixel projection before the mask alignment generalizes.
The current axis-aligned redraw of boxes around projected centers does not rotate or shear glyphs; feeding per-character quad corner coordinates into the mask would be a direct test of how much perspective cue the diffusion model actually needs.
The MAE-N metric is only as trustworthy as the underlying surface normal estimator, so orientation scores on images with weak normal estimates should be read cautiously.
A natural extension is to condition on region-specific normal maps rather than one global plane, which would let a single image contain text on multiple differently angled surfaces.

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.

Desk Editor's Note

read the letter

Bottom line: OrienText names a real gap in text-to-image editing — angled surfaces — and the basic idea of feeding surface normals into a ControlNet is a reasonable thing to try. That part is new relative to the cited baselines, and it's the paper's genuine contribution. The human evaluation, while tiny, at least suggests the method produces something visually preferable. So I don't read this as a purely incremental or lazy submission.

But the written core does not hold. Section 3.2's projection is internally inconsistent: Eq. 5 uses the original center C where the projection of C' requires n·C'/||n||². For a simple front-facing plane, their equations return (c_x, c_y, -1) instead of (c_x, c_y, 0). And even if that were fixed, the corners are derived by translating C_p by ±w/2 and ±h/2 along image x/y — no rotation, shear, or perspective warp. So the mask alignment, as described, cannot be the mechanism that orients text on an oblique banner. The stress-test note confirmed this on a second read.

The quantitative evidence is in the same condition. MAE-N is defined as mean cosine similarity while the text says a lower value means better preservation; that only makes sense if the code silently computes 1−cos or arccos. Table 1's AnyText row gets worse with mask alignment (1.8937 → 1.8955), contradicting the claim of universal improvement. The evaluation set is 60 self-curated images and no code, data, or training configuration is released. That leaves the central claim essentially unsupported.

What's here that's salvageable? The idea itself, plus a clean experimental setup in principle. But the paper needs a corrected geometric derivation (or an honest statement that the warp is a learned ControlNet transformation), a coherent metric definition, and a proper release before it can be fairly evaluated. As it stands, I would not accept it for publication, and I'd advise the authors to fix these issues before resubmitting anywhere.

For a reading group, it's a decent example of why internal consistency checks and metric sanity checks matter. Not something I'd cite in its current form.

Referee Report

3 major / 4 minor

Summary. The paper introduces OrienText, a method for text-to-image generation on perspectively angled surfaces. The authors propose conditioning a ControlNet-augmented diffusion model on surface normals estimated by the method of Bae and Davison (2024), and transforming the input character mask via a projection onto a surface-normal-defined plane. They evaluate on a self-curated 60-image dataset with a proposed surface-normal consistency metric (MAE-N) and a human study, reporting improvements over TextDiffuser, TextDiffuser-2, and AnyText.

Significance. The problem is real and timely: generating or editing text on surfaces such as billboards, packaging, and building facades requires perspective-aware rendering, and current text diffusion models often overlay frontal text. The core idea of conditioning on surface normals is reasonable, and the paper identifies a useful evaluation axis (perspective blending) that is typically neglected. However, the paper's central geometric construction is flawed, and the quantitative metric is confounded with the conditioning input. If the projection and metric were corrected, the approach might still be of interest, but as written the evidence does not support the claimed contribution. No code or data are released, limiting reproducibility.

major comments (3)

[Sec. 3.2, Eqs. (1)-(5)] The character-mask alignment procedure is internally inconsistent and does not perform the claimed perspective alignment. First, the corner points of the bounding box are computed by translating the projected center C_p by ±w/2 in x and ±h/2 in y, which produces axis-aligned rectangles in image coordinates; no rotation, shear, or foreshortening is applied, so the transformed mask C^a_mask cannot encode the orientation of a slanted surface. Second, the equations are inconsistent: with n=(0,0,1) and c_z=0, Eq. (3) gives C'=(c_x,c_y,-1), and Eq. (5) with t = n_x c_x + n_y c_y + n_z c_z over ||n||^2 gives t=0, so Eq. (4) yields C_p=(c_x,c_y,-1), not the orthogonal projection (c_x,c_y,0) onto the plane through the origin. The text in Section 3.2 says the projection is onto the plane defined by n and the origin, but the formula uses the original C rather than C' in the numerator, leading to the inconsistency. Since the paper's central claim is that OrienText orients text correctly via this mask alignment, this is a load-bearing error that invalidates the described mechanism.
[Sec. 4.2.1 and Table 1] The MAE-N metric is circular: it computes surface normals with the same estimator [Bae and Davison 2024] that is used as conditioning input to the model, and then compares normals before and after generation. Thus the metric largely measures whether the model preserved its own conditioning signal, which inherently favors normal-conditioned methods such as OrienText and does not provide an independent measure of text-surface alignment. Additionally, the definition 'MAE-Normal = mean(cosine_similarity(N, N'))' is not a mean angular error; cosine similarity is a similarity where higher values indicate better agreement, but the paper claims 'A lower MAE-N indicates that the surface normals remain similar.' The numerical values in Table 1 (e.g., 4.5243, 1.8937) are outside the range of cosine similarity, so either the equation is wrong or the table reports a different quantity. The metric as presented cannot support the quantitative conclusions.
[Table 1, 'Anytext' row] The text in Section 4.3 states that 'aligning character masks with surface normals improves performance across all methods,' but the AnyText row in Table 1 shows a worsening from 1.8937 (w/o mask alignment) to 1.8955 (w/ mask alignment). This direct contradiction undermines the universal-improvement claim, and the paper does not discuss or explain this exception. No confidence intervals or significance tests are reported, so even the improvements that are in the expected direction are not established statistically.

minor comments (4)

[Sec. 3.2] The symbol N is used both for the surface normal vector (Eq. 1) and for the full surface-normal map (Fig. 1), which is confusing; a distinct symbol such as N_map would clarify the presentation.
[Sec. 4.1] The training dataset (2,320 images) is derived from SCUT with affine augmentations, but the paper does not specify whether the surface-normal maps are computed on the original or augmented images, nor does it report the train/validation split or any hyperparameter sensitivity.
[References] TextDiffuser appears as both [Chen et al. 2024] and [Chen et al. 2023] with nearly identical titles; this duplication should be resolved, and the paper should clarify which version is used for comparison and as the base model.
[Supplementary, Fig. 2] The supplementary example explicitly shows failures for 3D and small text, which is consistent with the limitations mentioned in the conclusion, but the main paper does not discuss these failure cases in the quantitative evaluation.

Circularity Check

1 steps flagged · score 4.0 of 10

The quantitative MAE-N metric uses the same surface-normal estimator (Bae & Davison 2024) that supplies the model's conditioning input, making the metric partly a check that the model preserved its own input; the human study is small, so the central orientation claim is not independently supported by the reported numbers.

self definitional [Sec. 3.1 and Sec. 4.2.1 (Automated Evaluation; MAE-Normal definition)]
"We first compute the surface normals [Bae and Davison 2024] of the image before and after text generation using our proposed OrienText method. ... We then calculate the Mean Angular Error between the surface normals of the input and generated images, denoted as: MAE-Normal = mean( cosine_similarity(N, N’)), where N∈ R^{h×w×3} and N′∈ R^{h×w×3} represent the surface normals before and after the image generation, respectively. A lower MAE-N indicates that the surface normals remain similar, implying that the text characters are well-aligned with the underlying surface geometry."

The conditioning input is the surface-normal map N from Bae & Davison (Sec. 3.1). The metric defines good alignment as low difference between input and output normals, both computed by that same estimator (Sec. 4.2.1). So MAE-N largely measures whether the output preserved the conditioning map it was given. A model that leaves the background unchanged and does not perturb the estimator would score well regardless of actual character orientation. Since the model is trained with normal maps from this estimator (Sec. 4.1), the reported improvement partly verifies that the model kept its own input, not that text is correctly oriented against an external ground truth. The human study is independent but small (15 raters, 60 images).

full rationale

The central geometric projection in Sec. 3.2 is not circular, though it is a correctness risk: the corners are computed by translating the projected center by ±w/2 and ±h/2, leaving axis-aligned boxes, and Eq. 5 appears inconsistent with the projection described. Those issues are not self-referential reductions and do not raise the circularity score. The one genuine circularity is the MAE-N metric: the same surface-normal estimator both conditions the model and scores the result, so the metric is partly a check of input preservation. The human evaluation provides some independent evidence, which keeps the score moderate. No load-bearing self-citations or uniqueness imports are present; the self-citation to the authors' CustomText work is only a related-work enumeration.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method's core alignment is built on several arbitrary geometric choices, and the evaluation metric is entangled with the conditioning signal.

free parameters (3)

Unit depth d = 1
Assumed in Section 3.2 (Eq. 3) because depth is unavailable; the resulting projection height depends on this arbitrary value.
Projection plane origin = origin (0,0,0)
The plane is defined by n and the origin in Section 3.2; no camera or world coordinate system is specified, so this is arbitrary.
Bounding box center z-coordinate cz = unspecified (implicitly 0)
Equation 2 defines C=(cx,cy,cz) but no depth map exists; the center's z is never defined, affecting the projection in Eq. 5.

assumptions (4)

domain assumption Bae and Davison (2024) surface normal estimator is accurate for the target surfaces.
Both the conditioning input and the MAE-N metric rely on this estimator; a biased estimator would corrupt alignment and inflate metric scores.
ad hoc to paper Unit depth d=1 can stand in for missing depth information.
Section 3.2, Eq. 3: 'Due to the absence of depth information, we assume a unit depth (d = 1)'.
ad hoc to paper The projection plane passes through the world origin and the text lies on that plane.
Section 3.2, Eqs. 4-5 project onto a plane defined by the normal and 'the origin', with no camera or world coordinate system.
domain assumption Characters remain rectangular after projection.
The method preserves the bbox width/height and axis-alignment after the center projection, which ignores perspective distortion that would shear or rotate characters.

how reviews work

0 comments

Cite this review

Pith. "Pith review of OrienText: Surface Oriented Textual Image Generation." pith.science (2026). https://pith.science/paper/NPWY7U2F

@misc{pith2026250520958,
  author       = {Pith},
  title        = {Pith review of: OrienText: Surface Oriented Textual Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NPWY7U2F}},
  note         = {Machine review of arXiv:2505.20958}
}

read the original abstract

Textual content in images is crucial in e-commerce sectors, particularly in marketing campaigns, product imaging, advertising, and the entertainment industry. Current text-to-image (T2I) generation diffusion models, though proficient at producing high-quality images, often struggle to incorporate text accurately onto complex surfaces with varied perspectives, such as angled views of architectural elements like buildings, banners, or walls. In this paper, we introduce the Surface Oriented Textual Image Generation (OrienText) method, which leverages region-specific surface normals as conditional input to T2I generation diffusion model. Our approach ensures accurate rendering and correct orientation of the text within the image context. We demonstrate the effectiveness of the OrienText method on a self-curated dataset of images and compare it against the existing textual image generation methods.

Figures

Figures reproduced from arXiv: 2505.20958 by the authors.

**Figure 1.** Proposed Workflow: (a) Traditional T2I methods such as TextDiffuser [Chen et al. 2023] use Source image (S), Text-prompt (P), Region-of-Interest (ROI), and Character Mask (𝐶𝑚𝑎𝑠𝑘 ) and generate an image where the text often appears overlaid and frontal, lacking proper alignment with the background surface. (b) Our proposed OrienText method takes Surface Normal (N) as an additional input to the ControlNet-augmented T2… view at source ↗

**Figure 2.** Human Evaluation Results: (a) Average scores and corresponding variances across three quality parameters. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Screenshot of the survey interface where partici [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 1.** Figure 1: Example workflow for generating a product advertisement image: We first generate a base image using a diffusion model, then customize it with OrienText by adding or editing the required multiline text labels. Regeneration by OrienText (a) Original Image (b) Generated …

**Figure 2.** Figure 2 [PITH_FULL_IMAGE:figures/full_fig_p005_2.png]

**Figure 4.** Figure 4: Qualitative comparison of generated images using different methods such as TextDiffuser [2], TextDiffuser-2 [1], AnyText [3] and our proposed approach OrienText for disparate image editing tasks. SA Technical Communications ’24, December 3–6, 2024, Tokyo,Japan © 2024 C…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

5 extracted references · 2 canonical work pages

[1]

Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, and Qifeng Chen. 2024. Textdif- fuser: Diffusion models as text painters. Advances in NeurIPS 36 (2024)

work page 2024
[2]

Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei
[3]

Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie
[5]

arXiv preprint arXiv:2311.03054 (2023)

Anytext: Multilingual visual text generation and editing. arXiv preprint arXiv:2311.03054 (2023). Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citati...

arXiv 2023
[2023]

arXiv (2023)

TextDiffuser: Diffusion Models as Text Painters. arXiv (2023)

work page 2023

Pith tools

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

[1] [1]

Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, and Qifeng Chen. 2024. Textdif- fuser: Diffusion models as text painters. Advances in NeurIPS 36 (2024)

work page 2024

[2] [2]

Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei

[3] [3]

Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie

[4] [5]

arXiv preprint arXiv:2311.03054 (2023)

Anytext: Multilingual visual text generation and editing. arXiv preprint arXiv:2311.03054 (2023). Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citati...

arXiv 2023

[5] [2023]

arXiv (2023)

TextDiffuser: Diffusion Models as Text Painters. arXiv (2023)

work page 2023