{"id":"ff59ce4a-195e-46ab-8e69-335d53ded15f","arxiv_id":"2505.20958","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"OrienText uses surface-normal maps and a control branch on a diffusion model to render text that follows the angle of the surface it is placed on.","lead":"This paper introduces a way to make AI image generators place text on slanted surfaces, like billboards seen at an angle, so the text follows the surface geometry instead of looking flat. It works by feeding the generator a map of surface directions (normals) and shifting each letter's position to match that surface.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Sec. 3.2 mask alignment never rotates or shears character boxes around the surface normal, and Eqs. 1-5 are internally inconsistent, so the central claim that OrienText orients text correctly is unsupported.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing point: the geometric projection in Sec. 3.2. I agree. For the abstract's claim to be true, the surface-normal conditioning must actually change the orientation of generated text. The paper's only concrete mechanism for that is the character-mask transformation, and that transformation as written leaves boxes axis-aligned and has an algebraic error in the projection step. This is not a disagreement with consensus; it is an internal correctness failure of the method description. The human evaluation provides limited independent signal that the authors' images look better, and ControlNet conditioning on normals is a plausible idea, but neither establishes that the alignment algorithm orients text. The MAE-N metric is self-contradictory as written, and the AnyText ablation row contradicts the claimed universal improvement, further weakening the quantitative case. Because the central mechanism is unvalidated and, as printed, incorrect, the REJECT verdict should stand. A corrected derivation or released code with a clear per-character rotation/shear transform could change that assessment.","tokens_in":7669,"tokens_out":5795,"duration_ms":61395,"concrete_test":"Implement Eqs. 1-5 exactly on a synthetic single-character mask with an oblique normal, e.g., n=(0,1,1)/sqrt(2), then render C^a_mask and measure the angle of the character's bounding-box axis relative to the projected surface direction. If the box stays axis-aligned, or if its angle does not track the normal's slant, the alignment step cannot orient text. Also run the sanity check n=(0,0,1), c_z=0: the published equations give C_p=(c_x,c_y,-1), whereas an orthogonal projection onto the z=0 plane gives (c_x,c_y,0); this exposes the internal inconsistency. If code is released, rerun Table 1 with the corrected mask to see whether the claimed improvements survive.","verdict_should_be":"UNCHANGED","load_bearing_attack":"For the central claim to hold, C^a_mask must encode the text plane's orientation on the imaged surface. The published construction does not. Eqs. 1-5 translate and project the bounding-box center, but the four corners are then computed by translating C_p by +/-w/2 in x and +/-h/2 in y, leaving rectangles axis-aligned in image coordinates; no rotation, shear, or foreshortening is applied. A character on an oblique banner therefore receives the same upright box as a character on a fronto-parallel wall, so the mask cannot be the mechanism that produces the claimed orientation. The derivation is also internally inconsistent: with n=(0,0,1) and c_z=0, Eqs. 3-5 give C_p=(c_x,c_y,-1), not the orthogonal projection (c_x,c_y,0), because Eq. 5 uses C where the projection of C' requires n.C'/||n||^2. Unless Eq. 5 is a typo and the corner computation is replaced by a per-character affine/perspective transform, the method as written does not perform perspective alignment. The surrounding evidence does not repair this: MAE-N is defined as a mean cosine similarity while being called an angular error; Table 1's AnyText row worsens with mask alignment (1.8937 to 1.8955), contradicting the universal-improvement claim; and no code or data are released.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":7804,"tokens_out":3870,"duration_ms":40044,"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":[{"comment":"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.","section":"Sec. 3.2, Eqs. (1)-(5)"},{"comment":"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.","section":"Sec. 4.2.1 and Table 1"},{"comment":"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.","section":"Table 1, 'Anytext' row"}],"minor_comments":[{"comment":"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.","section":"Sec. 3.2"},{"comment":"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.","section":"Sec. 4.1"},{"comment":"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.","section":"References"},{"comment":"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.","section":"Supplementary, Fig. 2"}],"recommendation":"reject","confidential_remarks":"The paper is a short technical-communications-style manuscript, but the central geometric derivation in Sec. 3.2 is not merely incomplete; as written it cannot produce the perspective-oriented masks that the method claims. The evaluation metric is also circular because the same surface-normal estimator is used for conditioning and evaluation, and its definition is inconsistent with the reported numbers. These are load-bearing issues that would require a new alignment algorithm and a re-designed evaluation to address, which is beyond a standard revision. The paper does not release code or data, further limiting verification. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nBut 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.\n\nThe 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.\n\nWhat'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.\n\nFor 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.","headline":"Promising idea, broken write-up: the mask alignment equations don't actually orient text, and the metric is self-contradictory.","tokens_in":8494,"tokens_out":3899,"would_cite":false,"duration_ms":36005,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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…","keywords":["text-to-image generation","diffusion models","surface normal estimation","text rendering","perspective alignment","text editing","ControlNet conditioning","scene text generation"],"falsifier":"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.","tokens_in":7295,"feed_emoji":"🪧","tokens_out":7924,"duration_ms":70767,"temperature":0.7,"pith_summary":"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.","feed_headline":"Text bends to fit angled walls, signs, and packaging","feed_subtitle":"OrienText feeds surface normals into a diffusion model so letters follow the plane they sit on instead of floating flat.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the single-image surface normal estimation that OrienText uses as the conditioning signal and for the MAE-N evaluation.","marker":"[Bae and Davison 2024]"},{"why":"The base TextDiffuser model that OrienText augments with surface-normal control and the principal method compared in the human study.","marker":"[Chen et al. 2023]"},{"why":"Provides the TextDiffuser-2 baseline used in quantitative and human comparisons.","marker":"[Chen et al. 2024]"},{"why":"Provides the AnyText baseline used in quantitative and human comparisons.","marker":"[Tuo et al. 2023]"},{"why":"Supplies the SCUT dataset used to build the training set for the ControlNet-augmented model.","marker":"[Lyu and Liao 2018]"}],"fun_headline_variants":["Surface normals orient text on angled planes","OrienText feeds normals to bend text realistically","Text follows the surface via region-specific normals","Normals condition diffusion for correct text orientation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Surface normals orient text on angled planes","OrienText feeds normals to bend text realistically","Text follows the surface via region-specific normals","Normals condition diffusion for correct text orientation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1532,"prompt_tokens":885,"completion_tokens":647,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":589}},"tokens_in":501,"tokens_out":647,"duration_ms":7089,"temperature":1.0,"reasoning_tokens":589,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:43:10.863435+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the TextDiffuser-2 baseline used in quantitative and human comparisons."}],"review_version":1}