{"id":"e13a1056-f2d1-4c1e-950a-664a5c151d68","arxiv_id":"2608.03198","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A six-parameter differentiable brush model plus a diffusion enhancer lets one framework generate handwriting trajectories, rendered images, and polished offline images from text and style references.","lead":"This paper introduces a differentiable brush renderer that converts handwriting trajectories into stylized images, plus a framework that jointly generates online trajectories and offline images from text and style references. Real robotic calligraphy demos show the same trajectories can drive a robot arm to write.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The physical bridge to real handwriting is not validated: training and evaluation run inside the renderer's synthetic loop, the Eq. (3) velocity-pressure proxy is an uncalibrated assumption, and the only real-image evidence (Table 4) uses test-set-tuned noise start steps.","rationale":"The reader's weakest assumption is the surrogate fidelity of D_syn, and the paper's Limitations concede the brush parameters are renderer-level surrogates. My analysis agrees and sharpens the point: the online evaluation is also synthetic (the paired test split is rendered by the same R), so the only real-image quantitative support is Table 4, which is weakened by per-dataset test-set tuning of t0 and by the absence of error bars. The velocity-pressure proxy (Eq. 3) is the mathematical core that connects kinematics to appearance; its stroke-wise normalization discards absolute speed, so the renderer is not actually calibrated to physical pressure. The proposed check directly measures whether this proxy and the renderer transfer to real recorded handwriting. Because these issues affect the central claim but are addressable with additional experiments and code/data release, the appropriate outcome remains CONDITIONAL acceptance; no change from the reader's verdict is needed.","tokens_in":16950,"tokens_out":7661,"duration_ms":95633,"concrete_test":"Collect a small real paired dataset with a pressure-sensitive tablet: timestamped (x,y,measured pressure) trajectories for ~50 words, plus aligned scanned/photographed pages. First compute the per-stroke Spearman correlation between measured pressure and Eq. (3)'s predicted pressure; then render each trajectory twice with the same θ, once using measured pressure and once using the proxy, and compare binarized stroke width/ink density against the real page. If the proxy correlates poorly or its rendered output is no closer to the real page than a fixed-width baseline, the physical bridge fails on real data. Also rerun Table 4 with a fixed t0 selected on a held-out split to check whether FID/BFID gains persist.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the renderer's synthetic paired dataset is a faithful proxy for real trajectory-image data. This is not tested: G and O are trained and evaluated on images produced by the same R (Tables 2-3; the paired test split is rendered by R), so online results largely measure self-reconstruction in the renderer's own image manifold. The only quantitative evidence on real handwriting images is Table 4, where the renderer guides a pretrained diffusion model. But t0 is chosen per dataset (50/900/200 in Sec. 5.1) using the test set, so the reported FID/BFID gains may be inflated by test-set tuning; no error bars or validation-set selection are reported. The physical link itself is also weakly grounded: Eq. (3) converts a stroke-wise normalized displacement ||u_{t+1}-u_t||_2 into a pressure proxy under an inverse speed-pressure assumption. Because normalization is stroke-wise, absolute speed and absolute pressure are discarded; a uniformly slow heavy stroke and a uniformly fast light stroke can render identically. The renderer therefore does not establish a calibrated physical relation between kinematics and ink deposition. The robot demonstration is qualitative and does not measure whether predicted θ or pressure maps to real stroke width. Thus the claimed bridge from stroke dynamics to visual appearance for real handwriting is not yet supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unified online–offline handwriting generation framework built around a compact differentiable brush renderer R. The renderer maps an online stroke sequence S and six brush parameters θ to a raster image I_rend, using a velocity-derived pressure proxy, pressure-dependent stroke width, ink opacity, and max-composition blending (Sec. 3). The renderer is used to synthesize a large paired dataset D_syn from online trajectory data (IAM-OnDB, CASIA-OLHWDB), on which a text-to-stroke generator G and a brush parameter observer O are trained (Sec. 4). The rendered image is also used as a structural prior for zero-shot refinement by pretrained diffusion models (Sec. 4.4). Experiments report improved online trajectory metrics (DTW) over SDT, strong performance on the synthetic paired dataset, improved FID/BFID/KID when the renderer is attached to diffusion baselines on IAM and CVL, and a qualitative robot-writing demonstration (Sec. 5).","tokens_in":17388,"tokens_out":2347,"duration_ms":30196,"significance":"If the central claim holds, the paper makes a useful contribution: a differentiable, low-dimensional brush model that connects trajectory generation to image synthesis, enabling synthetic paired-data construction, zero-shot guidance of diffusion models, and direct robot execution. The renderer equations are fully specified and the real-data improvements of One-DM+Ours over One-DM (e.g., FID 25.89 vs 29.00 on IAM, 14.45 vs 19.45 on CVL) are concrete and reproducible in principle. The framework is modular and the robot demonstration, though qualitative, illustrates a plausible application. However, the evidence for a genuine physical bridge to real handwriting is currently incomplete: the synthetic paired evaluation is circular, the pressure proxy is uncalibrated, and the real-data gains rely on test-set-tuned hyperparameters. These are fixable with additional experiments and analysis, but they are load-bearing for the paper's main claim.","major_comments":[{"comment":"The evaluation of the brush parameter observer and the overall framework on the synthetic paired dataset is circular. D_syn is constructed by rendering online trajectories with randomly sampled θ using R, and the same R is used to generate both training and test pairs. The near-perfect LPIPS (0.001) and very low FID (11.81) for 'Our Renderer' are expected because the test images lie exactly on the renderer's image manifold. This does not validate that O recovers meaningful brush parameters for real handwriting images or that G generalizes beyond the synthetic domain. Please add an evaluation that breaks this loop, e.g., train on D_syn but evaluate on real trajectory–image pairs (even partially), or test generalisation to a different renderer with different composition rules or held-out physical parameters.","section":"Sec. 3.3, Table 3"},{"comment":"The pressure proxy is a kinematic heuristic that is not calibrated to any physical measurement. Because v_t is normalized stroke-wise in Eq. (3), absolute writing speed is discarded: a uniformly slow, heavy stroke and a uniformly fast, light stroke produce identical p_proxy sequences. The inverse speed–pressure assumption is plausible but unvalidated. The robot demonstration (Sec. 5.2) does not measure whether the predicted θ or pressure proxy maps to physical stroke width or contact force. Please provide a calibration experiment, e.g., compare predicted p_proxy against measured pen pressure from a sensor-equipped device or against stroke widths produced by a robot under controlled Z-axis displacement. This is essential to support the paper's claim of a 'physical bridge' between kinematics and appearance.","section":"Sec. 3.2, Eq. (3)"},{"comment":"The noise-injection start step t0 for the diffusion enhancer is tuned per dataset ('50 for DiffPen+Ours on both IAM and CVL, 900 for One-DM+Ours on IAM, and 200 for One-DM+Ours on CVL') and reported on the same test sets used for the final metrics. No validation split, selection procedure, or error bars are given. Since FID is sensitive to t0, the reported improvements of One-DM+Ours over One-DM may reflect test-set tuning rather than a robust benefit. Please report results for a range of t0 values, or fix t0 on a validation set and report test-set results with confidence intervals. Without this, the central quantitative evidence for the real-image contribution is not conclusive.","section":"Sec. 5.1, Table 4 and Sec. 5.1 text"},{"comment":"The robotic writing demonstration is qualitative and does not quantify the mapping from predicted brush parameters to physical outcomes. The text states that a writing tool is selected based on θ and that pressure proxy is mapped to Z-axis displacement, but no measurement is reported of whether the rendered stroke width, ink density, or edge sharpness matches the robot's actual writing. Given the paper's emphasis on physical executability and the acknowledged 'surrogate parameters' limitation, this missing calibration leaves the physical-bridge claim unsupported. Please add quantitative measurements, e.g., stroke width or ink coverage of the robot output compared with R(S, θ), or at least controlled ablations of the Z-axis mapping.","section":"Sec. 5.2, Appendix D"}],"minor_comments":[{"comment":"The kernel exponent γ is defined as γ_min + λ(1 − σ_sharp), but the ranges of γ_min, λ, and σ_sharp are not given until Appendix C.2. Please state the ranges in the main text for clarity.","section":"Sec. 3.2, Eq. (7)"},{"comment":"The loss notation L_params and L_render uses MSE on raw parameters and images; consider normalizing the parameter MSE since θ components have different scales and units.","section":"Sec. 4.3, Eq. (loss)"},{"comment":"The DTW comparison with SDT is not apples-to-apples because SDT is a single-character model and words are formed by concatenation. This is noted in the text, but the table would benefit from a column clarifying the concatenation setting.","section":"Sec. 5.1, Table 2"},{"comment":"The observer is trained with four reference images but at inference two are sampled from fifteen style references. The discrepancy is mentioned only implicitly; please state whether this mismatch affects performance and whether the number of references is a hyperparameter.","section":"Appendix B, 'Brush Parameter ObserverO'"},{"comment":"CER for One-DM+Ours on CVL (0.6285) is much worse than One-DM (0.2877); the text mentions 'backbone- and dataset-dependent trade-offs' but does not explain this large degradation. A brief explanation would be useful.","section":"Sec. 5.1, Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is potentially valuable but currently overclaims its physical grounding. The synthetic evaluation is circular, and the only real-data improvements depend on test-set-tuned t0. These issues are fixable within the scope of a major revision: add non-circular evaluation, calibrate or at least validate the pressure proxy on real data, and report robust t0 selection. I would not reject, but I would not accept without these changes. The robot demonstration, while appealing, needs quantitative validation to support the 'physical bridge' claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, what's worth knowing: the differentiable six-parameter brush renderer is a genuinely new formulation. It distills classical brush models into a compact gradient-friendly surrogate, and the max-composition rule is a sensible choice for self-overlapping strokes. The full pipeline—stroke generator, brush observer, renderer, diffusion refiner—is coherent, and the word-level extension of SDT is useful.\n\nThe most convincing evidence is Table 4: attaching the renderer to One-DM improves FID from 29.00 to 25.89 on IAM and from 19.45 to 14.45 on CVL, with similar BFID improvements. That's a concrete real-data benefit. The paper also deserves credit for stating its limitations clearly: the brush parameters are renderer-level surrogates, not calibrated to physical units, and the robot demo is explicitly qualitative.\n\nThe soft spots are real but not fatal. The synthetic paired dataset evaluation (Table 3) is largely circular: the renderer generates both the training and test images, so near-perfect LPIPS is expected. The online DTW numbers are also measured on rendered trajectories, so they mainly show the model can reconstruct the renderer's own manifold. The only real-image evaluation is Table 4, and there t0 is tuned per dataset (50/900/200) apparently on the test set, with no error bars or validation split reported. That makes the FID gains harder to fully trust, though the improvement is consistent across two datasets and two backbones.\n\nThe deeper concern is the physical bridge itself. Eq. (3) uses a stroke-wise normalized speed as a pressure proxy, which discards absolute speed and pressure. A uniformly slow heavy stroke and a uniformly fast light stroke can render identically. The paper acknowledges the parameters are surrogates, but the claim that the model 'bridges stroke dynamics and visual appearance' for real handwriting is not yet demonstrated. The robot execution maps the pressure proxy to Z-depth, but no measurement shows whether predicted θ or width transfers to real stroke appearance.\n\nWho is this for? Anyone working on handwriting generation, and readers interested in differentiable rendering as a tool for grounded synthesis. It deserves a serious referee. I'd ask for code/data release, error bars, validation-set t0 selection, and at least one real-image test of the observer (e.g., predict brush parameters from a real handwriting image and render the trajectory). With those, the paper could be solid.","headline":"A compact differentiable brush renderer that shows real but modest gains on real handwriting images, though the claimed physical bridge to real handwriting rests on synthetic-loop evaluation and an uncalibrated pressure proxy.","tokens_in":17796,"tokens_out":3619,"would_cite":false,"duration_ms":37264,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A six-parameter differentiable brush renderer lets one framework generate stroke trajectories, stylized images, and robot-executable handwriting.","keywords":["online handwriting","offline handwriting","differentiable rendering","brush model","stroke trajectory","diffusion models","robotic calligraphy","handwriting generation"],"falsifier":"Collect paired trajectory-image data from humans writing on a pressure-sensitive tablet with a camera, and test whether (a) the velocity-derived pressure proxy correlates with measured pressure, and (b) trajectories predicted by the stroke generator, when re-rendered, reproduce the real ink appearance; if either fails, the bridge is a synthetic-domain-only construction.","tokens_in":16887,"feed_emoji":"🖌️","tokens_out":11471,"duration_ms":103311,"temperature":0.7,"pith_summary":"This paper tries to establish that a six-parameter physical brush model, wrapped in a differentiable renderer, is enough to bridge online handwriting (stroke trajectories) and offline handwriting (pixel images) within a single framework. The renderer turns any stroke sequence into an image using a velocity-derived pressure proxy, a pressure-dependent footprint, and ink-opacity rules, and because every step is differentiable, gradients can flow from pixels back to trajectory coordinates. Around this renderer the authors build a text-to-stroke generator, a brush-parameter observer, and a diffusion-based refiner, and they train these on a synthetic paired dataset made by rendering existing online trajectories with sampled brush parameters onto real backgrounds. If the claim holds, online and offline handwriting generation stop being separate problems: one model would output an executable trajectory, a rendered image, and a refined realistic image, which matters for font design, data augmentation, and robotic calligraphy.","feed_headline":"Six brush parameters unite online and offline handwriting","feed_subtitle":"A differentiable renderer turns stroke paths into images, so one model gives both executable trajectories and realistic pixels.","key_machinery":"The differentiable brush renderer $\\mathcal{R}$, parameterized by $\\theta = (w_{\\mathrm{base}}, k_{\\mathrm{spread}}, \\rho_{\\mathrm{ink}}, \\sigma_{\\mathrm{sharp}}, p_{\\min}, p_{\\max})$, is the load-bearing object. For each stroke segment it computes a pressure proxy $p_t$ from inverse writing speed, an effective width $w_t = w_{\\mathrm{base}} + k_{\\mathrm{spread}} p_t$, a footprint kernel $K_t(q) = \\max(0, 1 - \\hat{d}_t(q)^\\gamma)$ with $\\gamma$ controlled by $\\sigma_{\\mathrm{sharp}}$, and an ink opacity $O_t = 1 - \\exp(-\\rho_{\\mathrm{ink}} p_t)$; per-segment $\\alpha$ maps are combined by maximum composition. Because every operation is differentiable, $\\mathcal{R}$ can be placed between a stroke","core_discovery":"The central claim is that the writing process can be compressed into six scalar brush parameters, $\\theta = (w_{\\mathrm{base}}, k_{\\mathrm{spread}}, \\rho_{\\mathrm{ink}}, \\sigma_{\\mathrm{sharp}}, p_{\\min}, p_{\\max})$, and that a renderer $\\mathcal{R}$ built from them maps a stroke sequence $S$ to a raster image $I_{\\mathrm{rend}}$ differentiably. The renderer estimates pressure from writing speed with the inverse proxy $p^{\\mathrm{proxy}}_t = p_{\\min} + (p_{\\max} - p_{\\min})(1 - \\tilde{v}_t)$, deforms the stroke footprint linearly via $w_t = w_{\\mathrm{base}} + k_{\\mathrm{spread}} p_t$, describes ink deposition with opacity $O_t = 1 - \\exp(-\\rho_{\\mathrm{ink}} p_t)$, and composites overlappin","pith_inferences":["The same six-parameter bottleneck could serve as a compact, physically interpretable style embedding for tasks beyond generation, such as writer style comparison or OCR data augmentation, where the explicit meaning of each parameter is a bonus over opaque deep features.","The inverse speed-pressure assumption is empirically testable: if pressure-sensitive tablets show that human writers do not systematically press harder when writing slowly, the renderer's pressure channel would need to be replaced by a learned or measured estimator, while the rest of the pipeline could remain intact.","Because the renderer is differentiable and low-dimensional, it can in principle be inverted to recover stroke trajectories from offline images, enabling trajectory extraction from scanned or historical handwriting—a direction the paper motivates but does not implement.","The maximum-composition rule encodes a physical prior about ink accumulation that could be injected into diffusion or GAN models as an inductive bias, potentially helping them render self-intersecting strokes without explicit renderer guidance."],"forward_implications":["One trained framework outputs a stroke trajectory, a directly rendered image, and a diffusion-refined image from the same text and style input, so offline realism and online executability are no longer separate generation tasks.","Existing online trajectory datasets (IAM-OnDB, CASIA-OLHWDB) become sources of paired trajectory-image training data, lowering the barrier to future joint models.","Using the rendered image as a structural prior consistently improves FID and BFID of diffusion baselines (One-DM, DiffPen) on both IAM and CVL, so the physical prior transfers across generator backbones and datasets.","The brush parameter observer extracts interpretable brush settings (width, ink density, edge sharpness) from style images, which can directly guide physical tool selection and Z-axis pressure control in robotic calligraphy.","Word-level trajectories can be composed into sentence-level sequences with fixed offsets, so the renderer's utility extends beyond the word-level training setup."],"supporting_citations":[{"why":"Establishes the hairy-brush model the paper distills: stroke as a trajectory of position and pressure, footprint deformation, and maximum composition of overlapping marks.","marker":"[42]"},{"why":"Identifies the visually critical brush factors (footprint, pressure-dependent deformation, ink supply) that the six-parameter model compresses.","marker":"[11]"},{"why":"Supplies the transformer-based trajectory estimation architecture that the stroke generator extends from single characters to word-level text.","marker":"[13]"},{"why":"One of the two diffusion handwriting baselines used as the zero-shot refiner and as a comparison; the renderer guidance improves its FID and BFID.","marker":"[12]"},{"why":"The other diffusion baseline used for refinement and comparison; renderer guidance improves its distributional realism on both datasets.","marker":"[32]"},{"why":"IAM-OnDB online handwriting dataset whose trajectories are rendered to build the synthetic paired dataset and to evaluate online generation.","marker":"[29]"},{"why":"CASIA-OLHWDB online dataset used together with IAM-OnDB to construct paired training data and evaluate online performance.","marker":"[27]"},{"why":"Provides the background composition strategy used to composite rendered ink onto realistic paper, cardboard, and wood textures.","marker":"[36]"},{"why":"IAM offline database supplies the test images for offline handwriting generation evaluation.","marker":"[31]"},{"why":"CVL offline database supplies the second test set for offline handwriting generation evaluation.","marker":"[23]"}],"fun_headline_variants":["Differentiable brush renderer unifies handwriting paths and pixels","Six brush parameters make handwriting renderable differentiably","From strokes to pixels: a differentiable brush model","Unified handwriting generation via differentiable brush rendering","Six parameters link stroke dynamics to ink appearance"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that online trajectories rendered through the six-parameter brush model with its velocity-derived pressure proxy, composited onto real backgrounds, make a faithful enough proxy for real paired trajectory-image data that the learned generators and observers transfer to real handwriting and robotic writing—a premise the paper itself softens in its Limitations by calling the parameters renderer-level surrogates rather than physical quantities.","fun_headline_variants_meta":{"raw":{"variants":["Differentiable brush renderer unifies handwriting paths and pixels","Six brush parameters make handwriting renderable differentiably","From strokes to pixels: a differentiable brush model","Unified handwriting generation via differentiable brush rendering","Six parameters link stroke dynamics to ink appearance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000721,"raw_usage":{"total_tokens":3125,"prompt_tokens":848,"completion_tokens":2277,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":2205}},"tokens_in":592,"tokens_out":2277,"duration_ms":17678,"temperature":1.0,"reasoning_tokens":2205,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T23:42:20.672868+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect paired trajectory-image data from humans writing on a pressure-sensitive tablet with a camera, and test whether (a) the velocity-derived pressure proxy correlates with measured pressure, and (b) trajectories predicted by the stroke generator, when re-rendered, reproduce the real ink appearance; if either fails, the bridge is a synthetic-domain-only construction.","supporting_citations":[{"cited_title":"ACM Siggraph Computer Graphics (1986)","cited_arxiv_id":null,"evidence_quote":"Establishes the hairy-brush model the paper distills: stroke as a trajectory of position and pressure, footprint deformation, and maximum composition of overlapping marks."},{"cited_title":"In: 10th Pacific Conference on Computer Graphics and Applications, 2002","cited_arxiv_id":null,"evidence_quote":"Identifies the visually critical brush factors (footprint, pressure-dependent deformation, ink supply) that the six-parameter model compresses."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)","cited_arxiv_id":null,"evidence_quote":"Supplies the transformer-based trajectory estimation architecture that the stroke generator extends from single characters to word-level text."},{"cited_title":"In: Proceedings of the European Conference on Computer Vision (ECCV) (2024)","cited_arxiv_id":null,"evidence_quote":"One of the two diffusion handwriting baselines used as the zero-shot refiner and as a comparison; the renderer guidance improves its FID and BFID."},{"cited_title":"In: Proceedings of the European Conference on Computer Vision (ECCV) (2024)","cited_arxiv_id":null,"evidence_quote":"The other diffusion baseline used for refinement and comparison; renderer guidance improves its distributional realism on both datasets."},{"cited_title":"In: International Conference on Document Analysis and Recognition (ICDAR) (2005)","cited_arxiv_id":null,"evidence_quote":"IAM-OnDB online handwriting dataset whose trajectories are rendered to build the synthetic paired dataset and to evaluate online generation."},{"cited_title":"In: International Conference on Document Analysis and Recognition (ICDAR)","cited_arxiv_id":null,"evidence_quote":"CASIA-OLHWDB online dataset used together with IAM-OnDB to construct paired training data and evaluate online performance."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025)","cited_arxiv_id":null,"evidence_quote":"Provides the background composition strategy used to composite rendered ink onto realistic paper, cardboard, and wood textures."},{"cited_title":"International journal on document analysis and recognition (2002)","cited_arxiv_id":null,"evidence_quote":"IAM offline database supplies the test images for offline handwriting generation evaluation."},{"cited_title":"In: International Conference on Document Analysis and Recognition (ICDAR) (2013)","cited_arxiv_id":null,"evidence_quote":"CVL offline database supplies the second test set for offline handwriting generation evaluation."}],"review_version":1}