- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper argues that stroke-based rendering (SBR) should consider not only reconstruction fidelity but also the organization of the strokes. Treating the output as an ordered stroke sequence (a "painting program"), HarnessPainter proceeds in three stages: (1) semantic ordering, which fits strokes in six stages toward coarse-to-fine targets (2) alignment, which rotates, bends, and translates strokes to follow a flow field derived from edge tangent flow and guide curves specified by a vision-language model (VLM) (3) consolidation, which merges compatible strokes into fewer curves. The VLM makes high-level decisions, local tools perform the geometric edits, and a "harness" accepts an edit only if it passes geometric checks and PSNR bounds.
Critical strengths
Viewing strokes as a program is an interesting perspective. The paper treats an SBR result as an executable stroke program whose organization matters in its own right, and refines it through VLM decisions that are executed and verified by tools. This viewpoint is the main reason I find the submission interesting.
Critical weaknesses
- The formulation and presentation make it hard to see what the method is and what is new.
Related work: Section 2 is very condensed. Each prior method is summarized in a short phrase relying on terms specific to the cited paper, and the paper's own contribution is described in similarly specialized terms (e.g., "stable IDs, protected context, replacement lineage, and rendered acceptance gates"). It is therefore hard to grasp the core idea of each prior work, what HarnessPainter builds on, and what is new. Beyond the works cited, stroke order and the alignment of strokes with contours have also been discussed in the non-photorealistic rendering literature.
Formulation: The abstract states that stroke organization is formulated "as an executable program refinement problem", but the program is simply an ordered list of stroke parameters, and no objective for a well-organized program is defined. The desired properties appear only implicitly in the acceptance checks.
Section 3: Each step is easy to follow, but the purpose of the procedures and the choice of numerical values are rarely explained. For example, "Locations with support at least 0.08 provide reliable guidance" relies on a quantity, "support", that is defined only verbally, with an unexplained threshold. The PSNR tolerances in Eqs. (3)–(4) and the many thresholds in Appendices A.1.3–A.1.4 are similar. The section reads more like an implementation with hard-coded constants than a method with motivated design choices.
Reproducibility: The VLM, which makes key decisions at every stage, is not identified, and its prompts and input/output formats are not provided.
- Parts of the manuscript appear unfinished.
Several passages read as working notes or plans. Appendix A.1.3 states that the conversion of the width values used in several thresholds to the renderer's half-width "remains to be confirmed". Appendix A.1.6 states that a clamping surrogate "and its threshold must be recorded" without reporting them. Appendix A.2 lists "planned extensions" for which "no measurements or qualitative outcomes are yet available". Some statements refer to things absent from the paper, such as "comparisons with fixed-rule control and one-pass planning", "generated states as targets", a "generator", and the "in-context" organization in the title.
- The evaluation does not substantiate the main claims.
Section 4.1 notes that stroke organization, artistic character, and drawing order matter beyond pixel fidelity, and the abstract claims "improved stroke coherence and process-level quality over existing SBR approaches". However, these qualities are neither defined nor measured. The evidence consists of visual comparisons and a single PSNR comparison in the ablation. Moreover, the only baseline is "unguided SBR", which is not described in the paper and appears to have been constructed by the authors. None of the SBR methods discussed in Section 2 is compared against.
Questions for authors
- What exactly is "unguided SBR"? Why are none of the existing SBR methods discussed in Section 2 used as baselines?
- Where is differentiable rendering used? The abstract mentions "differentiable SBR operators" and Appendix A.1.6 derives a gradient, but the alignment updates appear rule-based, and the loss and optimizer for the stage fitting and the appearance fit are not specified.
- How were the thresholds chosen, and how sensitive are the results to them?
- Which VLM is used, and could the prompts and input/output formats be provided?