PaperlayerView exact revision on arXiv

Source-linked AI summary

Inverse Garment Design with Visual Agents

Nanxiang Jiang, Yunuo Chen, Zeng Tao, Zizhang Li, Qi Guo, Jiajun Wu, Yin Yang, Chenfanfu Jiang

arXiv:2610.15582v1cs.AI

TL;DR

Inverse garment design aims to recover simulation-ready sewing patterns from a single image, while existing approaches face trade-offs in speed, generalization, or visual grounding. DressAgent combines an extended layered representation and dataset with coarse-to-fine agentic planning and closed-loop physical-simulation refinement. It reports stronger overall accuracy and preference than feed-forward methods and far less time than differentiable optimization.

  • Problem

    Single-image sewing-pattern recovery must handle arbitrary clothing layers, while differentiable optimization is time-consuming, feed-forward methods are distribution-limited, and one-shot VLM approaches struggle with subtle details.

    Source: §1 · p. 1

  • Method

    DressAgent uses an extended layered garment representation and dataset, then plans layer order and refines patterns through agents that edit, simulate, and judge against the reference.

    Source: §1 · p. 2

  • Results

    DressAgent achieves the best overall accuracy and human preference, generalizes beyond feed-forward methods, and takes far less time than differentiable optimization.

    Source: §1 · p. 2, §5 · p. 10

  • Takeaways & Limitations

    The paper demonstrates single-image recovery of simulation-ready sewing-pattern outfits with arbitrary layers and their relations.

    Source: §1 · pp. 1–2

    The current representation cannot express some free-form ornaments, and the method does not support garments with thickness such as padded down jackets.

    Source: I LIMITATIONS AND FUTURE WORK · p. 42

Abstract

from arXiv · show

We study the problem of inverse garment design, reconstructing simulation-ready 3D outfits with arbitrary layers from a single image, with visual agents. Unlike differentiable optimization that is time-consuming, or feed-forward methods limited by training data, we introduce DressAgent, a hierarchical agentic system that recovers garment sewing patterns through planning and closed-loop refinement grounded in physical simulation. To this end, we first present an enhanced representation based on GarmentCode for diverse styles and arbitrary dressing layers, and build the first multi-layer sewing pattern dataset with over 15K outfits for supervision and benchmarking. Inverse design then proceeds coarse to fine: In layer planning and coarse estimation, agents infer layer order and relations, peel each layer into its own view, and estimate coarse patterns. In fine-grained refinement, an agentic trio of Tailor, Critic, and Orchestrator refines patterns in a closed loop of editing, simulating, and judging against the reference with visual and measured evidence. Carefully designed tools and an evolving memory make this refinement converge on the target geometry. We also design an agentic generation method for high-fidelity textures. DressAgent outperforms feed-forward methods by a large margin, surpasses or matches differentiable optimization in a fraction of the time, and generalizes to complex layering, diverse poses, and body shapes.

1 INTRODUCTION

Inverse garment design seeks simulation-ready sewing patterns from a single in-the-wild image, but existing methods trade speed against generalization and fine-grained visual grounding. DressAgent combines agentic planning and closed-loop physical-simulation refinement with an extended garment representation and multi-layer dataset.

1 INTRODUCTION

  • Prior methods either incur high time cost, depend on training-data distributions, or use open-loop pattern generation that struggles with subtle details; few support arbitrary layers and their relations.Differentiable optimization is described as time-consuming and vulnerable to local minima, while feed-forward methods struggle to generalize to occlusion or complex layering.Source: §1 · p. 1
  • GarmentCode-XL extends GarmentCode with diverse styles and arbitrary dressing layers and their relations, while GarmentCodeData-XL provides the first multi-layer sewing-pattern dataset.Figure 2 reports 15,205 outfits, including 4,640 one-layer, 9,282 two-layer, 721 three-layer, and 562 four-layer examples.Source: §1 · p. 2, Figure 2 · p. 3
  • DressAgent achieves the best overall accuracy and human preference, generalizes beyond feed-forward methods, and takes far less time than differentiable optimization.Experiments also report that it outperforms feed-forward methods by a large margin and surpasses or matches differentiable optimization in a fraction of the time.Source: §1 · p. 2, Figure 1 · p. 2
  • DressAgent recovers arbitrary garment layers and their relations from one in-the-wild image through agentic planning and closed-loop refinement grounded in physical simulation.The system first decomposes layers and estimates coarse patterns, then Tailor, Critic, and Orchestrator iteratively edit, simulate, and judge patterns using visual and measured evidence.Source: §1 · p. 2

2 RELATED WORK

Related work covers garment representations, single-image sewing-pattern recovery, and agentic inverse graphics. The paper positions sewing patterns and code-based representations as simulation-ready, editable structures, and identifies planning with simulation-grounded refinement as its agentic contribution.

Garment Representations

  • Template-based representations preserve topology but are template-bound, while free-form fields support arbitrary shapes but lose garment structure and are hard to rig and simulate.Sewing patterns instead encode 2D panels and stitches, retaining editability and simulation readiness; GarmentCode writes patterns as code.Source: §2 · p. 3

Inverse Garment Design with Sewing Patterns

  • Differentiable optimization recovers geometry at high time and expertise cost, feed-forward methods lose accuracy on in-the-wild images, and VLM pattern estimation remains open-loop.The cited challenges include occlusion, unusual poses, complex layering, and inability to verify or refine subtle details.Source: §2 · p. 3

Agentic Systems for Inverse Graphics

  • Garments pose challenges beyond rigid inverse-graphics scenes because cloth deforms and has complex structures and relations; the paper presents the first agentic inverse garment design system grounded in physical simulation.The cited agentic inverse-graphics examples refine object poses or reconstruct Blender scenes through execution feedback.Source: §2 · p. 3

3 DRESSAGENT

DressAgent reconstructs layered outfits from a single image through a coarse-to-fine pipeline, combining an enhanced sewing-pattern representation, layer planning, closed-loop refinement, texture generation, and physics-based simulation.

3.1 GARMENT REPRESENTATION AND DATASET

  • GarmentCode-XL adds diverse panel constructors and ordered layers with piece relations, while GarmentCodeData-XL provides 15,205 outfits with programs, patterns, meshes, and rendered views.The dataset is used for subsequent training and benchmarking.Source: §3.1 · p. 4

3.2 STAGE 1: LAYER PLANNING AND COARSE ESTIMATION

  • Canonical front and back views and layer peeling expose each garment; code-derived visual questions form templates that an encoder-decoder maps to coarse programs.The workflow also fits a body to provide measurements, a simulation mesh, and cameras for aligning rendered garments with the canonical views.Source: Canonical A-pose Views · pp. 4–5

3.3 STAGE 2: FINE-GRAINED REFINEMENT

  • A Tailor edits and simulates patterns, a Critic scores garments and identifies measured discrepancies, and an Orchestrator directs the next refinement step.The Critic evaluates the simulated garment against the reference, while the Orchestrator reads checkpoint history and can continue, roll back, or finish.Source: §3.3 · pp. 5–6
  • Persistent memory retains patterns, observations, critiques, and rationales across iterations; checkpoints let the Orchestrator revert when Critic scores worsen.Rollback restores the pattern but leaves memory intact, so failed attempts remain available to guide later edits.Source: §3.3 · p. 6

3.4 POST-PROCESSING

  • An agentic method recovers high-fidelity garment textures using garment segmentation and pattern mapping, then selects tools for solid colors, woven textures, repeated motifs, logos, or other prints.The tools include PickColor, WeaveTile, RepeatPattern, PasteLogo, and ContinuePrint.Source: §3.4 · p. 7, Figure 5 · p. 7
  • An XPBD solver stitches seams and drapes each layer under gravity and frictional contact, reusing settled inner layers as colliders for outer layers.For reposing, the body is interpolated to the target pose and cloth is advected by nearby body points under gravity and contact.Source: §3.4 · p. 7

4 EXPERIMENTS

Experiments compare DressAgent with feed-forward and differentiable-optimization methods, then evaluate perceptual quality and ablate its models, iteration budget, memory, representation, and agent tools. The reported results favor DressAgent on most metrics and identify components that support refinement and garment-structure recovery.

4.1 IMPLEMENTATION DETAILS

  • The experiments use GPT-5.6-luna agents with medium reasoning effort and GPT-Image-2 image tools; comparisons include feed-forward and differentiable-optimization baselines.The encoder is a fine-tuned Qwen3.6-27B model, and the decoder is trained on GarmentCodeData and GarmentCodeData-XL splits.Source: §4.1 · p. 7

4.2 RESULTS

  • DressAgent achieves the best quality on most reported metrics and is significantly faster than differentiable optimization, while remaining robust across complex layering, poses, and body shapes.The comparison measures sewing-pattern structure and draped-garment geometry.Source: §4.2 · p. 7, Table 1 · p. 8
  • In a 200-case study with 65 participants, DressAgent leads by a large margin in geometry matching, simulation plausibility, and texture fidelity.Geometry is measured by best-match vote share; plausibility and texture use ratings from 1 to 5.Source: §4.2 · p. 8, Table 2 · p. 9

4.3 ABLATION STUDY

  • Using GPT-5.6-luna (medium) with a six-iteration budget balances quality and cost, as refinement closes most backbone differences and converges after about six iterations.The ablation benchmarks different backbones on 200 GarmentCodeData-XL test cases.Source: §4.3 · pp. 8–9, Figure 8 · p. 9
  • Evolving context memory lets agents progressively align skirt flare to the target; without it, edits become isolated guesses and the flare oscillates.The figure tracks skirt-flare adjustment over iterations.Source: §4.3 · p. 9, Figure 9 · p. 9
  • Ablations show that closed-loop refinement improves over one-pass estimates, GarmentCode-XL avoids the original representation’s design-space ceiling, and the Critic and Measure tool support accuracy.Without the Critic, the Tailor overrates flawed proposals; without Measure, accuracy drops slightly.Source: §4.3 · p. 9, Table 3 · p. 9

5 CONCLUSION

The conclusion presents DressAgent as a hierarchical system for reconstructing simulation-ready outfits from a single in-the-wild image. It combines a layered garment representation and dataset with coarse-to-fine pattern estimation, iterative refinement, and texture generation, and reports strong accuracy, preference, generalization, and speed relative to prior methods.

5 CONCLUSION

  • DressAgent reconstructs simulation-ready outfits from one in-the-wild image using a hierarchical framework with ordered-layer estimation and closed-loop pattern refinement.The Layer Planner decomposes outfits into ordered layers, the Visual Reader estimates each layer, and the Tailor, Critic, and Orchestrator refine patterns with evolving memory.Source: §5 · pp. 9–10
  • GarmentCode-XL represents diverse styles and arbitrary layers, and GarmentCodeData-XL provides the first multi-layer sewing-pattern dataset with over 15K outfits.The paper also presents agentic tools for generating high-fidelity textures.Source: §5 · pp. 9–10
  • Experiments report best overall accuracy and preference, generalization beyond feed-forward methods, and far less time than differentiable optimization.The conclusion frames these results as evidence that the framework advances inverse garment design.Source: §5 · p. 10

B GARMENT REPRESENTATION AND DATASET

GarmentCode-XL expands the garment language to represent diverse styles and layered outfits, while GarmentCodeData-XL and an outfit-level simulator support their reconstruction and simulation.

B GARMENT REPRESENTATION AND DATASET

  • GarmentCode-XL broadens garment styles and adds structured layered outfits through new panel constructors, expanded parameter ranges, and outfit-level fields.Source: B GARMENT REPRESENTATION AND DATASET · p. 17, Figure 10 · p. 17
  • GarmentCode-XL treats an outfit as ordered layers with typed relations, and uses that structure to determine inner-to-outer simulation order.Source: B GARMENT REPRESENTATION AND DATASET · p. 17
  • GarmentCodeData-XL contains 15,205 outfits: 4,640 one-layer, 9,282 two-layer, 721 three-layer, and 562 four-layer examples.Each example includes an outfit program, sewing patterns, body measurements, rendered views, and simulated meshes.Source: B GARMENT REPRESENTATION AND DATASET · p. 18
  • The layered simulator schedules garments by layer sequence and relations, draping over-garments after underlying ones and using settled garments as colliders.For tucked tops, it projects cloth below the waistline onto the body so the bottom can close over it.Source: B GARMENT REPRESENTATION AND DATASET · p. 18
  • For dynamic simulation, settled inner layers are replayed as moving colliders, and the layered drape is accepted only when every layer reaches static equilibrium without body contact or self-intersection.Source: B GARMENT REPRESENTATION AND DATASET · p. 18

C STAGE 1: LAYER PLANNING AND COARSE ESTIMATION

C.1 AGENT IMPLEMENTATION AND INTERACTION

This section describes how visual evidence from outfit images is translated into coarse garment programs through agent data flow, canonical views, and layer planning. It then details agent coordination and tool-mediated simulation, where measured and visual evidence guides edits, judgments, checkpoint decisions, and stopping conditions.

C.1 AGENT IMPLEMENTATION AND INTERACTION

  • The Layer Planner defines garment layers and descriptions; the Visual Reader fills each garment template, which the encoder-decoder converts into programs assembled as a coarse outfit.The same garment descriptions guide peeling and tell the Visual Reader which garment to inspect; the planner’s relations determine how the resulting programs are assembled.Source: §C.1 · p. 19

C.2 CANONICAL A-POSE VIEWS

  • Canonical front and back views preserve the person and garments while standardizing pose and framing, so an analytic 180° orbit aligns the views without an alignment network.The front view is generated in a canonical A-pose, and the matching rear view keeps the pose and framing fixed.Source: §C.2 · pp. 19–20

C.3 LAYER PLANNING

  • The planner orders garments inner-to-outer by shared-region coverage, and gpeel removes outer layers while completing garments beneath, stopping before the bare body is exposed.Planner and Visual Reader replies are schema-checked; invalid replies are returned with offending fields for correction.Source: §C.3 · pp. 21, 23

C.4 QUESTIONNAIRE SYNTHESIS

  • gques reads parameter-construction code, body measurements, and choice bins to write frozen visual questions that describe garment geometry in body-landmark terms.This addresses semantic misalignment when a vision-language model is asked to generate GarmentCode-XL programs directly from parameter names.Source: §C.4 · p. 23
  • A shared glossary stabilizes 30 categorical tokens, including 17 reused across parameters; 131 validated questions cover legal discrete values and 3–6 ordered numeric states mapped to intervals partitioning [0, 1].Numeric interval boundaries align with body landmarks, so visual choices and decoder targets refer to the same garment geometry.Source: §C.4 · pp. 24–25

C.5 TEMPLATE GENERATION WITH THE VISUAL READER

  • The Visual Reader answers only compatible questionnaire branches for each garment type and asymmetry state, keeping separate answer sheets and masking inactive or unobservable answers before decoding.Garments in the same layer can share one multimodal call; construction choices determine which geometry questions apply.Source: §C.5 · p. 25

C.6 ENCODER-DECODER FROM TEMPLATE TO PROGRAM

  • The encoder learns program semantics from 417 code-description units—122 component classes and top-level programs plus 295 methods—and is trained to connect descriptions and parameter-reading code to programs and visual questions.Together, the units cover all 131 design parameters and link each parameter to the geometry it controls.Source: §C.6 · p. 26
  • The decoder keeps the template’s 33 discrete choices fixed and autoregressively generates 98 numeric parameters, using a projection MLP and causal Transformer.The MLP maps 5,120-dimensional encoder conditions to width 512 for decoding.Source: §C.6 · p. 27

D.1 AGENT IMPLEMENTATION AND INTERACTION

  • Three GPT-5.6-luna agents coordinate through checkpoints: after each simulation the Orchestrator invokes the Critic, then continues with the Tailor, rolls back, or finishes; default budget is six simulations per layer.Agents retain session memory, with recent views inline and older ones compacted into notes; rollback restores the program but preserves histories and memories.Source: §D.1 · pp. 27–28

D.2 TOOL INTERFACE

  • Stage 2 combines paired target/render views, before-and-after strips, mask differences, eight measurements per garment, and attributed solver reports; only simulation consumes budget, while other tools are free.Tool access is role-dependent: Tailor can observe and modify, Critic can only observe, and Orchestrator manages control operations such as rollback.Source: §D.2 · pp. 28–29

D.3 PROGRAM EDITING WITH THE TAILOR

  • The Tailor uses construction edits to choose garment structure before numeric size edits, with Fit finding parameter values for target sizes and flagging interval crossings that can change shape.Its context includes critiques, measurements, solver reports, prior edit effects, active parameters, and their visual meanings.Source: §D.3 · p. 29
  • The Tailor edits garment parameters using before-and-after judgments, prioritizing the named garment when cloth passes through the body and avoiding edits that failed to help.Measurements from segmented masks and rendered drapes can be unreliable where edges are hidden or views are generated, so the Tailor weighs them against the images and explains which evidence it trusts.Source: §D.3 · p. 29

D.4 SIZE FITTING WITH THE FIT TOOL

  • The Fit tool converts a target drape measurement into a GarmentCode-XL parameter value by estimating the draft-to-drape relation from simulation history and probing draft space.This lets measurements set a precise edit target while visual intervals and rendered views preserve shape fidelity.Source: §D.4 · p. 30

D.5 DRAPE EVALUATION WITH THE CRITIC

  • The Critic scores each garment on five dimensions, ranks up to six discrepancies, and compares the current drape with the previous step to judge whether refinement is done.The dimensions are construction, proportion, fit, layering, and physical feasibility. It uses shared construction vocabulary and compares geometry while ignoring color, texture, lighting, and simulator seam lines; measurements and masks are evidence to weigh against pictures because hidden edges and generated views can be unreliable.Source: §D.5 · p. 30

D.6 SESSION CONTROL WITH THE ORCHESTRATOR

  • The Orchestrator continues, rolls back, or finishes after each critique, preserving the best physically valid checkpoint and directing the Tailor to address discrepancies in priority order.Cloth passing through the body must be fixed before delivery; continue while major addressable discrepancies remain, and finish when the fit is done, remaining differences exceed the program vocabulary, or the simulation budget is exhausted.Source: §D.6 · p. 31

D.7 AN EXAMPLE REFINEMENT ROUND

The round uses measured and visual critique to correct the skirt, then narrow and reshape the tee through fitted edits and simulation. The simulated result improves silhouette overlap and the Critic’s tee ratings, while leaving a neckline discrepancy.

D.7 AN EXAMPLE REFINEMENT ROUND

  • The Critic says the skirt edit aligns its front and back hems within measurement noise, while the tee remains broad, deep-necked, and long-sleeved.The front hem width is essentially matched; the back hem remains slightly narrow.Source: §D.7 · p. 31
  • The critique prioritizes tee width, then flags a wider tucked hem, a neckline about 2.9 cm too deep, and sleeves about 2.7–3.4 cm too low.The tucked edge is partly affected by the tuck, and the back neckline comparison is less reliable.Source: §D.7 · p. 32
  • The Orchestrator keeps the improved skirt unchanged and directs the next edit toward tee width, then neckline and sleeve length if stable.It leaves the skirt’s minor back-hem residual alone.Source: §D.7 · p. 32
  • The Tailor fits five requested sizes and simulates once; both sleeve parameters hit their range limits before fully reaching the sleeve-end target.It reports a remaining gap when a parameter reaches its limit.Source: §D.7 · p. 32
  • Front silhouette IoU rises from 0.851 to 0.937 after simulation, and the Critic raises the tee’s proportion and fit scores from 6 to 8.The neckline width remains the main discrepancy.Source: §D.7 · p. 32

E TEXTURE GENERATION

The paper routes garment textures among five classes and tools, then compares texture fidelity, layer-structure recovery, and synthesized-view faithfulness. DressAgent leads the reported texture and layer comparisons, while canonical views preserve garment identity and color.

E TEXTURE GENERATION

  • DressAgent’s texture method segments garments, maps observed pixels onto sewing patterns, then routes each garment among five texture classes and class-specific extrapolation tools.Routing uses a garment crop with other pixels masked, plus color-spread and repeat measurements.Source: E TEXTURE GENERATION · pp. 32–33, §E.1 · p. 33

E.1 TEXTURE ROUTING

  • The Layer Planner classifies each garment as solid, non-repeating fabric, placed artwork, exact repeat, or non-repeating print, and selects its corresponding tool.It also describes the print and, for placed artwork, reports a box for each piece.Source: §E.1 · p. 33

E.2 TOOL IMPLEMENTATION

  • PickColor uses the observed texel median; WeaveTile synthesizes a swatch, RepeatPattern aligns observed repeats, PasteLogo restores artwork, and ContinuePrint extends prints within the observed color gamut.The tool-specific procedures address different texture structures rather than applying one extrapolation rule to every garment.Source: §E.2 · pp. 33–34

F.1 BASELINE IMPLEMENTATION

  • All methods receive the same single input and are timed from image to final mesh on one workstation with one RTX PRO 6000 GPU per method.Baselines use official implementations and weights when available, otherwise reproductions from their papers.Source: §F.1 · pp. 34–35

F.2 QUANTITATIVE COMPARISON ON TEXTURE GENERATION

  • DressAgent achieves the best results on all reported texture metrics in both settings, with its largest gap on DINO, which responds strongly to prints and logos.The authors attribute the baselines’ weaker print performance to covering garments with tiled fabric swatches.Source: §F.2 · p. 35, Table 8 · p. 35

F.3 QUANTITATIVE COMPARISON ON LAYER STRUCTURE

  • DressAgent recovers the most correct layer counts and structures on synthetic data and in-the-wild photos; some baselines struggle beyond two layers.GarmentRecovery has zero count accuracy on three- and four-layer outfits, while DressCode often misjudges layer count and produces interpenetration after reposing.Source: §F.3 · p. 36, Table 9 · p. 36

F.4 FAITHFULNESS OF SYNTHESIZED VIEWS

  • The canonical view preserves garment color and identity: its color distance is near that of a real second photo and far below that of unrelated outfits.Its feature similarity is also clearly closer to the source outfit than to other outfits, with the smallest margin on CloSe.Source: §F.4 · pp. 36–37, Table 10 · p. 37

F.5 STRESS TEST: REFINEMENT FROM A PLAIN T-SHIRT

Starting from a plain T-shirt, the agent trio can rebuild garment construction and refine sizes, though the full pipeline’s coarse program gives refinement a stronger start. Additional examples show results across layering, poses, and body shapes.

F.5 STRESS TEST: REFINEMENT FROM A PLAIN T-SHIRT

  • After six simulations from a plain T-shirt, silhouette IoU reaches 0.754 on 4D-Dress and 0.755 on CloSe, versus 0.783 and 0.831 from the Stage-1 program.The trio rewrites construction and adds missing garments before adjusting sizes; the Critic and checkpoints help keep the loop from drifting.Source: §F.5 · p. 38, Figure 16 · p. 38
  • The plain-T-shirt refinement curve lags the Stage-1 initialization by about two iterations and ends 0.03 to 0.08 lower at the same simulation budget.The coarse program fixes construction in a forward pass, leaving more of the refinement budget for fit.Source: §F.5 · p. 38

F.6 ADDITIONAL VISUAL RESULTS

  • The paper presents additional in-the-wild visual results spanning different layering, poses, and body shapes.The stress test discards Stage 1’s program but keeps the canonical views, recovered body, layer plan, backbone, six-simulation budget, and agent trio.Source: §F.5 · p. 37, §F.6 · p. 38

G.1 ABLATION ON SYNTHESIZED VIEWS

The ablation compares synthesized-view supervision against weaker alternatives and reports that each added view improves recovery. The reported gains vary with pose difficulty, and peeling is especially relevant to multilayer supervision.

G.1 ABLATION ON SYNTHESIZED VIEWS

  • Each synthesized view improves recovery, with the canonical view bringing the largest gain and its benefit growing with pose difficulty.The gain is modest on near-A-pose CloSe scans and large on 4D-Dress dance poses.Source: §G.1 · p. 39
  • Adding the back view lowers error further, while peeling further improves recovery by supplying supervision for every layer.The no-peeling condition is tested only on 4D-Dress because CloSe has few multilayer outfits.Source: §G.1 · p. 39, Table 12 · p. 39

G.2 ABLATION ON AGENTIC TRIO DESIGN

Rollback lets the agentic trio undo harmful pattern edits and retain failed attempts as references. Without it, construction errors or tangled drapes persist to delivery.

G.2 ABLATION ON AGENTIC TRIO DESIGN

  • Without Rollback, failed construction edits or tangled drapes persist to delivery, reducing structure accuracy and IoU while quadrupling self-intersecting edges.Rollback returns the pattern to an earlier checkpoint and keeps the failed attempt in memory for later edits.Source: §G.2 · p. 39, Table 13 · p. 39

G.3 ABLATION ON SINGLE-AGENT REFINEMENT

The single-agent refinement baseline oscillates and plateaus, while the agentic trio improves progressively. The comparison also exposes self-evaluation bias and higher cost in the single-agent loop.

G.3 ABLATION ON SINGLE-AGENT REFINEMENT

  • The trio separates editing, judging, and session control, so an independent Critic evaluates edits and the Orchestrator can roll back bad steps.The single-agent baseline instead judges its own drape and rewrites the program itself.Source: §G.3 · p. 39
  • The baseline uses a TextGrad-style loop in which one agent evaluates the simulated drape, writes textual feedback, and rewrites the sewing-pattern program.Both variants start from the same coarse pattern and share the simulator, backbone, observations, tools, memory, and ten-simulation budget; only role decomposition differs.Source: §G.3 · pp. 39–40
  • After ten simulations, the single agent reaches IoU 0.820 versus 0.846 for the trio, with early plateauing and oscillation versus progressive convergence.The single agent's IoU is 0.812 after three simulations; its consecutive-step IoU changes average 0.063 versus 0.035 for the trio.Source: §G.3 · pp. 40–41, Figure 17 · p. 40, Table 14 · p. 40
  • Self-selection raises the single agent's IoU from 0.820 to 0.833, only halfway to the trio's 0.846, implicating its self-judgment rather than its evidence.The self-selection variant sees the same views and measurements as the Critic but differs in who scores each step.Source: §G.3 · p. 41
  • The single-agent loop emits 2.6× the output tokens and uses 1.8× the VLM time of the trio, yet needs more iterations for lower quality.The trio uses short, structured calls, reuses stable prompt prefixes through caching, and compacts older views into text notes.Source: §G.3 · p. 41

G.4 ABLATION ON TEXTURE GENERATION

The texture ablation shows that masking separates garments but does not correct missing or warped texture; the paper's method uses structured texture classes and extrapolation tools.

G.4 ABLATION ON TEXTURE GENERATION

  • Without instance masks, back-projection mixes garments, bleeding tee stripes onto pants and introducing residual skin and background along silhouettes.Figure 18 contrasts naive UV mapping without Grounded SAM against the masked variant.Source: §G.4 · p. 41
  • Masks separate garments, but fixed projection cannot extrapolate unseen cloth and maps the observed print onto flat panels in a warped form.The drape's folds and foreshortening are baked into the projected samples.Source: §G.4 · p. 41
  • The proposed texture method divides textures into five structured classes, each handled by an extrapolation tool, with the agent judging the class.The caption illustrates continuing stripes at their measured repeat, restoring a logo, and keeping pants uniformly black.Source: §G.4 · p. 41

H USER STUDY

  • The user study asks participants to choose the best geometry match and rate plausibility and texture from 1 to 5, with methods anonymized and shuffled.The questions use non-artistic participants for geometry selection and rating questions for plausibility and texture.Source: H USER STUDY · p. 41

I LIMITATIONS AND FUTURE WORK

The paper identifies three remaining boundaries: free-form ornament and cloth thickness are unsupported, while the full system prioritizes fidelity over API cost.

I LIMITATIONS AND FUTURE WORK

  • GarmentCode-XL cannot represent some free-form ornaments, such as a loose sash, although it expresses the vast majority of everyday outfits.Future work may let the agent write code to extend the representation using the input image as guidance.Source: I LIMITATIONS AND FUTURE WORK · p. 42, Figure 22 · p. 44
  • The current method does not support garments with thickness, such as padded down jackets.Future work may let the agent author thickness and material properties alongside the pattern.Source: I LIMITATIONS AND FUTURE WORK · p. 42
  • The full system prioritizes high-fidelity, simulation-ready generation over cost, with average API costs of $0.52, $0.75, $0.91, and $1.49 for one- to four-layer outfits.Cheaper or open-source models lower cost but also quality, reflecting a quality-cost trade-off.Source: I LIMITATIONS AND FUTURE WORK · p. 42

Related papers

Loading 2610.15582v1…