RADmesh: Text-Driven Mesh Deformation with Intelligent Remeshing

By Prahlad Menon 3 min read

Most text-to-3D methods either generate meshes from scratch or deform existing ones with visual losses. The problem? Drastic deformations without changing the triangulation quickly produce ugly, stretched triangles — even when large geometry changes are exactly what you want semantically.

RADmesh (Remesh-Aware Mesh Deformation) from UChicago, USC, and Technion solves this elegantly: it couples deformation with periodic isotropic remeshing, letting you grow large, detailed geometry from text prompts while keeping triangles neat and efficient.

The Core Insight

Previous work like Geometry in Style used dARAP (differentiable as-rigid-as-possible) deformations for text-driven shape editing. But these methods keep the triangulation fixed — fine for subtle changes, catastrophic for growing wings on a dog or adding horns to a head.

RADmesh’s key innovations:

  1. Extended dARAP with scaling — A 6-vector per vertex (3 for direction, 3 for scale) provides enough degrees of freedom for large deformations while staying controllable
  2. Periodic remeshing with state interpolation — Every 100 epochs, remesh using Botsch-Kobbelt method and interpolate the optimizer state onto new vertices
  3. Coarse-to-fine resolution schedule — Start with coarse triangles, progressively refine as geometry emerges

How It Works

Input: Base mesh + text prompt + (optional) vertex selection region
↓
Deformation optimization (dARAP with scale)
↓
Every 100 epochs: Remesh → Interpolate optimizer state → Continue
↓
Output: Deformed mesh with adaptive triangulation

The remeshing step is crucial. Without it, growing a wing from a small region creates stretched, unusable triangles. With coarse-to-fine remeshing, the triangulation naturally adapts to the emerging geometry.

Localized Deformations

One standout feature: RADmesh supports region-based editing. Select a small region on a mesh (say, a shoulder), give a text prompt (“eagle wing”), and grow detailed geometry from just that region. The rest of the mesh stays completely unchanged — geometry, triangulation, even UV coordinates and textures are preserved.

This enables iterative workflows: grow arms, then head, then accessories, building up complex characters from a simple base.

Visual Supervision with CSD

RADmesh uses Cascaded Score Distillation (CSD), an SDS variant, for text-to-geometry guidance via DeepFloyd IF. The vertex-based deformation representation is specifically designed to be robust to the noisy gradients these visual losses produce — a notorious challenge in generative mesh optimization.

Results

The paper shows impressive examples:

  • Growing symmetrical wings on a dog (Spot)
  • Adding detailed antlers to deer
  • “Detailizing” simple shapes (sphere → pineapple)
  • Iterative builds: torso → +arms → +head → +accessories

Each result has clean, isotropic triangles adapted to the generated geometry — not just a uniform upsample of the original mesh.

Code and Resources

The code includes example configs for both localized and whole-mesh deformations, with detailed hyperparameter guidance for different mesh types.

Why This Matters

RADmesh solves a real pain point in generative 3D: getting high-quality deformed geometry without manual retopology. For game artists, VFX teams, and anyone working with procedural mesh generation, this is a significant workflow improvement.

The coarse-to-fine remeshing insight is also generalizable — it should benefit other mesh optimization methods that currently suffer from fixed-topology limitations.

The acceptance as an ECCV 2026 Oral signals the community recognizes this as important work. With MIT-licensed code available, expect to see this technique integrated into creative tools soon.