GeoDrag: Diffusion-Based Drag Editor

• GeoDrag is a training-free, diffusion-based drag editor that precisely moves 2D handle-points in latent space using projected 3D priors.
• It incorporates a point fixation mechanism and copy-and-paste refinement to lock in edits and minimize drift during iterative optimization.
• Quantitative evaluations show that GeoDrag improves both accuracy and speed over previous methods, making it valuable for automotive, furniture, and architectural design workflows.

GeoDrag is a training-free, diffusion-based drag editor introduced as the point-based image-editing module at the core of GeoDiffusion. Its function is to move a set of 2D handle-points from source to target locations with high geometric accuracy and minimal collateral damage to the rest of the image. Within GeoDiffusion, it operates after 3D prior projection and style transfer: the framework first uses a class-specific 3D object as a geometric prior to define keypoints and parametric correlations in 3D space, ensures viewpoint consistency through a rendered image of a reference 3D object, and then hands the styled latent together with projected source–target keypoint pairs to GeoDrag for precise latent-space editing (Mueller et al., 25 Oct 2025).

1. Position within the GeoDiffusion framework

GeoDrag is not the entirety of GeoDiffusion but the drag-based editing submodule that realizes the final geometric modification in image space. The surrounding framework is designed for accurate 3D geometric conditioning in image generation, especially in iterative design workflows where object features must be controlled precisely. Traditional 3D editing approaches are described as time-consuming and demanding specialized skills, while image-based generative methods are characterized as lacking accuracy in geometric conditioning; GeoDrag addresses the last-mile editing problem by converting projected geometric intent into localized latent updates (Mueller et al., 25 Oct 2025).

The module inherits the alternating “drag + denoise” structure of GoodDrag, but introduces two specific mechanisms aimed at accuracy and convergence. The first is a Point Fixation Mechanism, in which a handle-point that has reached its target within a small threshold is excluded from further gradient updates so that it cannot drift. The second is Copy-and-Paste Refinement, in which, once a point is very close, its latent patch is copied to the exact target patch and the source patch is blurred, sharply closing any residual gap and speeding convergence. In the GeoDiffusion formulation, these two additions are the decisive modifications that distinguish GeoDrag from earlier latent drag editors (Mueller et al., 25 Oct 2025).

2. Optimization objective and point fixation

GeoDrag is formulated over diffusion latents. Let $z_t$ denote the latent at diffusion timestep $t$, and let $\{p_i^k\}$ be the 2D handle-point locations after $k$ gradient steps within timestep $t$, with targets $\{g_i\}$. GoodDrag’s core feature-alignment loss is given as

$$L_{\text{feature}}(z_t^k; \{p_i^k\}) = \sum_i \left\|F_{\Omega(p_i^k + \beta \cdot d_i^k, r_1)}(z_t^k) - sg[F_{\Omega(p_i^0, r_1)}(z_t^0)]\right\|_1 + \lambda \left\|(z_{t-1}^k - sg(z_{t-1}^0)) \odot (1-M)\right\|_1,$$

where $F_{\Omega(\cdot,r)}$ extracts a square patch of radius $r$ around a point in the U-Net feature map, $d_i^k = g_i - p_i^k$ is the residual direction, $t$0 is the step-size, $t$1 denotes stop-gradient, and $t$2 masks out the editable region so background stays fixed (Mueller et al., 25 Oct 2025).

GeoDrag modifies this objective through point fixation. For each point, the error is

$t$3

A lower threshold $t$4 defines entry into the ignored set $t$5: once $t$6, the point is treated as fixated. An upper threshold $t$7 determines when a previously fixated point rejoins optimization if it drifts beyond $t$8. Around each fixated point, GeoDrag builds a gradient-mask $t$9 that zeros out gradients in a small patch of radius $\{p_i^k\}$0. The resulting motion-supervision loss becomes

$\{p_i^k\}$1

This construction makes the editing dynamics asymmetric: active points continue to be optimized, while already-satisfied points are explicitly protected from further motion. A plausible implication is that GeoDrag converts a multi-handle optimization problem from a globally coupled update into a partially frozen one, thereby reducing destructive interactions among points that have already converged (Mueller et al., 25 Oct 2025).

3. Copy-and-paste refinement and latent-space locking

At the final gradient-based step of each timestep, GeoDrag applies a direct latent relocation to each fixated point. For every $\{p_i^k\}$2, let $\{p_i^k\}$3 and $\{p_i^k\}$4. The refinement is

$\{p_i^k\}$5

Here $\{p_i^k\}$6 optionally amplifies the copied latent patch, and the source patch is blurred by noise injection. During the subsequent denoising steps for the remaining $\{p_i^k\}$7 timesteps, GeoDrag also forcibly copies the original unedited latent patches $\{p_i^k\}$8 into $\{p_i^k\}$9 to guarantee exact placement (Mueller et al., 25 Oct 2025).

The timestep-level pipeline is correspondingly structured. Within each diffusion timestep $k$0, GeoDrag performs $k$1 iterations of latent optimization with the modified loss and gradient-mask, tracks each point’s new location via local feature matching, fixates points that are sufficiently close, copies latent patches for all fixated points to the exact target while blurring their sources, denoises one step, and then re-applies copy-paste for the remaining timesteps to guarantee no drift. The point tracker itself updates handle locations by local feature matching in a neighborhood of radius $k$2 around the current point, seeking the best match to the source feature patch (Mueller et al., 25 Oct 2025).

This pipeline clarifies a central property of the method: GeoDrag is training-free, but it is not one-step. Its behavior depends on iterative latent optimization within each diffusion timestep, followed by denoising and recurrent locking. The method’s efficiency therefore comes from tighter convergence and reduced re-editing of satisfied points rather than from eliminating the diffusion-time iterative loop.

4. Quantitative performance

The reported evaluations use 512×512 images. On the Geometry Guidance Dataset of 70 test cases covering cars, bikes, and airplanes, GeoDrag attains the best mean pixel distance of final handles to true targets, while also improving latency over GoodDrag. On DragBench, which measures general point-based editing rather than only geometry-guidance, GeoDrag again achieves the best mean distance while preserving essentially the same image fidelity. For non-vehicle examples such as chairs and buildings, the paper reports substantially larger gains in both accuracy and speed (Mueller et al., 25 Oct 2025).

Benchmark	GeoDrag	Reported comparison
Geometry Guidance Dataset	MD = 10.80 px, CLIP = 25.31, HPSv2 = 25.71, Time = 60.0 s	GoodDrag: MD = 11.77 px, Time = 66.3 s
DragBench	MD = 20.86 px, IF = 0.872	GoodDrag: MD = 24.29 px, IF = 0.871

The paper states that, on the Geometry Guidance Dataset, GeoDrag is “+8.2 % more accurate, –9.5 % faster than GoodDrag,” and on DragBench it is “+14.1 % more accurate, same image fidelity.” For non-vehicle examples, GeoDrag is reported to yield “+32.7 % accuracy, +39.9 % speed over GoodDrag, with small gains in CLIP/HPSv2.” A “fast” configuration reaches approximately $k$3 seconds per $k$4 edit, described as “38.5 % faster than GoodDrag” (Mueller et al., 25 Oct 2025).

These numbers position GeoDrag as a benchmarked solution for geometry-sensitive point dragging rather than a purely qualitative UI technique. The use of MD, IF, CLIP-score, and HPSv2 also indicates that evaluation is split between geometric endpoint accuracy, image fidelity, prompt alignment, and human-preference image quality.

5. Relation to prior and contemporaneous drag-based editors

Within its immediate lineage, GeoDrag is framed as an extension of GoodDrag. GoodDrag introduced alternating drag/denoise, but lacked both a “lock-in” mechanism for reached points and a patch-copy step. GeoDrag’s main claim over that baseline is that fixation prevents drift after target arrival and copy-paste closes the residual gap that gradient descent often leaves behind. The paper further states that DragDiffusion, DragonDiffusion, SDE-Drag, FreeDrag, EasyDrag, and DragNoise all lack a fixation mechanism and direct copy-paste, so they either converge slowly or overshoot or undershoot multiple points (Mueller et al., 25 Oct 2025).

The name GeoDrag is also used by a distinct later method, “Dragging with Geometry: From Pixels to Geometry-Guided Image Editing,” which is not the GeoDiffusion module. That work builds a unified displacement field from a geometry-aware field $k$5 based on monocular depth and a plane-aware field $k$6 based on pixel-distance decay, then integrates conflict-free Voronoi partitioning for multi-point drags. It operates in a one-step latent-relocation pipeline using Stable Diffusion 1.5, a Latent Consistency Model U-Net for 10-step inversion and generation, and Depth Anything V2. On DragBench it reports MD $k$7, IF $k$8, runtime $k$9 seconds, and memory $t$0 GB, with user study ranking best in $t$1 of cases (Pu et al., 30 Sep 2025).

Related work also extends drag editing into explicitly 3D representations. “Drag Your Gaussian” introduces a 3D drag-based editing method for 3D Gaussian Splatting using 3D masks, pairs of 3D control points, a multi-resolution triplane positional encoder, a region-specific positional decoder, and a Drag-SDS loss that combines latent-space SDS, image-space SDS, and LoRA regularization; its reported runtime is ten minutes per edit (Qu et al., 30 Jan 2025). “3DGS-Drag” likewise performs point-based 3D editing on 3D Gaussian Splatting via deformation guidance, diffusion guidance, and a progressive multi-step editing schedule, with reported editing time of 10 to 20 minutes on a single RTX 4090 GPU (Dong et al., 12 Jan 2026). Together, these papers indicate that “GeoDrag” denotes a family resemblance—geometry-aware dragging under point constraints—rather than a single canonical algorithm.

6. Practical controls, use cases, and operating regime

GeoDrag exposes a compact set of tunables. The main controls are $t$2, the number of gradient steps per timestep, typically exemplified as $t$3–$t$4; $t$5, the feature-step size; $t$6, the background-consistency weight; the lower and upper fixation thresholds $t$7 and $t$8 in pixels; the radii $t$9 for the feature patch, $\{g_i\}$0 for the gradient-mask, and $\{g_i\}$1 for the copy-paste patch; and the copy-paste amplification $\{g_i\}$2, with default $\{g_i\}$3–$\{g_i\}$4. The paper notes that decreasing $\{g_i\}$5 leads to more accurate but slower behavior, and that the blur parameter $\{g_i\}$6 trades off residual ghosting versus speed (Mueller et al., 25 Oct 2025).

The module is presented as applicable to both geometry-guided design and general point-based retouching. In the GeoDiffusion car editor, wheelbase or overall length can be adjusted by projecting 3D keypoints and then using GeoDrag to move 2D wheels and bumpers exactly. Other cited use-cases include furniture design, such as dragging chair legs or back supports, and architectural retouching, such as moving a building’s roof ridge or aligning architectural features without retraining (Mueller et al., 25 Oct 2025).

A common misconception is that training-free point-based editing implies weak geometric control. GeoDrag is explicitly positioned against that assumption: its geometric precision is obtained not through additional model training but through latent optimization guided by projected source–target keypoint pairs, fixation masks, and forced latent relocation. A second misconception is that all methods named GeoDrag are interchangeable. The literature instead contains at least two distinct usages: GeoDrag as the GeoDiffusion module for accurate latent-space dragging under projected 3D priors (Mueller et al., 25 Oct 2025), and GeoDrag as a standalone geometry-guided one-step image editor based on analytic displacement fields and depth cues (Pu et al., 30 Sep 2025).