SLaMP: Scheduled Latent Mixing for Part Editing

• SLaMP is a diffusion-based local editing method that uses scheduled latent mixing with spatial masks to generate 2D anchor images for part-level 3D Gaussian editing.
• It employs a time-dependent blending strategy where edited latents are integrated while progressively reinserting original latent context outside the target mask.
• Empirical results show improved CLIP and B-VQA metrics, indicating enhanced locality, controllability, and alignment compared to standard editing methods.

Searching arXiv for the cited work and closely related papers on SLaMP, RoMaP, and latent mixing. Scheduled Latent Mixing and Part (SLaMP) is a diffusion-based local image-editing method introduced as a component of RoMaP for part-level editing in 3D Gaussian Splatting. Its function is to generate high-quality part-edited 2D anchor images by mixing original and edited latents over the denoising trajectory under a spatial mask, then to use those anchors in a masked $\mathcal{L}_1$ term within a regularized Score Distillation Sampling (SDS) objective. In RoMaP’s formulation, SLaMP is motivated by the ambiguity of SDS for precise local modification and by the need to preserve non-target context while applying strong edits to a specified part (Kim et al., 15 Jul 2025).

1. Origin within part-level 3D Gaussian editing

SLaMP arises in the specific context of part-level local editing of 3D Gaussian Splatting. RoMaP identifies two obstacles that make such editing difficult: inconsistent multi-view part segmentation and the ambiguity of SDS. The paper states that prior 3D editing methods are more effective for instance-level edits than for part-level edits, because part masks derived from multiple 2D views are unstable and because SDS alone does not provide a sufficiently explicit modification direction for localized changes. The reported failure modes include weak edits, edits that drift to semantically correlated but unintended regions, and difficulty with drastic or low-probability changes such as altering only one eye or changing a nose into an unusual object (Kim et al., 15 Jul 2025).

Within that setting, SLaMP is introduced to generate high-quality part-edited 2D images that confine modifications to the target region while preserving contextual coherence. These images serve as anchors for a masked $\mathcal{L}_1$ loss. This design converts an otherwise implicit text-guided optimization problem into one with an explicit view-aligned visual target. A plausible implication is that SLaMP reduces underdetermination in local 3D editing by supplying a concrete intermediate supervision signal rather than relying on SDS alone.

2. Core mechanism: scheduled latent mixing under part masks

The name “Scheduled Latent Mixing and Part” describes the two defining operations. “Scheduled latent mixing” denotes time-dependent blending over the denoising or rectified-flow trajectory rather than a fixed mixing ratio. “Part” denotes the fact that this blending is spatially restricted by a 2D mask $\mathcal{M}_{2D}$, obtained from RoMaP’s robust 3D segmentation and projected into the current rendered view (Kim et al., 15 Jul 2025).

The scheduling logic is asymmetric over time. Early in denoising, the blending coefficient $\mathcal{F}_t$ is kept low so that the model can generate new target-part content without strong interference from the original image. After a switch timestep $t_s$, $\mathcal{F}_t$ is increased sharply so that the non-target region is pulled back toward the original latent, preserving identity, structure, and surrounding context. The paper explicitly states that setting $t_s$ too low disrupts the original context, whereas setting it too high hinders new content generation; $t_s$ is selected at a point where SSIM is stable while CLIP$_{dir}$ remains high (Kim et al., 15 Jul 2025).

Operationally, SLaMP edits the target region while progressively reinserting original latent information outside the target mask. This is not a post hoc RGB blending method. The mixing occurs in-process during iterative latent denoising, which is central to its preservation behavior.

3. Mathematical formulation

RoMaP defines the regularized SDS objective containing SLaMP as

$$\mathcal{L}_{R\text{-}SDS} = \lambda_1 \hat{\mathcal{L}}_{SDS}(c^\phi_{pr}, p_{\text{edit}}) + \lambda_2 \hat{\mathcal{L}}_{1}(c^\phi_{pr}, \mathrm{SLaMP}(c^\phi_{pr})).$$

Here $\mathcal{L}_1$0 is the render from prior-removed Gaussians at view $\mathcal{L}_1$1, $\mathcal{L}_1$2 is the edit prompt, $\mathcal{L}_1$3 is a masked SDS term, and $\mathcal{L}_1$4 is a masked anchor loss against the SLaMP-edited image. The paper states that the hat notation denotes masked losses leveraging $\mathcal{L}_1$5 and $\mathcal{L}_1$6 to restrict changes to intended regions (Kim et al., 15 Jul 2025).

The defining SLaMP latent update is

$\mathcal{L}_1$7

This equation states that the non-target region $\mathcal{L}_1$8 is blended with the original latent using the time-dependent coefficient $\mathcal{L}_1$9, while the target region remains governed by the edited latent. The supplementary algorithm gives the same mechanism in masked form:

$\mathcal{M}_{2D}$0

$\mathcal{M}_{2D}$1

The schedule itself is piecewise:

$\mathcal{M}_{2D}$2

The extracted text does not provide exact numeric values for $\mathcal{M}_{2D}$3, $\mathcal{M}_{2D}$4, or $\mathcal{M}_{2D}$5, but it does specify that the switch is sharp and is chosen empirically using SSIM and CLIP$\mathcal{M}_{2D}$6 (Kim et al., 15 Jul 2025).

4. Function inside the RoMaP pipeline

RoMaP consists of two main components: robust 3D mask generation with 3D-GALP and regularized SDS for local 3D editing. SLaMP belongs entirely to the second component. It is not used for 3D segmentation or for constructing the 3D mask itself. Instead, 3D-GALP first produces a consistent 3D part mask $\mathcal{M}_{2D}$7, which is then projected into each training view as $\mathcal{M}_{2D}$8. SLaMP uses that projected mask to generate a view-specific edited anchor image, and the regularized SDS objective uses the anchor in the masked $\mathcal{M}_{2D}$9 term (Kim et al., 15 Jul 2025).

The per-view workflow is: render the current prior-removed 3D Gaussian scene; project the robust 3D part mask into a 2D target mask; invert or initialize the rendered image in latent space using a rectified-flow-based editing backend; run prompt-guided denoising while blending original latent information outside the target mask according to the schedule; decode the edited latent into a 2D anchor image; then compare the current render to that anchor under masked $\mathcal{F}_t$0 while also applying masked SDS. The paper further states that the implementation uses PyTorch, is based on Threestudio, uses Stable Diffusion 3, leverages rectified flow inversion, and is reported on a single A100 (Kim et al., 15 Jul 2025).

This division of labor is important. SLaMP is a 2D part editor that supervises 3D optimization; it is not itself the final 3D editing objective, nor is it the segmentation mechanism that determines parts.

5. Empirical behavior and reported effects

RoMaP’s ablations provide the most direct quantitative evidence for SLaMP. In Table 3, the baseline reports CLIP $\mathcal{F}_t$1 and CLIP$\mathcal{F}_t$2 $\mathcal{F}_t$3; adding the mask yields CLIP $\mathcal{F}_t$4 and CLIP$\mathcal{F}_t$5 $\mathcal{F}_t$6; adding mask plus $\mathcal{F}_t$7 yields CLIP $\mathcal{F}_t$8 and CLIP$\mathcal{F}_t$9 $t_s$0; and the full method yields CLIP $t_s$1 and CLIP$t_s$2 $t_s$3. Since the $t_s$4 term is enabled by SLaMP-generated anchors, the reported jump from $t_s$5 in CLIP and from $t_s$6 in CLIP$t_s$7 is the clearest quantitative indication of SLaMP’s effect on edit alignment (Kim et al., 15 Jul 2025).

The supplementary material also reports a dedicated 2D part-editing comparison against RePaint, InstructPix2Pix, SD3-inpainting, Plug-and-Play, and RF-inversion. In that comparison, SLaMP reports CLIP$t_s$8 $t_s$9 and B-VQA $\mathcal{F}_t$0, compared with RePaint $\mathcal{F}_t$1, InstructPix2Pix $\mathcal{F}_t$2, SD3-inpainting $\mathcal{F}_t$3, Plug-and-Play $\mathcal{F}_t$4, and RF-inversion $\mathcal{F}_t$5. At the full-system level, RoMaP reports CLIP$\mathcal{F}_t$6 $\mathcal{F}_t$7 versus a best baseline of $\mathcal{F}_t$8, and B-VQA $\mathcal{F}_t$9 versus a best baseline of $t_s$0. The user study further reports Alignment $t_s$1, Fidelity $t_s$2, and Accuracy $t_s$3, where the accuracy criterion asks whether only the specified parts were edited correctly (Kim et al., 15 Jul 2025).

These results support the paper’s interpretation that SLaMP improves locality, controllability, and the ability to realize drastic local edits while preserving context.

6. Relation to adjacent latent-mixing research

SLaMP is most directly comparable to later or adjacent methods that use latent-space control to balance editability and preservation, but its combination of scheduling and explicit part masking is more specialized. LAMS-Edit introduces “Latent and Attention Mixing with Schedulers” for diffusion-based real-image editing by storing the full inversion trajectory and mixing both latent states and attention maps with separate schedulers. It also supports localized editing with a user- or prompt-derived mask using SAM. However, LAMS-Edit is framed as a general 2D image and style editor built on Stable Diffusion and Prompt-to-Prompt, whereas SLaMP is embedded in a 3D Gaussian editing pipeline and uses masked latent reinsertion specifically to construct anchor images for regularized SDS (Fu et al., 6 Jan 2026).

Earlier work on global latent mixing in semi-supervised medical image classification uses interpolation in input and hidden representation space, with random selection from a predefined set of eligible layers. That method is relevant as a precursor in showing that latent-space interpolation can regularize learning more effectively than input-only mixing, but it does not introduce explicit temporal scheduling or part-wise spatial control (Gyawali et al., 2020).

Mimicry, a system for targeted deep learning boundary testing, performs layer-wise StyleGAN style-vector replacement guided by classifier confidence. Its control is structured and feature-selective, and it uses coarse, middle, and fine layer groupings as latent factors. Yet it does not use explicit scheduled mixing coefficients over denoising timesteps, nor does it provide spatial part masks in the strong sense used by SLaMP (Weißl et al., 2024).

This suggests that SLaMP occupies a specific intersection of three lines of work: scheduled latent control, spatially localized editing, and view-aligned supervision for downstream 3D optimization.

7. Scope, misconceptions, and limitations

A common misconception is to treat SLaMP as a standalone 3D editing framework. In RoMaP, it is more precisely a 2D part-editing module that produces anchor images for 3D optimization. Another misconception is to equate it with generic latent interpolation. Its defining operation is not uniform global blending, but masked, time-dependent reinsertion of original latent information outside the target region (Kim et al., 15 Jul 2025).

The “Part” in SLaMP should also be understood narrowly. It denotes mask-guided locality rather than a broader semantic part-decomposition model. The method depends on a projected 2D part mask derived from robust 3D segmentation; it does not independently infer semantic part structure within the diffusion model. This suggests that mask quality is structurally important to performance.

The paper states several limitations. Performance may degrade when geometry is highly complex, when attention-based 3D segmentation is poor, when the original 3DGS is blurry or badly reconstructed, or when Stable Diffusion 3 cannot correctly interpret the scene. A plausible implication is that SLaMP is sensitive to prompt clarity, inversion fidelity, and mask correctness in addition to the underlying generative prior (Kim et al., 15 Jul 2025).

In that scope, SLaMP is best characterized as a scheduled, mask-aware latent editing mechanism designed to make local 3D Gaussian editing less ambiguous. Its distinctive contribution is not merely latent mixing, but the use of scheduled latent reinsertion to create part-faithful anchor images that stabilize and localize regularized SDS.