Stable Score Distillation

• Stable Score Distillation is a stability-oriented design pattern that decomposes the traditional score distillation sampler to reduce noise and improve optimization.
• The method enhances high-quality 3D generation by employing a timestep-dependent orchestration of mode-disengaging and variance-reduced mode-seeking terms for improved detail and plausibility.
• SSD for text-guided editing preserves source structure and strengthens prompt alignment through null-text stabilization and cross-prompt guidance, achieving state-of-the-art results.

Stable Score Distillation (SSD) refers to score-distillation formulations that modify Score Distillation Sampling (SDS) so that a pretrained diffusion model can supervise optimization without the characteristic failure modes of vanilla SDS. In the literature, the name is used both narrowly—for the 3D-generation method introduced in “Stable Score Distillation for High-Quality 3D Generation” (Tang et al., 2023) and the text-guided editing framework introduced in “Stable Score Distillation” (Zhu et al., 12 Jul 2025)—and more broadly for a family of methods that reduce over-smoothness, implausibility, unstable editing trajectories, geometric inconsistency, or poor prompt alignment in diffusion-guided optimization (Lukoianov et al., 2024, Wang et al., 2023, Kwak et al., 2024). This usage suggests that SSD is best understood not as a single canonical algorithm, but as a stability-oriented design pattern for score distillation.

1. Scope and historical position

SDS, popularized in DreamFusion-style text-to-3D optimization, uses a pretrained text-to-image diffusion model to provide gradients for a target parameterization such as a NeRF, a 3D Gaussian field, or an editable image. The core attraction of SDS is that it transfers a powerful 2D generative prior into downstream optimization without retraining the diffusion backbone. Its main weakness is that the resulting gradients are often noisy, semantically brittle, and poorly aligned with geometric or source-preservation constraints (Tang et al., 2023, Lukoianov et al., 2024).

Within that setting, the 2023 SSD paper addresses high-quality 3D generation by decomposing the SDS estimator into functional components and re-orchestrating them across diffusion timesteps (Tang et al., 2023). The 2025 SSD paper addresses text-guided 2D and 3D editing by anchoring editing to the source prompt, adding a null-text reconstruction branch, and explicitly strengthening prompt alignment for difficult edits such as style transfer (Zhu et al., 12 Jul 2025). Both works retain the basic premise of score distillation—optimizing a target representation under a frozen diffusion prior—but they differ in objective structure, intended task, and failure mode.

A broader research thread extends the same stabilizing impulse. Some methods reinterpret SDS as a high-variance approximation to a diffusion trajectory and replace the noise term accordingly (Lukoianov et al., 2024). Others treat SDS as a Monte Carlo estimator and introduce control variates for variance reduction (Wang et al., 2023), enforce multiview consistency through geometry-aware noising and gradient matching (Kwak et al., 2024), or separate shape and texture into multi-objective guidance terms (Xu et al., 12 Nov 2025). In that wider sense, SSD names both specific methods and a research agenda.

2. SDS baseline and the origin of instability

The standard SDS update used in text-to-3D can be written as

$$\nabla_\theta \mathcal{L}_{\text{SDS}} = w(t)\Big(\hat{\epsilon}_\phi^{\text{CFG}}(x_t;y,t)-\epsilon\Big)\frac{\partial x}{\partial\theta},$$

where $x=g(\theta,c)$ is the rendered image, $x_t$ is its noisy version, $\epsilon$ is Gaussian noise, and $\hat{\epsilon}_\phi^{\text{CFG}}$ is the classifier-free guided noise prediction of the frozen diffusion model (Tang et al., 2023). In practice, this simple estimator couples several effects that are beneficial in some regimes and harmful in others.

The 2023 SSD analysis decomposes the SDS estimator into three components: a mode-disengaging term, a mode-seeking term, and a variance-reducing term (Tang et al., 2023). In that decomposition, the classifier-free guidance difference

$$\hat{\epsilon}_\phi(x_t,t,y)-\hat{\epsilon}_\phi(x_t,t,\emptyset)$$

acts as a mode-disengaging term; the conditional prediction $\hat{\epsilon}_\phi(x_t,t,y)$ acts as a mode-seeking term; and the subtraction of $\epsilon$ acts as a crude variance reducer. The paper argues that over-smoothness arises from the mode-seeking term at large timesteps, where transient modes appear between conditional modes, while implausible colors and textures arise when the mode-disengaging term dominates at small timesteps (Tang et al., 2023).

Other analyses identify complementary sources of instability. “Score Distillation via Reparametrized DDIM” shows that, after a change of variables, SDS resembles a high-variance form of DDIM in which the correct trajectory-consistent noise is replaced by fresh i.i.d. Gaussian noise at each step; this breaks the underlying denoising trajectory and contributes to over-smoothing and unrealistic outputs (Lukoianov et al., 2024). “SteinDreamer” interprets SDS and VSD as Monte Carlo gradient estimators with high variance and frames variance reduction as the central problem (Wang et al., 2023). “Geometry-Aware Score Distillation” traces Janus artifacts to multiview inconsistency of 2D scores predicted from different viewpoints (Kwak et al., 2024). “Target-Balanced Score Distillation” identifies a separate instability in negative-prompt methods: target negative prompts improve local texture supervision but can weaken target semantics and distort shape (Xu et al., 12 Nov 2025).

Taken together, these analyses establish a common picture. Vanilla SDS is not merely noisy in an informal sense; it entangles prompt steering, mode selection, variance control, and geometry formation in a single estimator whose components are reliable on different parts of the diffusion trajectory and unreliable elsewhere.

3. SSD for high-quality 3D generation

“Stable Score Distillation for High-Quality 3D Generation” reformulates SSD as a timestep-dependent orchestration of the SDS components rather than a wholesale replacement of the score-distillation framework (Tang et al., 2023). Its central claim is that the mode-disengaging term is useful at large timesteps because it helps fast geometry formation and trap escaping, whereas a variance-reduced mode-seeking term is preferable at small timesteps because it refines details while staying closer to high-density conditional image modes.

The paper defines the mode-disengaging term as

$$h(x_t,t,y)=\hat{\epsilon}_\phi(x_t,t,y)-\hat{\epsilon}_\phi(x_t,t,\emptyset),$$

and replaces the fixed subtraction of $\epsilon$ with an adaptive variance-reduction coefficient

$x=g(\theta,c)$0

which yields the variance-reduced estimator

$x=g(\theta,c)$1

For large timesteps $x=g(\theta,c)$2, SSD uses only $x=g(\theta,c)$3. For small timesteps $x=g(\theta,c)$4, it uses a norm-matched version of $x=g(\theta,c)$5 so that the scale of the refinement term is compatible with the scale of the disengaging term (Tang et al., 2023).

This design directly targets the two diagnosed pathologies. At large timesteps it avoids transient-mode averaging by suppressing the raw mode-seeking term. At small timesteps it suppresses the tendency of the mode-disengaging ratio to drive the optimization away from natural-image modes, while also correcting the scale and direction mismatch created by subtracting raw Gaussian noise (Tang et al., 2023). The method is presented as simple, compatible with various 3D generation frameworks and 3D representations, and designed to be dropped into standard score-distillation pipelines.

The reported empirical effect is an increase in both plausibility and detail. In a user study over 100 prompts and 1000 total comparisons, SSD was preferred 53.2% of the time for plausibility and 61.4% of the time for finer details relative to DreamFusion, Fantasia3D, Magic3D, and ProlificDreamer (Tang et al., 2023). The article’s broader significance lies in its analytic decomposition: it makes explicit that “stability” in SSD can mean correct regime selection across timesteps, not merely smaller gradient norms.

4. SSD for text-guided 2D and 3D editing

The later editing-oriented method titled “Stable Score Distillation” addresses a different problem: how to modify an existing image or 3D scene according to a target prompt while preserving source structure, unedited regions, and editing coherence (Zhu et al., 12 Jul 2025). The paper attributes the failure of prior score-distillation editing methods such as DDS and CSD to complex auxiliary structures, conflicting optimization signals, and limited explicit source preservation.

Its starting point is a single classifier-free-guidance classifier anchored to the source prompt $x=g(\theta,c)$6. The core SSD score is written as

$x=g(\theta,c)$7

where $x=g(\theta,c)$8 is the current noisy latent, $x=g(\theta,c)$9 is the source latent, $x_t$0 is the target prompt, and $x_t$1 denotes null text (Zhu et al., 12 Jul 2025). The paper then decomposes this into a cross-prompt term and a cross-trajectory term,

$x_t$2

and adds a prompt-enhancement branch

$x_t$3

plus an image-space identity regularizer

$x_t$4

The final objective is

$x_t$5

The stated function of this architecture is precise. The cross-prompt term steers the current latent from the source prompt manifold toward the target prompt manifold. The cross-trajectory term ties the current source-conditioned prediction to the unconditional prediction at the source latent, thereby anchoring the optimization to the original content. The null-text branch serves as a constant reference that stabilizes optimization and preserves surrounding regions. The prompt-enhancement branch restores editing strength, especially for style transformations, which DDS-like differences often underperform (Zhu et al., 12 Jul 2025).

This editing SSD reports state-of-the-art 3D editing results in its benchmark. On the 3D editing evaluation in the paper, it reaches CLIP Similarity $x_t$6, CLIP Directional Similarity $x_t$7, and a user-study preference of $x_t$8, outperforming IN2N, DDS, GaussianEditor, and DGE (Zhu et al., 12 Jul 2025). On PIE-Bench for 2D image editing, SSD reaches CLIP image-text similarity $x_t$9, while SSD+CDS reaches structure distance $\epsilon$0 (Zhu et al., 12 Jul 2025). The paper also reports fast convergence with non-increasing timestep sampling, around 3000 iterations for NeRF and 1500 for 3DGS (Zhu et al., 12 Jul 2025). In encyclopedic terms, this version of SSD shifts the notion of stability from variance control alone to source-faithful editing trajectories.

5. Stabilization mechanisms in the wider score-distillation literature

A substantial body of adjacent work broadens the technical meaning of stability in score distillation. One line focuses on variance and trajectory consistency. “SteinDreamer” recasts SDS and VSD as control-variate estimators and introduces Stein Score Distillation, where Stein’s identity supplies zero-mean control variates built from arbitrary baseline functions; instantiated with MiDAS depth or normal priors, the method reports lower gradient variance and 14–22% fewer steps to reach a target CLIP distance, while being about 30% faster per iteration than VSD because it avoids a second diffusion model (Wang et al., 2023). “Score Distillation via Reparametrized DDIM” replaces SDS’s fresh i.i.d. noise with DDIM-inverted, prompt-conditioned noise, making the optimization much closer to a deterministic DDIM trajectory and reducing over-smoothing without training additional networks or requiring multiview supervision (Lukoianov et al., 2024).

A second line emphasizes geometry and cross-view consistency. “Geometry-Aware Score Distillation” introduces 3D consistent noising, geometry-based gradient warping, and a gradient consistency loss, with the explicit aim of making 2D score predictions more consistent across viewpoints and thereby reducing Janus artifacts at minimal computation cost (Kwak et al., 2024). “Repulsive Latent Score Distillation for Solving Inverse Problems” interprets SDS through a Wasserstein gradient-flow lens and adds pairwise kernel-based repulsion to prevent particle collapse, while an augmented latent–pixel variational distribution addresses latent ambiguity in Stable Diffusion inverse problems (Zilberstein et al., 2024). These methods treat stability as consistency across views or across posterior modes rather than only across timesteps.

A third line separates competing objectives instead of smoothing a single estimator. “Target-Balanced Score Distillation” identifies a structural trade-off between texture fidelity and shape accuracy when target negative prompts are used. It decomposes guidance into a shape term and a texture term, then combines them with MGDA and a time-dependent weighting schedule. On 43 prompts, it reports the highest CLIP score, $\epsilon$1, and a user preference of $\epsilon$2 among compared methods (Xu et al., 12 Nov 2025). “RewardSDS” and “RewardVSD” reweight noise samples by scores from a reward model, so that aligned high-reward samples contribute more strongly to the gradient; the paper shows improvements over SDS and VSD on text-to-image, 2D editing, and text-to-3D (Chachy et al., 12 Mar 2025). “Mean-Shift Distillation” takes a different route again, replacing SDS with a mean-shift mode-seeking proxy whose extrema are aligned with modes of the smoothed diffusion output distribution and presenting it as a drop-in replacement for SDS (Thamizharasan et al., 21 Feb 2025).

A fourth line concerns distilling the diffusion model itself rather than using it as a frozen optimization prior. “Guided Score identity Distillation for Data-Free One-Step Text-to-Image Generation” adapts Score identity Distillation to Stable Diffusion and introduces Long and Short Guidance for teacher and fake-score conditioning. In the data-free setting, the one-step distillation of Stable Diffusion 1.5 reaches FID $\epsilon$3 on COCO-2014 validation (Zhou et al., 2024). “Denoising Score Distillation” studies an especially harsh regime in which the teacher diffusion model is pretrained only on corrupted data, then distilled into a one-step generator; in a linear analysis it shows that Fisher-divergence-based distillation identifies the eigenspace of the clean covariance and can yield a generator closer to the clean distribution than the noisy teacher (Chen et al., 10 Mar 2025). These papers sit adjacent to SSD rather than inside its narrowest definition, but they reinforce the same theme: stability often emerges from restructured objectives, better-conditioned estimators, or more faithful trajectories.

6. Acronym collisions and conceptual boundaries

The acronym “SSD” is overloaded in the recent literature, and not every “SSD” paper concerns Stable Score Distillation in the diffusion-optimization sense. The distinction is technically important.

Name	Domain	Distinguishing idea
Stable Score Distillation (Tang et al., 2023)	Text-to-3D generation	Timestep-dependent orchestration of mode-disengaging and variance-reduced mode-seeking terms
Stable Score Distillation (Zhu et al., 12 Jul 2025)	Text-guided 2D and 3D editing	Single source-anchored classifier, null-text stabilization branch, prompt enhancement
Self Score Distillation (Liu et al., 2023)	3D head generation	Two CFG settings of one landmark-guided ControlNet supervise a 3D head model
Stein Score Distillation (Wang et al., 2023)	Text-to-3D generation	Stein-identity control variates reduce score-distillation variance
Guided Score identity Distillation (Zhou et al., 2024)	One-step text-to-image distillation	Data-free teacher–student distillation with Long and Short Guidance
Gen-SSD (He et al., 3 Apr 2026)	Chain-of-thought distillation	“SSD” means Generation-time Self-Selection Distillation, unrelated to diffusion score distillation

This terminological spread has two consequences. First, “Stable Score Distillation” in vision generally refers either to the 2023 estimator redesign for high-quality 3D generation (Tang et al., 2023) or the 2025 source-anchored editing framework (Zhu et al., 12 Jul 2025). Second, in broader discussion the phrase often functions as a shorthand for stability-oriented score distillation more generally, encompassing variance reduction, noise-path consistency, geometry-aware regularization, negative-prompt balancing, reward weighting, and multimodal posterior approximations (Wang et al., 2023, Lukoianov et al., 2024, Kwak et al., 2024, Xu et al., 12 Nov 2025, Chachy et al., 12 Mar 2025, Zilberstein et al., 2024). The most accurate encyclopedic reading is therefore task-dependent: SSD is simultaneously a named method, a family resemblance among SDS variants, and an overloaded acronym whose meaning must be fixed by context.