---
title: Synchronized Decoupled Guidance (SDG)
url: https://www.emergentmind.com/topics/synchronized-decoupled-guidance-sdg
type: topic
---

# Synchronized Decoupled Guidance (SDG)

Searching arXiv for recent papers on Synchronized Decoupled Guidance and related decoupled guidance methods.
Synchronized Decoupled Guidance (SDG) is an inference-time guidance mechanism for diffusion-based video generation that aims to improve physical plausibility without retraining the generator. In the formulation introduced in "Enhancing Physical Plausibility in Video Generation by Reasoning the Implausibility" [2509.24702], SDG operates after a physics-aware reasoning step constructs a counterfactual prompt that deliberately violates the governing physical law implied by the original prompt. The method then suppresses the implausible alternative during denoising by combining two designs—**synchronized directional normalization** and **trajectory-decoupled denoising**—so that physically implausible content is counteracted both early and consistently throughout sampling.

## 1. Concept and scope

SDG is defined in the context of a broader physical-plausibility problem in text-to-video diffusion models. The underlying claim is that contemporary video generators can produce highly photorealistic outputs while still violating basic physical commonsense, including cases where objects move without appropriate forces, fluids ignore gravity, phase transitions occur unrealistically, or object-surface interactions lack proper contact or deformation [2509.24702]. The method therefore targets physical fidelity rather than generic perceptual quality.

The key design choice is not to retrain the model to encode physics more explicitly, but instead to reason about **implausibility** and guide the model away from it at inference time. The full pipeline has two stages. First, a physics-aware reasoning module infers entities, interactions, environmental conditions, temporal evolution, and the governing physical principle from the user prompt, and then constructs a counterfactual prompt that preserves the same scene and entities while violating the inferred law. Second, SDG uses that counterfactual prompt as a negative branch during denoising and modifies the guidance rule so that the implausible branch is suppressed immediately and without inheriting bias from the positive branch [2509.24702].

A central point in the paper is that SDG is not equivalent to ordinary negative prompting. Negative prompting is treated as an insufficient baseline because its suppressive effect is delayed and its negative branch is evaluated on a trajectory already shaped by the positive condition. SDG is introduced specifically to address those two failure modes.

## 2. Physics-aware counterfactual prompting

The input to SDG is not an arbitrary negative prompt. It is the output of a **Physics-aware reasoning (PAR)** stage that uses an LLM to infer the relevant physical law behind the prompt and then construct a counterfactual prompt that deliberately violates that law [2509.24702]. This counterfactual is intended to be targeted, structured, and visually plausible rather than generic.

The paper describes PAR as extracting five elements from the original prompt: entities, interactions, environmental conditions, temporal evolution, and the governing physical principle. The counterfactual prompt is then constructed so that the same scene content is maintained while the underlying physical law is contradicted. This design matters because the subsequent guidance step depends on contrasting a physically plausible trajectory with a physically implausible but semantically related alternative. A plausible implication is that SDG’s effectiveness depends materially on the quality and specificity of the counterfactual prompt, because the guidance signal is only as informative as the contrast it encodes.

This reasoning-first design distinguishes SDG from methods that use negative prompts in a generic way. The paper’s ablation results support that distinction: removing physics-aware reasoning reduces performance, with Wan2.1-14B on PhyGenBench dropping from **0.50** for the full model to **0.47** without physics-aware reasoning [2509.24702]. The authors also note qualitatively that without PAR the counterfactual can violate the wrong thing, introducing irrelevant or nonsensical changes instead of targeting the intended physical process.

## 3. Mathematical formulation

The paper situates SDG within standard diffusion and classifier-free guidance (CFG) notation. Negative prompting is written as
\[
\hat{\epsilon}_t \leftarrow \epsilon_{\theta}(x_t, c(p_+), t) + w \cdot \big(\epsilon_{\theta}(x_t, c(p_+), t) - \epsilon_{\theta}(x_t, c(p_-), t)\big),
\]
with discrepancy
\[
\Delta_t = \epsilon_{\theta}(x_t, c(p_+), t)-\epsilon_{\theta}(x_t, c(p_-), t).
\]
The paper then rewrites negative prompting as
\[
\hat{\epsilon}_t \leftarrow \epsilon_{\theta}(x_t, c(p_+), t) + w \cdot \Delta_t.
\]
This establishes the baseline against which SDG is defined [2509.24702].

The first SDG component, **synchronized directional normalization**, replaces raw discrepancy magnitude with a normalized directional correction:
\[
\hat{\epsilon}_t \leftarrow \epsilon_{\theta}(x_t, c(p_+), t) + \lambda \cdot \frac{\epsilon_{\theta}(x_t, c(p_+), t) - \epsilon_{\theta}(x_t, c(p_-), t)}{||\epsilon_{\theta}(x_t, c(p_+), t)-\epsilon_{\theta}(x_t, c(p_-), t)||+\varepsilon}.
\]
Here, \(\lambda\) controls the strength and \(\varepsilon\) ensures numerical stability [2509.24702].

The second component, **trajectory-decoupled denoising**, evolves two separate latent trajectories,
\[
x_t^+ \quad \text{for } p_+,\qquad x_t^- \quad \text{for } p_-,
\]
rather than evaluating both prompts on the same latent state. Their branch-specific predictions are
\[
\epsilon^+ = \epsilon_{\theta}(x_t^+, c(p_+), t) + w \cdot (\epsilon_{\theta}(x_t^+, c(p_+), t)-\epsilon_{\theta}(x_t^+, \emptyset, t)),
\]
\[
\epsilon^- = \epsilon_{\theta}(x_t^-, c(p_-), t) + w \cdot (\epsilon_{\theta}(x_t^-, c(p_-), t)-\epsilon_{\theta}(x_t^-, \emptyset, t)).
\]
The combined correction is
\[
\hat{\epsilon}^+ = \epsilon^+ + \lambda \cdot \frac{\epsilon^+-\epsilon^-}{||\epsilon^+-\epsilon^-||+\varepsilon}.
\]
Once \(\hat{\epsilon}\) is obtained, it is inserted into the standard sampler update
\[
x_{t-1} = \alpha_t x_t + \beta_t \hat{\epsilon}_t + \eta_t,
\]
which makes SDG plug-and-play with standard diffusion sampling [2509.24702].

## 4. Failure modes addressed by SDG

The paper motivates SDG by identifying two specific weaknesses of negative prompting: **lagged suppression effect** and **cumulative trajectory bias** [2509.24702]. Lagged suppression refers to the observation that the negative prompt exerts only weak influence in early denoising, when coarse scene structure is formed. As a result, implausible physical structure can emerge first and only be corrected later, if at all.

The paper analyzes this using the discrepancy \(\Delta_t\) and a Jacobian-based latent-direction view. It defines
\[
J_t = \frac{\partial \epsilon_{\theta}(x_t, c, t)}{\partial x_t},
\]
and gives an eigen-direction expression for suppressive influence,
\[
\text{suppression}_{t}^{(-)} \propto v_{l,t}^{\top}(-w \Delta_t) = -w \langle v_{l,t}, \Delta_t \rangle = -w ||v_{l,t}||\cdot||\Delta_t||\cdot \cos(\theta),
\]
where \(v_{l,t}\) is the leading coarse-layout direction [2509.24702]. The practical interpretation given in the paper is that when \(x_t\) is still close to Gaussian noise, the positive and negative predictions are similar, so \(\Delta_t\) is small and the negative branch contributes little at precisely the stage when it would matter most.

Synchronized directional normalization addresses this by normalizing the correction direction and scaling it by a fixed \(\lambda\), thereby ensuring that the counterfactual signal has meaningful effect from the first denoising step. The paper characterizes this as making suppression **synchronized with generation rather than delayed** [2509.24702].

Cumulative trajectory bias is the second failure mode. If positive and negative prompts are evaluated on the same latent trajectory, then the negative branch inherits a state already shaped by the positive prompt’s accumulated semantics. This makes the negative branch less faithful to the intended counterfactual physics. Trajectory-decoupled denoising addresses this by separating \(x_t^+\) and \(x_t^-\), so that the counterfactual branch remains free from the positive branch’s accumulated bias. In the paper’s terms, the two SDG components are complementary: normalization makes suppression immediate, and decoupling makes it unbiased [2509.24702].

## 5. Empirical evaluation

The paper evaluates the method on **CogVideoX-5B** and **Wan2.1-14B** [2509.24702]. The benchmark suite consists of **PhyGenBench**, which contains **160 prompts**, **27 physical laws**, and **4 domains**—mechanics, optics, thermal, and material—measured by **Physical Commonsense Alignment (PCA)**, and **VideoPhy**, which evaluates real-world actions using **Semantic Adherence (SA)** and **Physical Commonsense (PC)**. Implementation details reported in the paper include resolutions of \(480\times720\) for CogVideoX-5B and \(480\times832\) for Wan2.1-14B, **25 frames per video**, **50 inference steps**, and all experiments conducted on **one NVIDIA RTX 5090 (32GB)** [2509.24702].

The main quantitative results show gains on both backbones. The strongest reported improvement is on PhyGenBench for Wan2.1-14B, where PCA improves from **0.40** to **0.50** [2509.24702].

| Backbone | Baseline | + Ours |
|---|---:|---:|
| CogVideoX-5B | SA 0.48, PC 0.39, PhyGenBench 0.47 | SA 0.49, PC 0.40, PhyGenBench 0.49 |
| Wan2.1-14B | SA 0.49, PC 0.35, PhyGenBench 0.40 | SA 0.52, PC 0.35, PhyGenBench 0.50 |

Domain-wise, the gains are especially strong in **mechanics** and **thermal physics**. For CogVideoX-5B, mechanics improves from **0.43** to **0.49**, optics from **0.55** to **0.58**, thermal remains **0.42**, material improves from **0.46** to **0.48**, and the average rises from **0.47** to **0.49**. For Wan2.1-14B, mechanics improves from **0.36** to **0.47**, optics from **0.53** to **0.60**, thermal from **0.36** to **0.51**, material from **0.33** to **0.40**, and the average from **0.40** to **0.50** [2509.24702].

The qualitative findings align with the metrics. The paper reports more physically realistic elastic bouncing, fluid interaction, refraction and scattering, melting and phase transitions, mixing behavior, and ignition propagation. Specific examples include a tennis ball bouncing more realistically, a highlighter adhering better to cardboard, a spoon in water showing more realistic ripples and refraction, butter melting progressively rather than remaining rigid, red and blue paint becoming gradually blended, and laser light in milk exhibiting more plausible scattering [2509.24702].

The method is also compared with base video models and physics-aware trained methods including **PhyT2V**, **DiffPhy**, **VideoREPA**, and **CogVideoX-5B + WISA**. The paper’s characterization is that SDG is competitive with trained physics-aware methods while remaining training-free, inference-time only, and plug-and-play [2509.24702].

## 6. Ablations, practical characteristics, and relation to other decoupled-guidance methods

The ablation study on Wan2.1-14B and PhyGenBench isolates the contribution of each component. The reported results are: **Baseline 0.40**, **w/o SDG 0.43**, **w/o Physics-aware reasoning 0.47**, **w/o Synchronized directional normalization 0.47**, **w/o Trajectory-decoupled denoising 0.48**, and **Full model 0.50** [2509.24702]. These values support the paper’s claim that physics-aware reasoning, synchronized directional normalization, and trajectory-decoupled denoising each contribute materially, and that the two SDG components are complementary rather than redundant.

From an implementation perspective, the method is described as **training-free**, **model-agnostic**, and **plug-and-play**, and the paper states that guidance strength \(\lambda = 30\) works best in general [2509.24702]. At the same time, the reported caveats are substantive. The approach depends on the quality of the LLM’s physics-aware reasoning; if the counterfactual prompt is poorly constructed, guidance may target the wrong law. The paper also notes that VideoPhy’s evaluator does not see the user prompt and may therefore miss prompt-specific fine-grained physical issues. A plausible implication is that SDG is most reliable when the prompt can be grounded into a clear physical process.

Within the wider literature, the term **decoupled guidance** is used in related but distinct ways. In EZIGen, personalized image generation is organized so that **text guidance** dominates early denoising to form a sketch latent and **subject guidance** dominates later denoising to transfer identity-specific appearance; the paper also adds **Iterative Appearance Transfer** by revisiting time steps [2409.08091]. In RHanDS, malformed-hand refinement separates **structure guidance** from a depth image and **style guidance** from a reference hand image, and reduces interference through distinct conditioning pathways and a **two-stage training strategy** [2404.13984]. These works share the high-level principle that competing condition signals should not interfere with one another, but they address different tasks and use different mechanisms. In that sense, SDG is a physics-aware video-generation instance of a broader decoupling paradigm: separate the sources of guidance, control their interaction explicitly, and use that separation to resolve task-specific conflicts.

Source: https://www.emergentmind.com/topics/synchronized-decoupled-guidance-sdg