Adaptive-Origin Guidance (AdaOr)

• Adaptive-Origin Guidance (AdaOr) is a method that interpolates among multiple candidate strategies to balance optimality and robustness in both stochastic routing and image/video editing.
• In stochastic routing, AdaOr extends the classical SOTA framework by averaging over successor paths, thereby improving detour robustness at the cost of some nominal optimality.
• In diffusion-based editing, AdaOr smoothly blends unconditional and identity predictions, achieving state-of-the-art performance on metrics like CLIP-Dir and DreamSim consistency.

Adaptive-Origin Guidance (AdaOr) is a methodology independently established in two research domains: stochastic route planning in road networks and continuous, instruction-driven editing in diffusion-based generative models. Despite disparate application contexts, both variants of AdaOr share the central principle of interpolating among multiple origins or candidate strategies to enable robust, graded, or continuously controllable outcomes. This article synthesizes the main contributions of AdaOr in stochastic network guidance (Manseur et al., 2016) and in image/video editing models (Wolf et al., 3 Feb 2026), detailing formalism, algorithmic procedures, implementation specifics, and empirical outcomes.

1. Foundations in Stochastic Network Routing

Adaptive-Origin Guidance as formulated in the context of robust routing (Manseur et al., 2016) extends the classical Stochastic-on-Time-Arrival (SOTA) framework for finding high-probability, time-constrained routes in networks with uncertain link delays. SOTA computes, for each node $i$ and budget $t$, the maximum probability $R_i(t)$ of reaching a designated sink node within $t$ units, using a dynamic program based on a Bellman recursion over arcwise travel-time distributions. Traditional SOTA selects the single best outgoing arc per node, remaining indifferent to the multiplicity or quality of alternative detours, and consequently can prescribe brittle routes vulnerable to path failures.

AdaOr modifies the Bellman recursion to account for detour robustness. For every predecessor–current node pair $(k,i)$, and each state $(t, y)$ (remaining time, realized upstream time), the contributions to reliability from all successors $j \in \Gamma^+(i)$ are computed,

$$A_{kij}(t,y) = \int_0^t P_{ij}(\mathrm{TT}_{ij}=w\,|\,\mathrm{TT}_{ki}=y)\, u_{ij}(t-w, w)\, dw.$$

These are sorted to form a sequence $$B_{ki}^{(1)} \ge B_{ki}^{(2)} \ge \dots \ge B_{ki}^{(K)}$$, where $K$ is the maximum out-degree. AdaOr introduces weights $t$0, $t$1, and defines the robust reliability as

$t$2

This weighted mean operator interpolates between the purely optimistic (classical max) and robustness-prioritizing (averaging) strategies, favoring nodes with multiple high-quality detours.

2. Continuous Edit Control in Diffusion-Based Models

In diffusion-based image and video editing, AdaOr addresses the lack of smooth, continuous control over the intensity of text-guided edits—a limitation of conventional Classifier-Free Guidance (CFG). In standard CFG, the unconditional prediction $t$3 acts as a fixed origin of guidance, but at low guidance strengths this origin may steer samples toward distortions unrelated to the input, precluding gradual transitions between input and edited output (Wolf et al., 3 Feb 2026).

AdaOr introduces an identity prediction $t$4, conditioned on an explicit “identity” instruction corresponding to perfect reconstruction of the source input. The method linearly interpolates between $t$5 and $t$6 as a function of the edit-strength parameter $t$7 using a schedule $t$8 (typically $t$9), forming an adaptive origin: $R_i(t)$0 The final guided prediction at strength $R_i(t)$1 is

$R_i(t)$2

where $R_i(t)$3 is the CFG scale and $R_i(t)$4 is the prediction with edit instruction $R_i(t)$5.

Boundary conditions explicitly recover identity reconstruction ($R_i(t)$6) and standard CFG-based editing ($R_i(t)$7). This mechanism ensures that varying $R_i(t)$8 yields a faithful, continuous progression from input to edit.

3. Algorithmic Procedures

3.1 Successive Approximation in Network Routing

AdaOr in transportation implements its weighted Bellman equation via a successive approximation scheme. The value function $R_i(t)$9 is initialized across discretized state grids, and at each iteration, for every $t$0:

1. Compute $t$1 for all $t$2.
2. Sort $t$3 to get $t$4.
3. Update $t$5 as the weighted sum over sorted $t$6 with weights $t$7. Iterations continue until the supremum norm between iterations falls below threshold $t$8. If $t$9 and $(k,i)$0 (number of iterations) are modest, the overall complexity is $(k,i)$1, with $(k,i)$2 the number of discretized time steps.

3.2 AdaOr for Diffusion Inference

Diffusion-based AdaOr builds on standard sampling procedures (e.g., DDIM, Euler–Maruyama):

1. For each diffusion step $(k,i)$3:
   • Compute $(k,i)$4 (null instruction),
   • Compute $(k,i)$5 (current edit),
   • Compute $(k,i)$6 (identity instruction),
   • Interpolate the adaptive origin $(k,i)$7 with $(k,i)$8,
   • Compute the final guided prediction,
   • Update the latent.
2. At $(k,i)$9, decoding produces the original input; at $(t, y)$0, the fully edited version.

Training requires only a new “(id)” token, included in 10% of instruction-labeled batches with source-target identity matching, using the same denoising objective as standard models. No specialized continuous-edit dataset is needed.

4. Empirical Evaluation and Key Outcomes

4.1 Network Routing

Manseur et al. (Manseur et al., 2016) demonstrate AdaOr on a 5×5 grid network with Gamma-distributed link delays, including bivariate correlations. AdaOr policies with lower $(t, y)$1 ('robustness weight') select successors with richer detour options. Quantitative experiments show that, as robustness increases (lower $(t, y)$2), the nominal on-time arrival probability declines, but resilience to path disruption increases—capturing the "robustness overhead." For instance, at $(t, y)$3, classical and robust strategies yield on-time reliability values differing by $(t, y)$4, while time budgets differ by $(t, y)$5 to achieve $(t, y)$6.

4.2 Diffusion Model Editing

AdaOr (Wolf et al., 3 Feb 2026) is benchmarked against continuous-editing baselines including FreeMorph, Kontinuous Kontext, Concept Sliders, and SAEdit across metrics such as second-order smoothness ($(t, y)$7), CLIP-Direction alignment, DreamSim consistency, and edit linearity. Results (summarized in the tables below) show AdaOr achieves state-of-the-art smoothness, directionality, and consistency, especially compared to standard CFG which exhibits random artifacts at low edit strength and morph-based baselines which suffer from semantic entanglement.

Method	$(t, y)$8	CLIP-Dir$(t, y)$9	DreamSim$j \in \Gamma^+(i)$0	Linearity$j \in \Gamma^+(i)$1
FreeMorph	0.26	1.71	0.23	0.10
Kontinuous Kontext	0.12	1.75	0.32	0.08
Standard CFG	0.61	1.48	0.27	0.12
CFG–(id)	0.27	1.65	0.30	0.05
Linear scheduler	0.14	1.99	0.36	0.07
AdaOr (ours)	0.12	1.89	0.36	0.07

On the human-focus benchmark, AdaOr also achieves highest CLIP-Dir and DreamSim consistency, with competitive smoothness and linearity compared to Concept Sliders and SAEdit.

User studies (36 participants, 10 tuples each) confirm AdaOr’s subjective superiority over morph- and context-based alternatives for smoothness, quality of intermediate steps, and overall edit preference.

5. Implementation Considerations and Limitations

The road network AdaOr approach incurs pseudo-polynomial complexity due to the cubic scaling with discretized time steps, though practical values remain manageable since $j \in \Gamma^+(i)$2 (out-degree) and $j \in \Gamma^+(i)$3 (iterations) are bounded in real networks.

Diffusion-based AdaOr introduces only minor overhead: three prediction passes per diffusion step (unconditional, conditional, identity). Training only requires a single additional instruction token and standard synthetic triplets, leveraging existing datasets. The most effective scheduler in reported experiments is $j \in \Gamma^+(i)$4, though linear schedules are also viable with slightly reduced smoothness.

Both AdaOr variants inherit the failure modes of their respective base frameworks: representational limitations in diffusion backbones, and statistical modeling assumptions (e.g., Markovian, parametric time distributions) in network routing. The price of robustness or edit smoothness is an explicit reduction in the nominal optimality for those favoring maximum probability or minimum distortion.

6. Broader Perspectives and Contextualization

Adaptive-Origin Guidance represents a principled strategy for interpolating among origin policies or prediction anchors to balance competing objectives: robustness versus nominal optimality in routing, and semantic consistency versus edit strength in generative models. In both contexts, AdaOr decomposes continuous decision control into interpretable, parameterized blends of extreme strategies. This abstraction aligns with other methods leveraging multi-anchor blending or averaged value function updates for stability, although AdaOr is distinct in deploying explicit, learnable or adaptive origins within standard optimization or inference regimes.

AdaOr’s road network instantiation foregrounds the trade-off between reliability and adaptability in the presence of stochastic failures by leveraging weighted successor averaging. In generative editing, the method resolves CFG’s arbitrary artifacts at low strength by grounding the origin at the true input, restoring continuous semantics to the edit-strength slider.

These results situate AdaOr as a generic blueprint for robust and continuous control in structured decision and generation tasks, with demonstrated empirical advantages and minimal integration cost within existing architectures (Manseur et al., 2016, Wolf et al., 3 Feb 2026).