Continuous Speculative Decoding for Autoregressive Image Generation

Conference: ECCV 2026
arXiv: 2411.11925
Code: None
Area: Image Generation
Keywords: Speculative Decoding, Autoregressive Image Generation, Diffusion Models, Inference Acceleration, Continuous Distribution

TL;DR

This paper is the first to extend speculative decoding from discrete to continuous distributions. It enhances the draft-target acceptance rate through denoising trajectory alignment, and resolves the sampling bottleneck of the correction distribution (which lacks an analytical expression) via acceptance-rejection sampling. It achieves over 2x wall-clock speedup on multiple continuous visual AR models including MAR, xAR, and Harmon while maintaining generation quality, completely training-free.

Background & Motivation

Background: Continuous visual autoregressive (AR) models (e.g., MAR, xAR, Harmon) generate images token-by-token through a diffusion denoising process. This avoids the training instability and information loss associated with discrete VQ tokenizers, leading to outstanding performance in image generation quality. However, similar to LLMs, their token-by-token serial decoding inference is extremely slow, presenting a major bottleneck for practical deployment.

Limitations of Prior Work: Speculative decoding is a mature acceleration technique in the LLM domain where a small draft model quickly generates candidate tokens, which are then parallelly verified by a large target model, deciding acceptance or rejection based on the probability ratio \(p(x)/q(x)\). Existing works (LANTERN, SJD) have extended this to discrete visual AR models, but all these methods rely on direct computation of discrete probability distributions and cannot be directly applied to continuous distribution scenarios.

Key Challenge: Under continuous distributions, there are two fundamental obstacles. (1) Extremely low acceptance rate: Because the draft and target models learn different data distributions, the denoising trajectory of the draft diverges significantly from the target's expected trajectory. This results in an incredibly small probability ratio \(p(Y)/q(Y)\) for shared paths, on the order of \(5.33 \times 10^{-23}\), making the acceptance rate nearly 0%. (2) No analytical expression for the correction distribution: If a token is rejected, it must be resampled from the correction distribution \(p'(x) = \mathrm{norm}(\max(0, p(x)-q(x)))\). However, in continuous spaces, this distribution involves high-dimensional Gaussian product integration \(Z = \int \max(0, p-q) \, dY\), which cannot be calculated analytically or sampled directly.

Key Insight: Replace the infeasible exact criterion with an approximate acceptance criterion \(p(Y_p)/q(Y_q)\), which evaluates modeling probabilities of both models on the same token \(x_0\) rather than forcing them along the identical trajectory. Then, improve the acceptance rate and eliminate exponential terms in the probability ratio via denoising trajectory alignment (sharing reparameterized noise \(\varepsilon_t\)). Finally, use acceptance-rejection sampling combined with a clever upper-bound factor \(M = 1/Z\) so that \(Z\) cancels out, reusing the simplified results of trajectory alignment to avoid extra model inference. These three designs are closely linked to form a complete continuous speculative decoding scheme.

Core Idea: Share random noise \(\varepsilon_t\) between the draft and target models during the reparameterization of diffusion denoising. This reduces the expected squared distance of the two trajectories by \(2 \cdot \mathrm{tr}[\sqrt{\Sigma_t^q \Sigma_t^p}]\), thereby substantially boosting the acceptance rate. Simultaneously, leverage this alignment to eliminate complex integral terms in the correction distribution sampling, achieving highly efficient resampling with zero extra inference overhead.

Method

Overall Architecture

The core problem to be solved by this method is how to accelerate inference under the premise of keeping the output distribution of the target model \(M_p\) unchanged, given a continuous AR model as target \(M_p\) and a structurally identical but smaller draft \(M_q\). The overall pipeline follows the standard draft-and-verify paradigm of speculative decoding, but adapts two key stages for continuous distributions: (1) acceptance criterion—using the approximate ratio \(p(Y_p)/q(Y_q)\) instead of the infeasible \(p(Y)/q(Y)\); (2) post-rejection resampling—using acceptance-rejection sampling instead of directly sampling the correction distribution.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input: Prefix + Conditions"] --> B["Draft Model Autoregressively Generates<br/>γ Candidate Tokens"]
    B --> C["Target Model Parallelly Verifies<br/>Forcing Shared ε<sub>t</sub> (Denoising Trajectory Alignment)"]
    C --> D["Calculate Approximate Acceptance Ratio<br/>p(Y<sub>p</sub>)/q(Y<sub>q</sub>)"]
    D --> E{"Token-by-Token Determination<br/>Ratio > r ~ U(0,1)?"}
    E -->|"Accept"| F["Keep Draft Token<br/>Continue to Next Position"]
    E -->|"Reject"| G["Acceptance-Rejection Sampling<br/>Sample from Correction Distribution p'"]
    F --> H["Output Sequence"]
    G --> H

The framework uses a small portion of tokens prefilled by the target model as a reliable prefix (resolving inconsistency in early AR steps). The draft model autoregressively generates \(\gamma\) candidate tokens based on this prefix, recording the reparameterized noise \(\varepsilon_t^q\) during the denoising trajectory of each token. The target model then verifies all candidates in parallel using the same prefix, forcing the use of the same \(\varepsilon_t\) (i.e., \(\varepsilon_t^p = \varepsilon_t^q\), achieving denoising trajectory alignment) in the denoising process, calculates the acceptance ratio for each position, and decides acceptance or rejection token-by-token. For the first rejected position, a new token is resampled from the correction distribution using acceptance-rejection sampling, and the final merged sequence is output.

Key Designs

1. Approximate Acceptance Criterion and Denoising Trajectory Alignment: Solving the Fundamental Problem of Tiny Probability Ratios

Directly applying the acceptance criterion \(p(Y)/q(Y)\) from discrete speculative decoding is completely infeasible. In continuous diffusion models, generating token \(x_0\) involves a \(T\)-step denoising process \(x_T \to x_{T-1} \to \dots \to x_0\), where these \(T\) intermediate states form a trajectory \(Y = [x_0, x_1, \dots, x_T]\). Theoretically, the probability ratio should compute the joint probability assigned to the identical trajectory \(Y\) by target and draft models. However, the trajectory \(Y\) generated by the draft is highly unlikely to fall into the high-probability region of the target. The ratio for any specific trajectory is miniscule, accumulating to \(p(Y)/q(Y) \approx 5.33 \times 10^{-23}\) after \(T\) steps, yielding a 0% acceptance rate.

This paper instead proposes an approximate acceptance criterion \(p(Y_p)/q(Y_q)\): given the same final token \(x_0\) (generated by the draft), the target model reruns the denoising process with \(x_0\) as the destination using its own model parameters to obtain trajectory \(Y_p\) (such that \(x_0^p = x_0^q = x_0\)), and then computes the ratio of joint probabilities under their respective model trajectories. Intuitively, this is equivalent to comparing the likelihood assigned to "the same token \(x_0\)" by the two models, rather than requiring them to follow the exact same path.

Relying solely on the approximate criterion is insufficient—if the two trajectories independently sample noise \(\varepsilon_t^p \neq \varepsilon_t^q\), the distance between tokens remains large. This paper proposes denoising trajectory alignment: forcing \(\varepsilon_t^p = \varepsilon_t^q\), i.e., using the identical random noise at each denoising step, during reparameterization sampling \(x_{t-1} = \sqrt{\Sigma_\theta(x_t, t)} \cdot \varepsilon_t + \mu_\theta(x_t, t)\).

Theorem 1 (Reparameterization Proximity): Let \(\varepsilon_t^p = \varepsilon_t^q\), the difference in expected squared distance of two steps is: $\(\mathbb{E}_{\neq}[||x_{t-1}^q - x_{t-1}^p||^2] - \mathbb{E}_{=}[||x_{t-1}^q - x_{t-1}^p||^2] = 2 \cdot \mathrm{tr}\left[\sqrt{\Sigma_t^q \Sigma_t^p}\right] \geq 0\)$

After alignment, the common exponential term \(\exp\{\frac{1}{2}\varepsilon_t^T\varepsilon_t\}\) in the Gaussian PDF cancels out in the numerator and denominator, dramatically simplifying the acceptance criterion to a compact form that only involves the determinants of variances: $\(\frac{p(Y_p)}{q(Y_q)} = \frac{p_\theta(x_0|x_1^p)}{q_\theta(x_0|x_1^q)} \cdot \Sigma, \quad \Sigma = \frac{\prod_{t=2}^T \sqrt{|\Sigma_t^q|}}{\prod_{t=2}^T \sqrt{|\Sigma_t^p|}}\)$

where \(\Sigma\) accumulates along the intermediate denoising steps and can be calculated efficiently. Experimental results verify that alignment reduces the average token distance from 2.56 to 1.13 and boosts the acceptance rate from 7% to 30%+. Without alignment, visual distortions and artifacts appear in generated images, which are significantly eliminated by alignment. Thus, alignment ensures both acceleration and generation quality.

2. Token Prefilling: Compensating for Prefix Inconsistency in Early Autoregressive Steps

In addition to the inconsistency at the denoising trajectory level, inconsistency also exists between the draft and target models along the autoregressive dimension. In the first few steps of generation, both models generate tokens conditioned on their own predicted prefixes. Discrepancies in the prefix embeddings cause the output distributions to diverge, resulting in extremely low acceptance rates (around 5%, see Fig. 9 in the paper) in early stages. As the AR steps progress and the prefix is gradually filled with tokens verified by the target model, the distributions converge and the acceptance rate recovers.

The concept of Token Prefilling is straightforward: prior to starting speculative decoding, the target model itself generates a small portion of prefix tokens (e.g., 5% of total steps), ensuring that subsequent draft verification is always based on a consistent target prefix. This does not add extra latency because under a low acceptance rate, speculative decoding degenerates to token-by-token decoding by the target model anyway, so prefilling simply performs this step ahead of time.

Experiments show that 5% prefilling increases the acceptance rate from 0.25 to 0.30 (MAR-H + MAR-B, \(\gamma=32\)) while maintaining the speedup ratio (unchanged at 1.63×). 15% prefilling further improves the acceptance rate to 0.33, but the speedup ratio starts to decline slightly to 1.61×—as prefilled tokens do not participate in speculative acceleration, the returns diminish.

3. Acceptance-Rejection Sampling: Resolving Direct Sampling Infeasibility of the Correction Distribution and Avoiding Duplicate Model Inference

When a draft token is rejected, a new token must be resampled from the correction distribution: $\(p'(Y) = \frac{\max(0, p(Y) - q(Y))}{Z}, \quad Z = \int_{Y'} \max(0, p(Y') - q(Y')) dY'\)$

Since \(Z\) involves integration of a high-dimensional Gaussian product, it lacks an analytical expression, making direct sampling impossible. This paper employs acceptance-rejection sampling: sampling candidates from the target distribution \(p(Y)\) and accepting them with probability \(\alpha_s = p'(Y) / (M \cdot p(Y))\).

The key trick lies in selecting the upper bound factor. Since \(\max(0, p-q) \leq p\), setting \(M = 1/Z\) precisely cancels out \(Z\) in \(\alpha_s\): $\(\alpha_s = \frac{\max(0, p(Y)-q(Y))/Z}{p(Y)/Z} = \frac{\max(0, p(Y)-q(Y))}{p(Y)}\)$

However, this still requires running the full diffusion denoising to sample from \(p(Y)\), which is highly costly. To address this, the simplified results from denoising trajectory alignment (Design 1) are reused to derive Corollary 1 (Easy-to-Compute Rejection Threshold): $\(\alpha_s = \frac{\max(0, \Sigma \cdot p_\theta(x_0|x_1^p) - q_\theta(x_0|x_1^q))}{\Sigma \cdot p_\theta(x_0|x_1^p)}\)$

Here, \(x_0\) only needs to be sampled from the single-step Gaussian distribution \(p_\theta(x_0|x_1^p)\) (one forward pass) without running the full \(T\)-step denoising process. Empirical measurements show that the runtime of acceptance-rejection sampling is a tiny fraction of the total model inference time (0.0047 seconds vs 51 seconds of model inference for \(\gamma=32\)). Resampling usually accepts candidates within 2-3 iterations.

A Complete Example

Taking MAR-H as target, MAR-B as draft, and \(\gamma=4\) as an example, we demonstrate a single speculative decoding step (corresponding to Algorithm 1 in the paper):

1. Prefix Preparation: A 5% prefix of tokens has been prefilled by MAR-H to serve as the common condition for subsequent draft/target runs.
2. Draft Generation: Guided by the prefix, MAR-B autoregressively generates 4 candidate tokens \(\{x_1,x_2,x_3,x_4\}\), recording \(\varepsilon_t^q\) for each step of diffusion denoising sampling.
3. Parallel Target Verification: Conditioned on the identical prefix, MAR-H runs 5 forward passes in parallel (prefix, prefix\(+x_1\), \(+x_{1:2}\), \(+x_{1:3}\), \(+x_{1:4}\)). Each denoising step uses exactly the same \(\varepsilon_t\) as in step 2 (trajectory alignment) to obtain target distribution parameters at each position.
4. Token-by-Token Determination: For each candidate \(x_i\), compute the acceptance ratio \(r_i = \Sigma \cdot p_\theta(x_i|x_1^p) / q_\theta(x_i|x_1^q)\) and compare it to \(u_i \sim U(0,1)\). If \(r_i > u_i\), it is accepted; otherwise, it is rejected. Suppose \(x_1, x_2\) are accepted (\(r_1=2.1, r_2=1.5\)), and \(x_3\) is rejected (\(r_3=0.6, u_3=0.8\)).
5. Resampling: Perform acceptance-rejection sampling at position \(x_3\)—sample candidate \(x_t\) from \(p_\theta(x_0|x_1^p)\), compute \(\alpha_s = \max(0, \Sigma \cdot p - q)/(\Sigma \cdot p)\), and accept \(x_t\) if \(\epsilon \sim U(0,1) < \alpha_s\). In practice, the candidate is accepted after 2-3 iterations.
6. Output: The final output = prefix \(+ [x_1, x_2, x_t]\). In this step, 3 valid tokens are obtained (instead of 1 in the serial execution), saving 2 serial forward passes of the target model.

Loss & Training

This method is completely training-free: it does not require any fine-tuning, distillation, or architectural modifications for either the draft or target models, and directly uses official pretrained weights. Speculative decoding theoretically guarantees that the output distribution is identical to the original target model (lossless acceleration), and the speedup is derived entirely from the parallel draft-and-verify mechanism during inference.

Key hyperparameters: (1) draft length \(\gamma \in \{4, 8, 16, 32\}\)—larger values offer higher speedup potential but lead to a decaying acceptance rate, serving as the central lever for acceleration; (2) prefilling ratio—5% is recommended, as overly large values do not participate in acceleration and yield diminishing returns; (3) diffusion sampler—supports DDPM and DDIM variants, maintaining 2.30x speedup with DDIM in 100 steps. All experiments were conducted on a single A100 GPU with batch size of 1-256.

Key Experimental Results

Main Results

Speedup ratio and quality metrics (wall-clock, single NVIDIA A100):

Model Pair (Target + Draft)	Resolution	Max Speedup	Corresponding \(\alpha\)	Impact on Quality
MAR-H (943M) + MAR-B (208M)	256×256	2.33× (bs=256, \(\gamma\)=32)	0.19	FID 2.35→2.36±0.05, almost identical
xAR-H (1.1B) + xAR-B (172M)	256×256	2.72× (bs=256, \(\gamma\)=32)	0.22	FID 1.79→1.82±0.07, almost identical
Harmon-H (1.5B) + Harmon-B (0.5B)	256×256	2.05× (bs=32, \(\gamma\)=32)	0.17	FID 8.39→8.38, CLIPScore 34.8→34.7
Harmon-H (1.5B) + Harmon-B (0.5B)	512×512	2.54× (bs=32, \(\gamma\)=32)	0.15	All metrics almost identical

The speedup ratio positively correlates with the batch size (larger batches yield smaller draft-to-target inference time ratios \(c\), making the parallel verification advantage more prominent) and the draft length \(\gamma\); the acceptance rate negatively correlates with \(\gamma\) (longer drafts imply greater difficulty in predicting later tokens). MAR and xAR show no significant differences in FID/IS compared to the original models (all within standard error ranges), while Harmon exhibits only a minimal decrease in the fine-grained Color Attri. metric of Geneval (0.48→0.44), with other metrics remaining stable.

Ablation Study

Effect of denoising trajectory alignment (MAR-H + MAR-B):

\(\gamma\)	\(\alpha\) (w/o Alignment)	\(\alpha\) (w/ Alignment)	Token Distance (w/o Alignment)	Token Distance (w/ Alignment)
32	0.07	0.30	2.56	1.13
16	0.07	0.33	2.36	0.91
8	0.13	0.31	2.22	0.82
4	0.14	0.32	2.17	0.80

Alignment reduces the token distance from >2 to approximately 1, boosting the acceptance rate from ~7% to ~30%. Visually, without alignment, images exhibit clear distortions and artifacts, which are significantly eliminated after alignment—proving that alignment secures both the acceptance rate and generation quality.

Ablation on the Token Prefilling Ratio (MAR-H + MAR-B, \(\gamma=32\), bs=256):

Prefilling Ratio	\(\alpha\)	Speedup
0%	0.25	1.63×
5%	0.30	1.63×
15%	0.33	1.61×

A 5% prefilling ratio effectively improves the acceptance rate without compromising the speedup ratio; 15% further improves the acceptance rate, but the speedup ratio starts to decline slightly.

Key Findings

• Denoising trajectory alignment is the most significant single component: Removing alignment causes the acceptance rate to drop drastically from 30% to 7% (\(\gamma=32\)), introducing severe visual artifacts. Alignment serves as both a means to accelerate and a guarantee of quality.
• The optimal draft length is \(\gamma=32\): This provides the maximum speedup benefit (2.33x). Further increasing \(\gamma\) is constrained by the capability of the small model, which leads to a decaying acceptance rate and potentially lower comprehensive gains.
• Acceleration is more effective under large batch sizes: 2.33x at bs=256 vs 1.44x at bs=1. This occurs because large batch sizes reduce the relative draft-to-target inference time ratio \(c\), allowing for greater amortization benefits in parallel verification.
• Texture complexity determines the acceptance rate: Tokens in background and simple texture regions have high acceptance rates, whereas details-rich regions exhibit low acceptance rates (as confirmed by heatmap visualization). The modeling capability gap of the small draft model on fine textures is a natural bottleneck.
• DDIM sampling is equally effective: Replacing DDPM with DDIM (100 steps) still achieves a 2.30x speedup (\(\gamma=32\)), demonstrating that the method is insensitive to different diffusion samplers.

Highlights & Insights

• Three birds with one stone via sharing \(\varepsilon_t\): Forcing the two models to use identical reparameterized noise at each denoising step is an extremely simple operation at the code level, requiring only random seed synchronization. Yet, it simultaneously achieves: reducing trajectory distance to boost acceptance rates, canceling the exponential terms in the probability ratio to simplify calculations, and providing an easy-to-compute rejection threshold for acceptance-rejection sampling. Having a single design reutilized three times is the core systemic ingenuity of this work.
• Elegant construction of the upper bound factor \(M=1/Z\): In acceptance-rejection sampling, one generally needs to find an \(M\) large enough to satisfy \(M \cdot p(Y) \geq p'(Y)\). This paper directly sets \(M=1/Z\), causing \(Z\) to elegantly cancel out within \(\alpha_s\), thereby side-stepping the need to compute the complex integration. This trick is not required in discrete speculative decoding (where \(Z\) can be trivially calculated via summing), making it a refined solution unique to the continuous domain.
• Practical value of being completely training-free: It requires no distillation, fine-tuning, architectural modifications, or weight changes, and can be directly integrated into the inference pipeline of any existing continuous AR model to achieve 2x+ speedup. This zero-barrier, plug-and-play property minimizes the gap between laboratory research and engineering deployment, demonstrating direct utility in model deployment scenarios.
• Cross-domain transferability of the framework: Any autoregressive generation task with continuous distributions (e.g., audio generation, video generation, molecular generation) can directly reuse the three core designs (trajectory alignment, prefilling, and acceptance-rejection sampling), provided a draft-target model pair is available. The appendix already discusses the theoretical compatibility with GMM distributions, DDIM samplers, and other variants.

Limitations & Future Work

• Speedup ceiling constrained by model size ratio: Currently, the draft/target size ratio is not sufficiently disparate (MAR-B 208M vs MAR-H 943M, with an inference time ratio \(c \approx 0.38\)), which is much higher than typical LLM speculative decoding where the draft model accounts for only 5% of the target's inference time. The authors anticipate that speedup will increase substantially with a larger target (e.g., 7B/13B) and smaller draft (e.g., 97M/125M), though open-source continuous visual AR models at this scale are currently lacking.
• Experiments limited to diffusion-denoising distributions: While the appendix discusses compatibility with GMM and other distributions, no empirical validation is provided. All experiments were conducted using DDPM/DDIM; validation on GMM outputs (e.g., GIVT, DiCoDe) or other continuous distribution types remains to be done.
• Unclear degradation boundary of generation quality under high compression: The paper lacks a quantitative analysis of the acceptable speedup ceiling across different texture complexities. Low acceptance rates in highly textured areas could potentially lead to quality degradation in extreme scenarios (such as images with dense texture throughout).
• Rudimentary prefilling strategy: The current implementation uses a fixed prefilling ratio (5%) without dynamic adjustments based on image content or real-time acceptance rates. Adaptive prefilling strategies (e.g., dynamically transitioning to speculative decoding based on early AR step acceptance rates) represent an avenue worth exploring.
• Potential to stack acceleration in orthogonal directions: Diffusion sampling acceleration (e.g., reducing denoising steps via DDIM/DPM-Solver) and speculative decoding (reducing AR steps) act on different stages of the inference pipeline. Stacking them can further improve the overall speedup ratio, which is not systematically explored in this paper.

Related Work & Insights

• vs. Discrete Speculative Decoding (Leviathan et al., 2023; Chen et al., 2023): LLM speculative decoding is the direct predecessor of this work. The core difference is that LLMs output discrete softmax distributions, where \(p(x)/q(x)\) can be directly looked up and the normalization sum of the correction distribution is trivial. The contribution of this paper lies in addressing a series of sub-problems introduced by the two fundamental differences in the continuous domain: "probabilities cannot be directly obtained" and "correction distributions cannot be directly sampled."
• vs. Discrete Visual Speculative Decoding (LANTERN / SJD): LANTERN and SJD introduce speculative decoding to discrete visual AR models using VQ tokenizers. This work is the first to skip the information bottleneck of VQ tokenizers and implement speculative decoding directly in the continuous diffusion space, representing a paradigm shift.
• vs. Distillation Alignment (DistillSpec / Online Speculative Decoding): Aligning the draft distribution closer to the target distribution via distillation is another line of work for boosting the acceptance rate. The training-free approach in this paper is orthogonal to it, and they can be combined—first using distillation to improve draft quality, and then using the continuous speculative decoding framework of this paper to achieve lossless acceleration.
• vs. Fast Diffusion Model Sampling (DDIM / DPM-Solver / Rectified Flow): These methods reduce the number of denoising steps per token (i.e., reducing \(T\)). In contrast, this paper reduces the number of autoregressive steps (\(N \to N/\text{speedup}\)). Since they target orthogonal dimensions of inference, they can theoretically be stacked directly.

Rating

• Novelty: ⭐⭐⭐⭐☆ This paper is the first to extend speculative decoding from discrete to continuous distributions, presenting a coherent setup with three designs (trajectory alignment, prefilling, and acceptance-rejection sampling) targeting two core continuous hurdles. However, the draft-and-verify paradigm itself is inherited from LLM literature, making the top-level novelty moderate, but the contribution lies in the exquisite construction of the adapter layer.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ The paper provides meticulous wall-clock and FID/IS/CLIPScore/Geneval evaluations on 3 models (MAR/xAR/Harmon), 2 resolutions (256/512), 4 draft lengths, and multiple batch sizes. Ablations cover alignment, prefilling, temperature, CFG, DDIM, and mask generation comparisons, with theoretical bias analysis and visual heatmaps provided in the appendix.
• Writing Quality: ⭐⭐⭐⭐☆ The problem definition is clear, the discrete-continuous comparison is intuitive (Fig. 2), the theoretical derivations are complete (with both theorem and corollary proofs provided in the appendix), and the algorithm is depicted with pseudocode. Though the presence of many experimental tables makes reading slightly less fluent, the overall structure is highly logical.
• Value: ⭐⭐⭐⭐⭐ The inference speed of continuous visual AR models is a critical bottleneck for practical deployment. The training-free 2x+ speedup scheme proposed in this paper has high practical value. The framework is generalizable to continuous autoregressive generation tasks (image/audio/video/molecules), and the "shared noise" concept of denoising trajectory alignment can also inspire research on accelerating other continuous generative models.

Related Papers

• [AAAI 2026] Annealed Relaxation of Speculative Decoding for Faster Autoregressive Image Generation
• [ICCV 2025] Grouped Speculative Decoding for Autoregressive Image Generation
• [CVPR 2026] SJD-PAC: Accelerating Speculative Jacobi Decoding via Proactive Drafting and Adaptive Continuation
• [CVPR 2026] Multi-Scale Local Speculative Decoding for Image Generation
• [ICML 2026] Speculative Coupled Decoding for Training-Free Lossless Acceleration of Autoregressive Visual Generation