Domino introduces a Markov-chain sampler for faster masked diffusion generation
Grafy Research Lab unveils Domino, a first‑order Markov chain framework that improves few‑step masked diffusion by modelling intra‑step token dependence.
Source: Grafy Research LabMasked diffusion models have become a popular way to generate sequences—text, images, or other modalities—by progressively uncovering masked positions over many denoising steps. The standard sampling procedure treats each position revealed in a single step as independent, which works well when many steps are available but degrades sharply when the sampler is forced to unmask many tokens at once.
What problem does Domino address?
Traditional masked diffusion samplers reveal a small number of tokens at each denoising iteration, relying on a reverse kernel that assumes independence among tokens uncovered in the same step. When developers attempt to speed up generation by using only a few steps, the independent assumption leads to incoherent or low‑quality outputs because the model ignores how newly revealed tokens influence each other. Domino targets this gap by explicitly modelling the joint distribution of tokens uncovered together.
How does Domino model intra‑step dependencies?
The core idea is to treat the set of positions revealed in a single sampling step as a sequential chain. The first position is sampled using the posterior distribution of a pre‑trained masked diffusion model (MDM), preserving the original model's knowledge. Subsequent positions are then generated by a first‑order Markov transition that conditions on the previously revealed token within the same step. This creates a joint distribution that captures the conditional dependence among tokens uncovered together, rather than treating them as independent draws.
What are the practical benefits for creators and developers?
By accounting for token‑to‑token influence within a step, Domino can achieve comparable or better generation quality with far fewer denoising iterations. For developers building real‑time AI applications—such as interactive storytelling, rapid image synthesis, or low‑latency code completion—this translates into reduced compute cost and faster response times without sacrificing fidelity. Creators can also experiment with more aggressive step reductions, opening the door to novel workflows like instant image‑to‑video transformations.
How can the research be explored with Grafy tools?
Researchers and practitioners can dive into the Domino paper directly using Grafy Reader, which lets you upload the PDF, ask natural‑language questions, and retrieve specific sections such as the Markov‑chain formulation or experimental results. To prototype a few‑step sampler that incorporates Domino’s chain, developers can use Grafy Code to prototype Python code that wraps an existing masked diffusion model and adds the Markov transition layer. The integrated environment lets you run the code in the cloud, switch between model providers (OpenRouter, Vertex AI, etc.), and compare outputs side‑by‑side.
Does Domino replace existing diffusion pipelines?
No. Domino is presented as a complementary sampling strategy that can be plugged into any pre‑trained masked diffusion model. The underlying diffusion process and the learned denoising network remain unchanged; only the reverse sampling kernel is altered to follow the Markov chain. Existing pipelines that rely on many steps can continue as before, while those that need speed‑ups can adopt Domino for a few‑step regime.
What are the limitations and future directions?
The authors note that the Markov‑chain approximation still assumes a first‑order dependency, which may not capture higher‑order interactions among a large group of tokens revealed together. Extending the framework to higher‑order chains or learning richer transition operators is an open research avenue. Additionally, empirical validation is currently limited to certain modalities; broader testing across text, audio, and 3‑D generation will help gauge its generality.
Source: Grafy Research Lab, "Domino: Joint Unmasking via First‑order Markov Chain for Few‑Step Masked Diffusion" (arXiv:2611.18645, 2026‑11‑17).
Frequently asked questions
Does Domino require retraining the underlying diffusion model?
No. Domino works with a pre‑trained masked diffusion model; it only modifies the sampling procedure by adding a Markov‑chain step, leaving the original model weights untouched.
Can I use Domino for image generation in Grafy Creator?
Domino itself is a sampling technique, not a generation app. However, you can integrate the Domino sampler into a custom workflow using Grafy Code and then feed the generated images into Creator for further editing or up‑scaling.
Is the Markov‑chain approach limited to text tokens only?
The paper describes the method in a modality‑agnostic way, applying it to any masked diffusion model. While experiments focus on token‑based data, the same principle can be extended to pixel or voxel positions in image or 3‑D generation.


