Nikita Starodubcev (@NikStarodub) header
Nikita Starodubcev (@NikStarodub) profile picture

research @Yandex, ex intern @Adobe

Saint Petersburg, Russia
Joined June 2026
Photo by Nikita Starodubcev (@NikStarodub)Photo by Nikita Starodubcev (@NikStarodub)Photo by Nikita Starodubcev (@NikStarodub)Photo by Nikita Starodubcev (@NikStarodub)Photo by Nikita Starodubcev (@NikStarodub)Photo by Nikita Starodubcev (@NikStarodub)Photo by Nikita Starodubcev (@NikStarodub)Photo by Nikita Starodubcev (@NikStarodub)
Great work on text diffusion models by my friends @free001style, @sudakov_il, and @DmitryBaranchuk. A must-read for anyone interested in this fast-growing field! TL;DR: Want to boost your text diffusion model performance? Here is a two-part recipe: MMD + Diffusion Representation Space = Better performance across all NFEs. I especially like the broad improvements across discrete, continuous, small, and large models. BUT is it only about quality?
Introducing a simple method to post-train both discrete and continuous diffusion LMs. It's based on Maximum Mean Discrepancy (MMD) that matches token-level feature distributions. Our method improves generation and speed on OWT/GSM8K and scales to 16B models. 🧵
1
4
242
I think diffusion-model representations are important yet still underexplored as a general training signal. Representation space is a very powerful, sometimes even necessary, component of training or fine-tuning. For example, Drifting Models (arxiv.org/abs/2602.04770) simply don't converge without an external encoder. BUT we almost always use features from discriminative models. Wouldn't it make more sense to use generative features when solving a generative task? The authors give a hint that the answer might be yes. It would be interesting to understand what makes these diffusion features so effective for training. I’d love to see this pushed further, for example by training a few-step model from scratch in this representation space.
1
2
41
Register tokens were introduced to fix artifacts in Vision Transformers (ViTs) (arxiv.org/abs/2309.16588) In Diffusion Transformers (DiTs), we don't observe such outliers, so registers shouldn't help. But they do... And especially for pixel-space DiTs. Simply adding a few empty tokens can improve FID from 3.52 to 2.69.
5
20
2
133
9,739
We put this to use. Registers improve image quality and add a useful signal, so why not amplify it? Register Guidance uses the model's prediction without registers as the negative direction, just as CFG uses the unconditional prediction. It combines naturally with CFG and consistently sharpens details and structure
1
3
718
And the final one – Register Guidance in action. More in the paper :) paper: arxiv.org/abs/2605.16147 project page: quickjkee.github.io/register… code: github.com/quickjkee/registe…
1
11
696