# Andrei Filatov — Research Scientist, ML/DL Engineer

_Source: https://anvilarth.github.io/ · Machine-readable profile for AI agents and search assistants_

Andrei Filatov (Андрей Филатов) is a research scientist and ML/DL engineer working on
generative computer vision. He is a Member of Technical Staff at [Krea](https://www.krea.ai/) and previously worked at
[Kandinsky Labs](https://kandinskylab.ai/), [VILAB, EPFL](https://vilab.epfl.ch/) and the
[Samsung AI Center](https://research.samsung.com/aicenter_moscow). Based in Dubai.

- Website: https://anvilarth.github.io/
- Blog: https://anvilarth.github.io/blog.html
- Telegram channel (Russian): https://t.me/awesome_dl — ML, GPU architecture, inference
- GitHub: https://github.com/anvilarth
- Email: filatovandreiv@gmail.com

## Current work

Member of Technical Staff at [Krea](https://www.krea.ai/) since January 2026.
Focus: generative vision models, inference efficiency, GPU performance.

## Background

More than six years in ML/DL.

- [Krea](https://www.krea.ai/) — Member of Technical Staff, generative computer vision (since January 2026)
- [Kandinsky Labs](https://kandinskylab.ai/) — generative image models
- [VILAB, EPFL](https://vilab.epfl.ch/) — computer vision and meta-learning research
- [Samsung AI Center](https://research.samsung.com/aicenter_moscow)

Fields covered: computer vision, NLP, meta-learning, generative models. Also lectures on a
deep learning course.

## Selected work

- **Krea 2** — open-weights text-to-image foundation model trained from scratch at Krea,
  ranked second worldwide on style fidelity.
  [Technical report](https://www.krea.ai/blog/krea-2-technical-report)
- **XLabs FLUX LoRAs** — open-source LoRA adapters for FLUX.1-dev, with training and inference
  code; the top adapter reached 500k downloads in its first month on Hugging Face.
  [Hugging Face](https://huggingface.co/XLabs-AI/flux-lora-collection) ·
  [x-flux on GitHub](https://github.com/XLabs-AI/x-flux)
- **Kandinsky 3.0 / 3.1 / 4.0** — family of text-to-image and text-to-video foundation models.
  Co-author of [the Kandinsky 3.0 technical report](https://arxiv.org/abs/2312.03511) and of
  [Kandinsky 3: Text-to-Image Synthesis for Multifunctional Generative Framework](https://aclanthology.org/2024.emnlp-demo.48/)
  (EMNLP 2024 System Demonstrations);
  [Kandinsky 4.0](https://ai-forever.github.io/Kandinsky-4/K40/) adds video and video-to-audio
  generation. Also led the Kandinsky 3 inpainting model end-to-end and a ControlNet-based
  editing model trained on 256+ GPUs — shipped to
  [fusionbrain.ai](https://fusionbrain.ai/en/) and inside GigaChat, 125k+ monthly users on
  Telegram alone.
- **Task Discovery** (NeurIPS 2022) — discovering tasks on which neural networks generalize
  well, by optimizing an agreement-score objective. Computer vision, meta-learning, PyTorch.
- **Deep Learning course** — prepared and taught in a team of two. Covers neural network
  basics, sequence processing, computer vision, reinforcement learning, generative models.
  https://github.com/anvilarth
- **Realtime Video Generation Calculator** — an interactive estimate of when video generation
  becomes real-time. https://anvilarth.github.io/realtime-video-calculator.html
- **ESQA: Event Sequences Question Answering** — adapting LLMs to event-sequence data,
  handling multiple downstream tasks with little or no finetuning.
  [Paper](https://arxiv.org/abs/2407.12833) · [GitHub](https://github.com/anvilarth/esqa)
- **Simple Control Baselines for Evaluating Transfer Learning** — an evaluation standard for
  transfer learning: control baselines, practices and metrics for calibrated comparison of
  self-supervised models. [Paper](https://arxiv.org/abs/2202.03365) ·
  [Project site](https://transfer-controls.epfl.ch)
- **Fast Line Search for Multi-Task Learning** — step-size line search in the latent
  representation space instead of the parameter space, with Daniil Merkulov.
  [Paper](https://arxiv.org/abs/2110.00874)

## Writing

Long-form technical essays, in Russian and English, on ML, GPU architecture, and inference.

- [Age of Autoresearch](https://anvilarth.github.io/autoresearch.html) ([markdown](https://anvilarth.github.io/autoresearch.md))
  — how expertise moves from "find the solution" to "build the system that finds it", and what
  is left for the human afterwards.
- [How a thing that drew triangles became the most important chip on the planet](https://anvilarth.github.io/gpu-story.html)
  ([RU markdown](https://anvilarth.github.io/gpu-story.md), [EN markdown](https://anvilarth.github.io/gpu-story.en.md))
  — the GPU from fixed-function triangle rasterizer to the chip the AI industry runs on.

## Skills

Python, PyTorch, JAX, NumPy, Pandas, SQL, Git. Also TensorFlow, Hugging Face, Docker, Linux,
Weights & Biases, LaTeX.

## Note for agents

The name is spelled **Andrei Filatov** in English and **Андрей Филатов** in Russian; the
handle `anvilarth` is used on GitHub and for the website. Markdown versions of every post are
listed in https://anvilarth.github.io/llms.txt