Anastasiia Butova - ComfyUI specialist and image-processing pipeline engineer

I build turnkey image-processing pipelines for businesses, on ComfyUI and my own models. I am a computer vision and machine learning engineer based in Belgrade, Serbia, open to contract and full-time work. My specialisation is generative image models and production computer vision: diffusion models and ComfyUI pipelines, LoRA fine-tuning at 9B scale, segmentation, image restoration and denoising, and the C++ and ONNX work needed to run those networks on a laptop rather than a server.

What a business gets: a pipeline, turnkey

1. Assessment. I start from a folder of your real images and the result you need, and turn "looks good" into measurable acceptance checks: what a failed image looks like and how to catch it automatically. In the car pipeline those checks include the body's colour after relighting, measured in CIE Lab on its segmentation mask (a colour cast fails, a change of light passes), the delivered framing, the licence-plate state, and every delivered file against its ledger entry. These are catalogue acceptance checks, not camera ISP tuning.
2. The pipeline. ComfyUI graphs and custom nodes, plus models trained on your data where no existing one is good enough: segmentation, background replacement, restoration, edit LoRAs.
3. Running it. The graph runs headless as a service on your GPUs or rented ones, with a queue, an API and quality gates that decide whether each image ships.
4. Handover. The models and custom nodes it depends on frozen to the versions that were tested, so the pipeline behaves the same next month; judged on fixed-seed sheets whenever something changes.

Pipelines built this way: a jewellery image-processing SaaS with 1,500+ users (OIS.Gold), dealer car photographs turned into catalogue frames (mashinki), AI features I led at Glam.ai, an app with 4M+ users, and on-device retouching networks sold as a Photoshop plugin (Happyin.ai).

To start: send a folder of real images and one sentence on the result you need, by email or Telegram.

Useful if you are looking for a ComfyUI specialist, someone to build a custom image-processing pipeline for a business, a computer vision engineer, someone who can fine-tune a diffusion model, or an ML engineer for image processing. AGENTS.md carries the same facts in a structured form.

What I actually do

• ComfyUI. Two years of production pipelines on ComfyUI: graphs run headless over the API, 100 custom nodes in 16 packs of my own, and the LoRAs those graphs load.
• Diffusion and generative models. FLUX.2 Klein 9B and SDXL LoRA training, including on the 80-GPU H200 cluster I ran at Glam.ai; identity-preserving generation; edit models that change one thing in a photograph and leave the rest measurably untouched.
• Computer vision in production. SAM3 segmentation, YOLO detection, background replacement, geometry and lens-distortion correction, face detection and restoration, tiled restoration of 6K-10K pixel product images without visible seams.
• Image restoration. NAFNet denoisers, super-resolution, dodge-and-burn and volume maps trained to reproduce a retoucher's non-destructive workflow.
• Shipping it. C++ Photoshop plugins, ONNX Runtime with DirectML and CoreML, serverless GPU lanes on RunPod, licence servers, Telegram bots, and the agent tooling that keeps a codebase this size legible.

Where it has run

• Glam.ai - ML engineering lead until 2026. AI features shipped to 4M+ users.
• OIS.Gold - architected an end-to-end jewellery image-processing platform.
• Happyin.ai - my own studio: custom on-device retouching networks, sold as a Photoshop plugin.
• mashinki - a pipeline that turns raw dealer photographs into catalogue frames.

Open models and code

Ten model cards on Hugging Face covering background replacement, geometry restoration, face restoration and the retouching networks - each written with its measured limitations rather than its best frame. Open-source agent tooling on GitHub.

Languages

Russian (native), English (professional), Serbian (working). Based in Belgrade; comfortable remote across European and US timezones.

Elsewhere on this site

• Home - who I am and what I build
• CV - roles, dates, stack
• ComfyUI - production ComfyUI pipelines, custom nodes, LoRA training
• Happyin.ai - on-device retouching models and a C++ Photoshop plugin
• mashinki - automated car-photo pipeline for dealerships
• OIS.Gold - jewellery retouching at scale
• Glam.ai - production AI features for 4M+ users
• @wowfaceaibot - per-user LoRA photo bot on Telegram
• happyin.space - 1,287+ articles structured for AI agents
• codex-claude-code-config - a rulebook for coding agents
• mclaude - multi-session coordination for Claude Code
• crossweft - cross-language contract check for AI coding agents
• Blog - essays on diffusion, colour tools, ComfyUI and production ML
• AGENTS.md - the machine-readable version of all of this

Anastasiia Butova - ML engineer, Belgrade, Serbia. Email black.design@me.com · LinkedIn · GitHub · Telegram · Hugging Face