Task: Research which models generate layouts and infographics better

black-forest-labs/flux.2-pro is doing really terribly with texts and fonts )))

Here is the draft interface with multiple model selection that I ended up with:

And now let's try running all these models with a single prompt:

Modern website for a research engineer.

This is neither a blog nor a portfolio.

It is a public journal of engineering activity that displays ongoing research, projects, tasks, working notes, and development progress in real time.

The main character of the site is Nikolai Lanets, a research engineer in the fields of artificial intelligence, agentic systems, memory architectures, and knowledge graphs.

Style:

* light theme (but not necessarily pure white)
* premium tech design
* research lab aesthetic
* a combination of GitHub, Linear, Notion, Stripe, and modern AI platforms
* minimalism without empty space (yet without heavy element crowding)
* medium information density
* high-quality typography
* neat cards
* strict visual hierarchy
* the feel of a working engineering system

Hero section:

Large name:
Nikolai Lanets

Subtitle:
research engineer

Description:
A public journal of projects, tasks, research, and engineering solutions.

Separate block:
available for new challenges

Nearby, show main research areas:

* Agent Systems
* AI Research
* Memory Architectures
* Knowledge Graphs

Below, place several sections styled as a modern research activity dashboard.

"Active Projects" section:
project cards with statuses, start dates, and a short description.

"Active Tasks" section:
research and engineering tasks displayed as a working backlog.

"Recent Worklogs" section:
a feed of recent engineering actions and decisions made.

Load section:
number of commercial tasks, number of personal research tasks, and current commitments.

Visual impression:

Not a personal website.
Not a blog.
Not a portfolio.

This is a researcher's operating system, open for public viewing.

Very high UI/UX level.
Quality on par with the best SaaS products of 2026.
Figma interface layout.
Full-fledged desktop web application.

I'll clarify once again that ChatGPT generated this based on the text from my website's homepage, I only tweaked it slightly. I wonder which models will produce what kind of design and at what cost/time.

Here are more results.

Gemini3_1_Flash_Image. $0.07

BLACK_FOREST_LABS_FLUX_2_KLEIN_4B. $0.016

BLACK_FOREST_LABS_FLUX_2_PRO. $0.06

BLACK_FOREST_LABS_FLUX_2_MAX. $0.13

black-forest-labs/flux.2-max takes a very long time to think, literally a few minutes (5-10, maybe even longer).

Судя по всему найти подходящую модель локальную для этого крайне сложно. На сколько я понимаю, обычно приходится специально под конечный задачи готовить модели, LoRa доучивать и т.п. Так же сильно упрощает процесс наличие готовых референсов. Но мне сейчас хочется именно по описанию сайта и его функционала получать макеты, потому что в том же lovable это довольно долго делается.

Added this image generation interface to haih-agent with the ability to choose the model, quality, and aspect ratio

It is tailored for OpenRouter, so it provides access to many models at once.

I fed the text from my website's homepage into ChatGPT and asked it to write a prompt for the image generator so I could play around and see which model produces the best result. But since there are several models and I don't want to wait for each one individually to compare the results later, I'll tweak the interface a bit so you can select multiple models at once, send several parallel requests, and then view all the results together.

Here are the first results.

Gemini3_1_Flash_Image. Cost $0.104

BYTEDANCE_SEED_SEEDREAM_4_5. $0.116

GOOGLE_GEMINI_3_PRO_IMAGE_PREVIEW. $0.145

Gemini2_5_Flash_Image. $0.039

OPENAI_GPT_5_4_IMAGE_2. $0.341