Dinsky Plus Generator

Dinsky Plus is a procedural art generator that draws Kandinsky-style abstract compositions - colored circles, rectangles, and lines scattered across a canvas with no diffusion model involved. Give it a size and a seed and it paints you a fresh abstract image in milliseconds. No checkpoint, no sampler, no VRAM. That's the appeal: it's instant, reproducible, and infinite.

The practical uses are more varied than they sound. These make great background textures to composite AI subjects onto, useful as style-reference inputs, or a fast source of abstract art for dataset experiments. And because it's seeded, the same seed always gives the same painting - which makes it handy as a controllable element in bigger workflows rather than just random noise.

How it works

The mechanism is refreshingly simple: it creates a blank RGB canvas in PIL, then loops num_shapes times, each iteration picking a shape type (weighted by circle_weight, rectangle_weight, line_weight), a random color from the selected palette, and random geometry, then draws it. The weights are normalized into probabilities, so cranking circle_weight to 10 vs 1 for lines makes the canvas mostly circles. The seed drives both Python's random and numpy's, so everything is reproducible.

The kandinsky palette is the namesake and the best default - a warm set of corals, golds, and teals that genuinely evokes the painter. There are also warm, cool, monochrome, and vibrant options.

The inputs that matter

• width / height - canvas size (64–4096). Default 1024².
• seed - the reproducibility knob. Change it, get a different painting; keep it, get the same one forever.
• num_shapes - shape count (1–1000). 100 is a nice abstract density; 1000 is basically confetti.
• min_size / max_size - shape size range.
• circle_weight / rectangle_weight / line_weight - the three shape-type probabilities. This is the personality control.

Output is a single IMAGE at your requested size.

Installing it

Dinsky Plus is part of DJZ-Nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/MushroomFleet/DJZ-Nodes
cd DJZ-Nodes
pip install -r requirements.txt

Restart ComfyUI, or install via Manager (search "DJZ-Nodes"). It only needs Pillow and numpy - both ship with ComfyUI, so it works out of the box.

Common issues

Not much can go wrong here, which is the point. The one thing to watch: line shapes are capped at a width of min_size // 10, so if min_size is tiny your lines come out hairlines. And there's no "background" input - the canvas is always solid white, so if you want a specific backdrop you'll need to composite the output yourself. For a version that anchors shapes to explicit coordinates instead of scattering them, the pack also has Dinsky Plus V2.