🎨 Colorful Starting Image

Every txt2img run starts from the same thing: a field of gaussian noise. Which means the big shapes in your image are decided by the sampler, and you influence them with words. The 🎨 Colorful Starting Image node gives you a different door - it draws an abstract composition of random shapes with numpy, hands you the image and a matching mask, and lets the sampler refine it. Push it through VAE Encode and a KSampler at moderate denoise and you get abstract wallpaper that still has legible structure, because the structure was there before the model touched it.

The author's own framing is "use it as a starting latent for fun and varying shapes, great for abstract wallpaper creation." That's accurate and refreshingly un-hyped. It's not a ControlNet replacement and it won't make your portraits better.

What it draws

Shape generation is a comma-separated whitelist in shape_string - rectangle, ellipse, circle, line, spline, dot, stripes, triangle, polygon, arc, concentric_circles - with components deciding how many get drawn and component_scale capping how big the largest one can be (a shape's max width is width × component_scale, so 0.5 on a 1024px canvas means roughly 512px across).

Colour comes from three places that layer on each other:

• color_palette - random, random_color, muted, grayscale, binary, neon, pastel or colorized (grayscale tinted with a single hue across every shape, which is the one that looks most intentional).
• color_harmony - complementary, analogous, triadic, tetradic, or none. Genuinely useful, and it does nothing at all when you pick the binary palette, because black and white have no hue to harmonize.
• fill_mode - none for flat shapes, gradient for a two-colour vertical blend, blocks for vertical strips. blocks plus colorized is where it stops looking like a test pattern.

Then the composition knobs: positioning_bias (scattered, center_weighted, edge_weighted, grid_aligned, random_weighted, or a compass corner), arrangement (spiral, burst, grid or none), and size_distribution (uniform, prefer_small, prefer_large). All three take random if you'd rather let it choose. shape_opacity (0.1-1.0) is what turns a collage into a layered one, and background_color accepts colour names, hex codes, random, or one of the noise variants (noise, noise_large, noise_color, noise_color_large) - those last ones are surprisingly good as a texture base.

For finishing there's a global grain pass (noise_level, noise_scale, noise_color), a warp (wave, noise_field, swirl) with warp_intensity, and a final blur_radius. seed makes the whole thing exactly reproducible, which matters more than you'd think - you'll want to go back to the composition that worked.

Outputs

Two: image (the render) and mask (white where shapes were drawn, black where the background shows through). The mask is the sleeper feature. It's a free region mask for inpainting or compositing without running a segmentation model.

Using it

The practical loop is: image → VAE Encode → KSampler with denoise around 0.6-0.85 → VAE Decode → whatever. Low denoise keeps the geometry and just textures it; high denoise lets the model eat the shapes and you're basically back to txt2img with an unusual push. Set components low (3-8) and component_scale high for big blocks of colour; go the other way for a busy confetti field.

Install

ComfyUI Manager → ComfyUI-mnemic-nodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/MNeMoNiCuZ/ComfyUI-mnemic-nodes

Restart. Nothing to download - this node is pure Python/Pillow. Note the pack as a whole installs transformers, opencv-python, tiktoken and a stack of other requirements you don't need for this node, which is the price of a grab-bag utility pack (it's currently v3.0.0, converted to the newer V3 node API with automatic workflow migration and a ? help page on every node).

Where it goes wrong

Nothing renders, canvas is empty. components is at 0, or shape_string got emptied out. Both fail quietly.

Everything looks like mud. shape_opacity well below 1 combined with noise_level turned up. Start with opacity at 1 and noise at 0, get a composition you like, then season it.

The image is a single flat colour. color_palette: binary with shape_opacity: 1 and a small component_scale - you've got white shapes on black doing very little. Turn on fill_mode or switch the harmony rule.