🎨 Colorful Starting Image

Most "starting image" tricks in Stable Diffusion boil down to feeding the sampler colored noise instead of pure Gaussian noise, hoping the extra structure nudges the composition somewhere interesting. Colorful Starting Image does that on purpose and with actual control: it draws a random arrangement of shapes - rectangles, circles, splines, stripes, arcs, whatever you list - onto a canvas, and hands you that as an image (and a mask) you can push through img2img at a low-ish denoise, or just save straight out as abstract wallpaper. The pack's own description is blunt about the second use case: it's genuinely good for that.

How it works

It's a pure image generator - no model, no GPU work, just procedural drawing. You tell it how many shapes to place (components), how big they're allowed to get (component_scale, as a fraction of the image width), and which shapes are in the pool (shape_string, a comma list like "rectangle, ellipse, circle, line, spline, dot, stripes, triangle, polygon, arc, concentric_circles"). Everything else is styling on top of that: a color palette family, an optional harmony rule to keep the colors from clashing, a fill mode (solid, gradient, or striped blocks), where shapes cluster on the canvas, whether they're rotated, and an optional warp/noise/blur pass at the end for texture. seed locks the whole thing down - same seed, same settings, same image, every time.

The inputs and outputs that matter

For a first pass, you really only need a handful of the many knobs here:

• width / height - canvas size, same as any latent.
• components - how many shapes get drawn. Higher = busier, more detailed image.
• shape_string - which shapes are eligible. Trim the list if you want a specific look (just circles, say).
• color_palette - random, random_color, muted, grayscale, binary, neon, pastel, or colorized. This is the single biggest lever on the final mood.
• background_color - a color name, a hex code, or one of the special values (random, noise, noise_large, noise_color, noise_color_large).
• seed - for reproducibility.

Everything past that - color_harmony, fill_mode, shape_opacity, positioning_bias, arrangement, size_distribution, allow_rotation, the noise_* and warp_* fields, blur_radius - is there for when you want to go deep on the look, and each field's tooltip in the node itself spells out exactly what it does. Outputs are image and mask, ready to wire into a VAE Encode or straight into a Save Image node.

How to install it

Through ComfyUI Manager: search for ComfyUI-mnemic-nodes, install, restart. Or by hand:

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

Restart ComfyUI afterward. No model downloads, no external API - this node is self-contained image drawing, so there's nothing extra to configure for it specifically (a couple of other nodes in this pack need a Groq API key, but this one doesn't touch any of that).

Common issues & troubleshooting

It looks nothing like a "starting latent" should. That's expected if you feed the raw output at full denoise - it'll just get erased and replaced with whatever the checkpoint wants to draw. The trick with any shape-based starting image is a partial denoise (img2img strength somewhere in the 0.5–0.8 range depending on how much of the shape structure you want to survive) so the sampler builds on top of the composition instead of ignoring it.

Results feel too chaotic or too flat. That's noise_level, warp_intensity, and components fighting each other - turn noise_level and warp_intensity down toward 0 for a cleaner, more graphic look, or push components down if it feels overcrowded. There's no wrong answer here since the whole node is meant to be played with; the defaults are just a reasonable middle ground, not a "correct" setting.

Getting a plain black or white square. Check background_color and shape_opacity - a low opacity on a small number of components against a solid background can end up mostly empty. Bumping components or shape_opacity back up fixes it.