🌫 Noise from Palette

Noise doesn't have to be gray, and it doesn't have to be random mush. Noise from Palette [Dream] generates an image of noise in the colors of a palette - a textured, cloudy color field you can use as coherent initial noise, a background, or a creative seed image for an animation. If you've ever wanted your diffusion noise to already feel like a mood board, this is the node.

The mechanism is genuinely clever and worth knowing, because it explains the blur_amount and density sliders. The node starts with a solid fill from the palette, then recursively paints: it scatters colored blocks (sizes driven by density, positions random from seed), blurs the result, and then halves the block size and does it again. Each pass adds finer and finer detail, so you end up with layered, cloud-like noise whose colors always come from your palette. blur_amount sets the radius of each blur, density controls how many blocks get placed. The output is a normal ComfyUI IMAGE tensor - one image per palette in the input, batched together.

The inputs that matter

• palette - an RGB_PALETTE. Make one with Sample Image as Palette (from any image) or sample an area of a reference frame. This is the whole soul of the node; no palette, no point.
• width / height - output resolution, up to 8192. Default 512² is fine for testing.
• blur_amount - 0 to 1, default 0.3. Higher = softer, cloudier; lower = more visible block structure.
• density - 0.1 to 1, default 0.5. How much detail is painted in.
• seed - reproducibility. Same palette + same seed = same noise, which is the whole trick for animation consistency.

Installing it

Part of Dream Project Animation Nodes - ComfyUI Manager (search "Dream Project Animation Nodes") or:

cd ComfyUI/custom_nodes
git clone https://github.com/alt-key-project/comfyui-dream-project.git
cd comfyui-dream-project
pip install -r requirements.txt

Restart ComfyUI. Watch the numpy<2.0 pin.

Common issues

The biggest confusion is expecting this to be actual latent noise for a sampler - it isn't, it's an RGB image. You'd feed it through VAE Encode or use it as an image input, not drop it on a latent slot. Also, because it's recursive, high resolution plus high density can get slow - that's normal, drop the size for iteration. And the pack-wide note: the README declares the pack unmaintained (author moved to comfyui-dream-video-batches); this generator still runs fine on current ComfyUI. Need region-specific colors instead of one palette for the whole image? That's Noise from Area Palettes, next in the menu.