Image to Noise
Turn a real photo into noise you can seed a diffusion with
- images
- image
Image to Noise takes a real photo and converts it into noise - not random static, but noise derived from that specific image, which the README describes as "useful for init blending or init input to theme a diffusion." The idea: instead of starting a generation from pure Gaussian randomness, you start from noise that still carries some character of a real source image, nudging where the diffusion process lands without going as far as a full img2img pass.
Where this fits conceptually
In a standard img2img setup, you feed a real image in directly and control how much the model deviates from it with a denoise strength - low denoise stays close to the source, high denoise mostly ignores it. Image to Noise sits adjacent to that idea but works differently: rather than handing the model your actual image to partially preserve, it converts the image into noise itself, which you can then use as the starting point for sampling or blend with an existing noise source. It's a subtler kind of influence than img2img - less "keep this composition," more "let the texture and tonal character of this image bias where things start."
How it compares to Image Perlin Noise
The suite also ships Image Perlin Noise, which generates noise procedurally with no relationship to any real image. The two are complementary: reach for Perlin noise when you want organic-but-characterless noise, and reach for Image to Noise when you specifically want the noise to carry some trace of a real photo - its color distribution, its rough tonal layout - into the generation's starting point.
Installing it
Through ComfyUI Manager: search "WAS Node Suite", install, restart. Manually:
cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/was-node-suite-comfyui
cd was-node-suite-comfyui
pip install -r requirements.txt
Pure image processing - no model weights to download for this node.
Troubleshooting
If the effect on your final generation seems negligible, remember this node influences the starting noise, not the sampling process itself - its impact is naturally subtle and easy for a strong prompt or a high number of steps to override entirely. If you want the source image's influence to actually show, you may be better served by conventional img2img with an explicit denoise setting instead, which gives you a much more direct dial.
As with every node in this suite: WAS Node Suite has had no active development since the author stepped back in December 2023. If this node - or the whole pack - fails to import after a ComfyUI core update, that's almost always a stale pinned dependency breaking the suite at startup (BLIP support is the recurring one), not this node's own logic. Check your console log first.
Inputs (7)
| Name | Type | Default | Description |
|---|---|---|---|
| images | IMAGE | The image whose colours the noise is built from. A batch is scrambled frame by frame. | |
| num_colors | INT | 162–256 | How many colours the image is reduced to before its pixels are shuffled. 2 gives noise in two tones, 16 keeps the broad palette, 256 keeps almost every shade of the original. |
| black_mix | INT | 00–20 | How many passes of random black pixels are laid over the noise. Each pass blacks out about half of what is left, so 1 is a coarse half-and-half speckle and 4 is nearly black. 0 adds none. Every pass draws once per pixel, so high values on a large image are slow. |
| gaussian_mix | FLOAT | 0.00–1024 | Radius of a blur mixed back into the noise, in pixels, which softens it into clumps instead of single dots. 0.0 skips it; 2.0 gives soft grain. Most of the original noise is kept whatever the radius. |
| brightness | FLOAT | 1.000–2 | Brightness of the result. 1.0 leaves it as it is, 0.5 halves it, 0.0 is black and 2.0 is twice as bright. |
| output_mode | COMBO | `batch` emits one image holding every scrambled frame, which is what the image nodes expect. `list` emits the frames as a list on the same socket instead, for a downstream node that reads one. | |
| seed | INT | 00–18446744073709550000 | Which shuffle is used. The same seed and the same input always give the same noise; change it to scramble differently. Any whole number; `0` is as good a seed as any. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| image | IMAGE | The scrambled noise, one frame per input frame, with an alpha channel taken from the shuffle. |