Latent Channel Scale x16 (CCN)
A volume knob for each of Wan's 16 latent channels
- latent
- latent
Latent Channel Scale x16 is the sixteen-channel version of the Scale node: sixteen multiplier sliders, ch_00 through ch_15, each multiplying one channel of a Wan-style latent by a factor. Where the 4-channel Scale works on SD 1.5/SDXL latents, this one exists because Wan's VAE outputs 16 channels - and if you only have four sliders, you can only touch a quarter of them.
The mechanism is identical to the 4-channel version: multiply, don't add. Above 1 amplifies a channel's contribution, below 1 damps it, negative inverts, 0 zeros it out entirely. All sixteen default to 1, which is the identity - a node you've just dropped in does nothing until you move a slider, which is exactly what you want from a tool like this.
The reality of using it
Wan's 16-channel latent layout is essentially undocumented. Nobody has published a reliable "channel 9 controls X" map, and the author isn't pretending otherwise - the docstrings on the offset variants say so plainly. So this node is a mapping tool: pick a channel, multiply it by 0.7 or 1.3, generate on a fixed seed, and see what moved. Sixteen unknowns, one at a time. It's slow, it's boring, and occasionally you find a channel that does something no prompt ever could.
Scale is arguably the better of the two dials for this kind of exploration, because multiplication is easier to reason about than offset: "make this axis 20% louder" maps directly to a factor, whereas an offset's effect depends on the channel's current magnitude. If the Offset x16 node is your exploratory shovel, this is the refinement tool you use once you know which channels matter.
The honest gotchas
Same family behavior as the other channel nodes: it returns a bare {"samples": ...} latent, so noise_mask and friends don't survive a pass through it - keep it out of masked/inpainting paths. And remember the pre-sampler placement: these nodes shape the latent before denoising, so put them between your latent source and the sampler, not after. If you're on a 4-channel model, the plain Latent Channel Scale is the version you want; this one is Wan-specific.
Install
cd ComfyUI/custom_nodes
git clone https://github.com/valkymaera/ComfyCollectorNodes
restart, or ComfyUI Manager → ComfyCollectorNodes. It appears as Latent Channel Scale x16 (CCN) under ComfyCollectorNodes/Latent. Pure torch, no models, no extra dependencies - sixteen tiny sliders and a whole world of undocumented model behavior to explore.
Inputs (17)
| Name | Type | Default | Description |
|---|---|---|---|
| latent | LATENT | — | |
| ch_00 | FLOAT | 1.00 | — |
| ch_01 | FLOAT | 1.00 | — |
| ch_02 | FLOAT | 1.00 | — |
| ch_03 | FLOAT | 1.00 | — |
| ch_04 | FLOAT | 1.00 | — |
| ch_05 | FLOAT | 1.00 | — |
| ch_06 | FLOAT | 1.00 | — |
| ch_07 | FLOAT | 1.00 | — |
| ch_08 | FLOAT | 1.00 | — |
| ch_09 | FLOAT | 1.00 | — |
| ch_10 | FLOAT | 1.00 | — |
| ch_11 | FLOAT | 1.00 | — |
| ch_12 | FLOAT | 1.00 | — |
| ch_13 | FLOAT | 1.00 | — |
| ch_14 | FLOAT | 1.00 | — |
| ch_15 | FLOAT | 1.00 | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| latent | LATENT | — |