Latent Channel Offset x16 (CCN)

The four-channel story breaks the moment you load a Wan model, because Wan's VAE produces 16 latent channels, not 4. Latent Channel Offset x16 is the same idea as the 4-channel version - add a value to each channel before the sampler - except you get sixteen sliders, ch_00 through ch_15, one per channel.

Here's the uncomfortable truth the node's own docstring leads with: "Channel meanings in 16-channel latents are less documented. Experimentation encouraged!" For SD 1.5/SDXL you have a rough map (channel 0 = cyan-red/brightness, channel 3 = luminance, etc.). For Wan, nobody has published a confident equivalent, and the author isn't going to pretend otherwise. What you're doing with this node is mapping the latent by hand: offset channel 5 by +0.5, generate, see what changed, repeat. It's archaeology on your own model.

How to use it

Wire a latent in, nudge the sixteen ch_XX sliders, get a latent out. Same flow as the 4-channel version, just more knobs. A few practical habits for the mapping grind:

• Change one channel at a time with a fixed seed, and log what you see. With sixteen unknowns you can't afford to move three at once.
• Keep offsets small - ±0.5 is already meaningful; ±2.0 on the wrong channel will wreck a run.
• Use the x16 Scale sibling to multiply channels instead of adding, once you've found a channel that does something you like - multiplication is easier to reason about for "make this axis stronger."

The interplay of two effects - which channels map to which visual properties and how offset vs. scale interacts - is exactly the kind of thing that's undocumented, model-specific, and occasionally delightful. That's the deal with this node: you're not buying a solution, you're buying sixteen experimental dials.

The honest gotchas

Same as the 4-channel version: the node returns a bare {"samples": ...} latent, so noise_mask and other extra keys don't survive it - keep it out of masked/inpainting paths. And remember this offsets before sampling; if you're applying it after, you've already missed the moment it matters. Like every node in this pack's latent family, no model downloads and no dependencies beyond torch.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/valkymaera/ComfyCollectorNodes

restart, or ComfyUI Manager → ComfyCollectorNodes. It shows up as Latent Channel Offset x16 (CCN) under ComfyCollectorNodes/Latent. If you're on SD 1.5/SDXL, the 4-channel version is the one you want; if your latent has 16 channels, this is the only flavor of the idea that touches all of them.