ποΈ Gimbal Channel Band Scaler
Gain sliders for individual latent channels
- latent
- scaled_latent
- telemetry
Every now and then you look at a render and think "this needs more texture" or "the color is too loud" - and you reach for a prompt rewrite, which is like fixing a typo by rewriting the paragraph. GimbalChannelScale is the surgical alternative: it gives you an independent gain slider for each latent channel, so you can amplify the texture-carrying channels or attenuate the saturation channels without touching the rest of the image.
How it works
One latent input, then a set of per-channel gains: ch0_gain, ch1_gain, ch2_gain, ch3_gain, plus remaining_ch_gain which (per the tooltip) applies to channels 4β15 on FLUX/SD3. Every gain defaults to 1.0 (identity) and runs from β5 to +5. The node multiplies each channel's values by its gain - a channel at 1.5 is boosted 50%, a channel at 0 is zeroed out, a negative gain inverts it. Outputs are scaled_latent and a telemetry dict.
The mental map the pack pushes (and it's an approximation, not gospel) is that on SDXL the four channels roughly split into: ch0 luminance/macro lighting, ch1 warm-cool balance, ch2 chroma saturation, ch3 high-frequency detail. So a small boost on ch3_gain is your "add surface detail" dial, and dropping ch2_gain toward 0.8 is your "tone down the saturation" dial. On FLUX there are 16 channels and you're mostly moving remaining_ch_gain, which is blunter - you can't address individual channels 4β15, just all of them at once.
When you'd reach for it
Inside the split/merge loop as an alternative to heavy editing: split, scale one band, merge. Or as a standalone gain knob on a full latent before a refinement sampler - the typical session is "bump ch3 a touch, drop ch2 a notch, re-sample at low denoise." It's a nice debugging tool too: scale a channel to zero and you can see exactly what that channel was contributing.
Installing
Standard Gimbal install: ComfyUI Manager β search "Gimbal-comfy" β install β restart, or git clone https://github.com/FormAndNoise/Gimbal-comfy into ComfyUI/custom_nodes. Dependencies are torch/numpy/pillow, already in ComfyUI - no downloads, no keys. The README's clone URL is stale; FormAndNoise/Gimbal-comfy is correct.
Where people get burned
The traps are the same as any per-channel edit. Aggressive gains (2+ on a single channel) produce latents that drift out of distribution, and you get deep-fried artifacts or color casts instead of "more detail" - keep gains in the 0.7β1.5 range and re-sample at low denoise rather than full strength. And the channel semantics genuinely vary by checkpoint, so the "ch3 = detail" mapping is a starting point, not a law. Test with one channel boosted to 0 and you'll see what that channel actually carries in your model. Also remember this node doesn't know about the split/merge banding - if you want band-style control, pair it with GimbalChannelSplit/Merge. It's a small, honest utility; it won't change your life, but it beats prompt roulette for color and detail nudges.
Inputs (6)
| Name | Type | Default | Description |
|---|---|---|---|
| latent | LATENT | β | |
| ch0_gain | FLOAT | 1.00-5β5 | β |
| ch1_gain | FLOAT | 1.00-5β5 | β |
| ch2_gain | FLOAT | 1.00-5β5 | β |
| ch3_gain | FLOAT | 1.00-5β5 | β |
| remaining_ch_gain | FLOAT | 1.00-5β5 | Applied to channels 4-15 on FLUX/SD3 |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| scaled_latent | LATENT | β |
| telemetry | DICT | β |