deepDeband (Banding Removal)

Every smoothing filter can make banding disappear - but a good one knows banding from sky. That distinction is the whole reason deepDeband exists. Where the pack's classical node applies generic math (blur the flats, add dither), deepDeband is a trained model from a real ICIP 2022 paper ("Deep Gradient-Domain Image Debanding," Raymond Zhou et al.), taught on 51,490 pairs of pristine and banded 256×256 image patches. It's seen what banding looks like, and it removes exactly that, rather than softening everything to hide it.

So when do you use it? When you have actual gradient banding - skies, studio backdrops, film grain beds, color-graded footage - and the classical node isn't clean enough. It's the middle option in this pack: more powerful than the math, less setup than ABCD, and it works on both 8-bit and 16-bit input in ComfyUI's usual float32 [0, 1] space. One caveat the author flags hard: it was trained on real images and video frames. Feed it a synthetic gradient, a 3D render, or a vector graphic and it can introduce banding instead of removing it. Plan accordingly.

How it works

Read the node source and the pipeline is a tidy deep-learning affair. The image gets reflection-padded up to a 256-multiple canvas (a neat mirror-tiling trick so edges don't get mangled), pushed through a U-Net in [-1, 1] normalized space, and reassembled. Two model variants pick the strategy:

• deepDeband-w (default, recommended) - processes overlapping tiles and fuses them with weighted bilateral-style blending masks. Smoothest result, no visible seams. This is the one the paper recommends.
• deepDeband-f - processes patches directly, faster, but on big images you can see tile boundaries.

The strength slider then blends the result against your original: 1.0 is full debanding, and the tooltip's advice to drop to 0.3–0.7 when you want to keep intentional gradients is exactly right - sometimes the banding is mild and you don't want the model to "fix" a sky that's supposed to look that way.

The inputs that matter

• model - deepDeband-w unless you're previewing speed. Leave it.
• strength (0.0–1.0, default 1.0) - start at 0.7 for footage, 1.0 for stubborn banding. Back off if detail starts to smear.
• tile_size (128–512, default 256) and tile_overlap (0–128, default 128) - these only matter for images bigger than the tile. The model was trained on 256px patches, so 256 is the sweet spot; the default 128px overlap matches the original deepDeband-w's 50% and is the right quality setting. If memory is tight, drop overlap toward 64.

One honest note from the source: the node's description claims it detects and preserves your input bit depth, but the detection routine in the code just returns "8-bit" for everything - everything runs through ComfyUI's float32 [0, 1] space anyway. In practice it doesn't matter much, because the pack's Save16BitTIFF writes whatever precision it receives. Just know the "automatic 16-bit preservation" in the tooltip is aspirational, not magical.

Installing it and the weights

The node ships in the Bit Depth Enhancer pack. ComfyUI Manager (search "Bit Depth Enhancer") or:

cd ComfyUI/custom_nodes
git clone https://github.com/subraoul/ComfyUI_Bit-Depth-Enhancer.git
cd ComfyUI_Bit-Depth-Enhancer
pip install -e .

Restart ComfyUI, then fetch the checkpoints from the original deepDeband GitHub repo (under pytorch-CycleGAN-and-pix2pix/checkpoints/):

ComfyUI/models/bit_depth_enhancement/deepdeband/
├── deepDeband_w.pth   # from deepDeband-w/latest_net_G.pth
└── deepDeband_f.pth   # from deepDeband-f/latest_net_G.pth

You only need the generator weights (latest_net_G.pth) - skip the discriminator. Rename them to match exactly. Unlike ABCD, this node uses ComfyUI's proper models_dir lookup, so portable installs are fine. If the files are missing the node just logs an error and returns your image unchanged, which is polite but easy to miss - check the console.

Common issues

• Model not found - wrong filename (deepDeband_w.pth, note the underscore) or the folder doesn't exist. Create models/bit_depth_enhancement/deepdeband/ yourself; ComfyUI won't.
• Artifacts on renders - the synthetic-image warning is real. Don't deband a 3D render with this; use the classical node there.
• Seams on large images - you dropped the overlap, or you're on deepDeband-f. Raise tile_overlap back to 128.
• Looks mushy - strength too high on fine detail. Drop to 0.5 and blend.
• On Apple Silicon it'll happily use MPS, and there's a CPU fallback if your GPU chokes.

Wire debanded_image into Save16BitTIFF (or any save node) and you're done. For real-world footage banding, this is the node in the pack that actually knows what it's doing.