Nodes/ComfyUI-CustomNodePacks/Control AOV — Multi-Control Fusion (C2C)
ComfyUI Node

Control AOV — Multi-Control Fusion (C2C)

One node, every ControlNet pass you need — depth, canny, pose, normal, motion, and an ID matte, fused for a union control

By Code2Collapse·Created 9 months ago·Updated about 4 hours ago· 59
Control AOV — Multi-Control Fusion (C2C)
  • image
  • depth
  • canny
  • pose
  • normal
  • id_matte
  • blended
  • combined
  • channel_packed
  • depth
  • canny
  • pose
  • normal
  • motion
  • id_matte
  • info
◄blend_modescreen►
◄preview_layouthorizontal_3►
◄depth_modeloff►
◄normal_modeloff►
◄depth_sizesmall►
◄depth_custom_ckpt(auto)►
◄pose_modeloff►
◄pose_ckpt►
◄id_matte_modeloff►
◄edge_modelinternal_canny►
◄run_cannytrue►
◄canny_low100►
◄canny_high200►
◄canny_aperture3►
◄depth_invertfalse►
◄preproc_resolution512►
◄run_motionfalse►
◄depth_weight1.00►
◄canny_weight1.00►
◄pose_weight1.00►
◄normal_weight0.00►
◄match_toimage►
◄resizeoff►
◄width1024►
◄height1024►
◄scale1.00►
◄divisible_by16►
◄fitcrop►
◄resize_filterlanczos►
◄pad_color#000000►
◄allow_downloadfalse►

ControlNet changed everything about local generation, and the modern version of the game is stacking conditions - depth for spatial layout, canny for edges, pose for structure, normal for surface, motion for video, an ID matte for identity. That's normally a chain of separate preprocessor packs. ControlAOVC2C collapses it into one node that emits every pass at once, plus a channel-packed image (depth=R, canny=G, pose=B) designed for union ControlNets, plus a preview blend. It's the "all the AOVs, in one stop" node of the C2C Control suite.

The mental model is VFX render passes. In a 3D pipeline you render depth, normals, and ID as separate "arbitrary output variables" and recombine them in comp. This node does the same for ControlNet conditioning: each pass comes out of its own output, so you can feed them to separate ControlNets and stack them for maximum spatial lock, or use the convenience outputs to feed a single union control.

How it works

The node is a fusion hub with two philosophies of preprocessing. Canny and optical-flow motion run internally - OpenCV Canny (no model) and optical flow on your frame batch. Everything else is either vendored (runs inside the pack: DepthAnything v1/v2/v3, MiDaS, Depth Pro, DepthCrafter, NormalCrafter, ViTPose) or delegated to comfyui_controlnet_aux (the depth_anything_*, dwpose, openpose, hed, lineart options). You pick per-pass: leave a pass off and wire in an external map (from your own depth node, a SAM matte, whatever) - the node treats those as passthrough AOVs. Weights per pass (0–2) let you tune how hard each condition bites.

The inputs that matter

  • image - source frames; preprocessors run on these.
  • depth_model / normal_model / pose_model / id_matte_model / edge_model - per-pass backends, mostly off by default so you wire in what you need.
  • run_canny + canny_low / canny_high (100/200) - the internal edge pass.
  • run_motion - optical-flow pass, needs ≥2 frames.
  • preproc_resolution (512) - resolution handed to delegated preprocessors.
  • depth_invert - flip near/far when your source convention doesn't match the ControlNet's.
  • The weight inputs (depth_weight, canny_weight, pose_weight, normal_weight).

Outputs: blended, combined (side-by-side/grid preview), channel_packed, and the separate depth, canny, pose, normal, motion, id_matte passes, plus an info string describing per-pass status.

Installing it

Part of Code2Collapse/ComfyUI-CustomNodePacks. ComfyUI Manager → search "CustomNodePacks", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Code2Collapse/ComfyUI-CustomNodePacks.git

Be ready for model downloads: the vendored depth/normal/pose backends pull weights from HuggingFace on first use, and the delegated options need comfyui_controlnet_aux installed. This is the heaviest node in the pack.

Gotchas

First: the delegated options quietly depend on another pack (comfyui_controlnet_aux) - if you pick dwpose or lineart without it, expect a missing-module error. Prefer the vendored backends to stay self-contained. Second, run_motion needs at least two frames or it has nothing to compute - wire a video batch, not a single image. Third, depth_invert matters more than people think: a "far=bright" depth source fed to a ControlNet expecting "near=bright" gives you inverted spatial cues and garbage structure. And finally, the combined preview is for your eyes, not the model - for actual conditioning use the separate passes or channel_packed, not the pretty blend.

Category🐺 C2C/🧰 Core/Control

Inputs (37)

NameTypeDefaultDescription
blend_modeCOMBOscreenHow the 'blended' overlay combines passes. screen = least-destructive; linear_dodge clips; multiply darkens.
preview_layoutCOMBOhorizontal_3Layout for the 'combined' output: horizontal_3 = depth | pose | canny; grid_2x2 = depth|pose // canny|original.
imageoptIMAGESource frames — preprocessors below run on this.
depth_modeloptCOMBOoffDepth backend. VENDORED (run inside this pack): da_v2/da_v1/midas (transformers), da3 (Depth-Anything-3), depth_pro, depthcrafter (video), dvd (DVD — deterministic Wan-2.1 VIDEO depth, EnVision-Research; needs the DVD ckpt in models/DVD/, CC BY-NC). The depth_anything_* options delegate to comfyui_controlnet_aux. 'off' = wire an external depth map.
normal_modeloptCOMBOoffNormal backend (vendored): sobel_from_depth (no model) or normalcrafter (video). 'off' = wire an external normal map.
depth_sizeoptCOMBOsmallDepthAnything backbone: small=ViT-S (fastest, test default) → giant=ViT-G (best). v1 has no giant (falls back to large). Ignored by metric/zoe.
depth_custom_ckptoptCOMBO(auto)Pick your own depth model, or '(auto)' = use the backend/size default.
pose_modeloptCOMBOoffPose backend. vitpose = WanV2 ViTPose chain (uses models/detection/*.onnx). dwpose/openpose/... delegate to comfyui_controlnet_aux.
pose_ckptoptSTRINGCustom ViTPose .onnx pose model filename (none auto-detected; type one or leave blank for default).
id_matte_modeloptCOMBOoffID/segmentation matte. sam_auto = automatic SAM segmentation (controlnet_aux SAMPreprocessor). 'off' = wire an external matte.
edge_modeloptCOMBOinternal_cannyinternal_canny = OpenCV (no model). Others delegate to controlnet_aux.
run_cannyoptBOOLEANtrueUsed only when edge_model = internal_canny.
canny_lowoptINT1000–255—
canny_highoptINT2000–255—
canny_apertureoptCOMBO3Sobel aperture for internal Canny (odd 3/5/7).
depth_invertoptBOOLEANfalseInvert depth (1 - depth). Use when source is 'far=bright' but the ControlNet expects 'near=bright'.
preproc_resolutionoptINT51264–4096Resolution passed to delegated preprocessors.
run_motionoptBOOLEANfalseOptical-flow motion-vector pass (needs an image batch ≥ 2 frames).
depth_weightoptFLOAT1.000–2—
canny_weightoptFLOAT1.000–2—
pose_weightoptFLOAT1.000–2—
normal_weightoptFLOAT0.000–2—
match_tooptCOMBOimageInternal: align every AOV pass to this source's size before resize/output.
resizeoptCOMBOoffResize ALL outputs in-node. off = keep source size; width/height = resize to an exact width x height; scale = multiply size by 'scale'.
widthoptINT10240–16384Target width (resize=width/height). 0 = derive from height + aspect.
heightoptINT10240–16384Target height (resize=width/height). 0 = derive from width + aspect.
scaleoptFLOAT1.000.05–8Scale factor applied to source size (resize=scale).
divisible_byoptINT161–512Snap output W/H to a multiple (VAE-safe: 16=Wan, 64=SD). 1 = no snapping.
fitoptCOMBOcropAspect handling: stretch = distort to fit; pad = letterbox with pad_color; crop = centre-crop to fill.
resize_filteroptCOMBOlanczosResample filter. lanczos = sharpest; nearest = hard pixels.
pad_coloroptSTRING#000000Pad colour (hex), used only when fit=pad.
depthoptIMAGEYour own depth map (Depth Anything, DepthCrafter, ZoeDepth...). Wins over depth_model; depth_invert and depth_weight still apply.
cannyoptIMAGEYour own edge map. Wins over edge_model.
poseoptIMAGEYour own pose render (DWPose / OpenPose / ViTPose). Wins over pose_model.
normaloptIMAGEYour own normal map (NormalCrafter...). Wins over normal_model.
id_matteoptIMAGEYour own ID / segmentation matte (SAM, cryptomatte). Wins over id_matte_model.
allow_downloadoptBOOLEANfalseLet the model backends download missing weights from Hugging Face (some are many GB). Off: only installed copies are used (a models folder or the Hugging Face cache) and the info output says what is missing.

Outputs (10)

NameTypeDescription
blendedIMAGESingle image, all passes OVERLAID via the chosen blend mode.
combinedIMAGESingle image, depth | pose | canny shown SIDE-BY-SIDE (or grid) — the clearest 'all 3 at once' view.
channel_packedIMAGEdepth=R, canny=G, pose=B — lossless packing for union ControlNets.
depthIMAGEDepth AOV (passthrough).
cannyIMAGECanny AOV (run internally if 'image' wired).
poseIMAGEPose AOV (passthrough).
normalIMAGENormal AOV (passthrough).
motionIMAGEMotion-vector AOV (optical flow, run internally on a frame batch).
id_matteIMAGEID / segmentation matte AOV (passthrough).
infoSTRINGPer-pass STATUS + recommended max-control wiring.