Forge ControlNet Stack

ControlNet is how you tell the model where things go - pose, edges, depth - while the prompt decides what they are. If you're building a ComfyUI graph that actually renders on SD WebUI Forge (through the ForgeSampler node in this same pack), you can't use ComfyUI's native ControlNet nodes; Forge never sees them. This node is the matching adapter: it turns your images and settings into a CN_STACK that ForgeSampler injects into Forge's ControlNet handler when it fires the generation call.

The headline feature is capacity. It exposes twenty units, cn_1 through cn_20, and every one carries the full A1111-era ControlNet toolkit: image, model, preprocessor, weight, start/end guidance, resolution, thresholds, control mode, and resize mode. You'll rarely use more than two or three at once, but the headroom is nice for stacked pose + depth + canny compositions - which is exactly the setup ControlNet was famous for enabling.

How it works. When the node loads, it asks Forge for its model list (/controlnet/model_list) and preprocessor list (/controlnet/module_list) to fill the dropdowns. For each slot that has both an image and a model, it runs the preprocessor on Forge's side: it POSTs your image to Forge's /controlnet/detect endpoint, which returns the processed map (line drawing, depth map, whatever the module produces). Then it packages a per-unit dict - model, weight, the base64 image, processor_res, thresholds, guidance_start/guidance_end, control mode, resize mode - into the stack. ForgeSampler drops that into alwayson_scripts.ControlNet.args on the API call, and Forge does all the ControlNet math. ComfyUI is the remote control, not the engine.

The inputs that matter per slot:

• cn_N_image and cn_N_model - the only two you must set for a unit to activate. No image plus model, no unit, no effect.
• cn_N_preprocessor - default "none". Set it when you're feeding a raw photo and need pose/canny/depth detected; leave it at "none" when you're feeding a ready-made map.
• cn_N_weight - 0 to 2, default 1.0. This is your main dial; if a condition is fighting the prompt, it's usually this.
• cn_N_start / cn_N_end - when during sampling the condition applies (0 to 1). Useful for letting the prompt take over in the final steps.

The rest you can ignore until they bite: resolution (512, the preprocessor's working resolution), threshold_a/threshold_b (100/200 - meaningful mainly for threshold-style modules like canny), control_mode (Balanced / prompt-more-important / ControlNet-more-important), and resize_mode ("Crop and Resize" is the sane default).

Outputs are cn_stack (into ForgeSampler's cn_stack input) plus cn_1_preview through cn_20_preview images of the processed maps. Wire a couple of those into a PreviewImage node. It's free, and it's the best debugging habit in ControlNet: the canny preview should look like an actual line drawing before you spend 30 steps hoping.

Install & gotchas. Same pack, same install - ComfyUI Manager ("ComfyUI-ForgeAPI") or git clone https://github.com/kreonxv/ComfyUI-ForgeAPI into custom_nodes, then restart. No model downloads: the ControlNet models are the ones already in Forge's folder.

Two gotchas worth flagging. First, the /controlnet/* endpoints only exist if Forge actually has ControlNet available (the sd-webui-controlnet extension, which Forge can run). Without it you get empty model lists, and detect fails - the code then silently passes your original image through instead of preprocessing it, so a "none" model with no map does nothing useful. Second, same rule as every node in this pack: Forge must be running and reachable when ComfyUI boots, or the dropdowns come up empty. Start Forge, restart ComfyUI, then build your stack.