Use ControlNet in Forge without guessing

Yes. The inspected original Forge commit contains sd_forge_controlnet under extensions-builtin and labels the accordion ControlNet Integrated. That interface does not bundle every ControlNet weight file.

Not for the integrated workflow described here. Installing another ControlNet extension adds a second code path and makes diagnosis harder. First prove the built-in original-Forge unit.

The default weights directory is models/ControlNet. Forge can also scan the ControlNet extra path in Settings or a path supplied with --controlnet-dir. Preprocessor downloads use a separate models/ControlNetPreprocessor directory.

The integrated interface is present even when no compatible weights have been installed. Put the publisher’s control weights in the effective ControlNet directory, refresh that model inventory, and inspect the first console error if the file remains absent.

The preprocessor derives a control map such as edges, pose, or depth from an image. The ControlNet model consumes that map during diffusion. A useful workflow needs the map type, control model, and base checkpoint family to agree.

Use a model whose publisher explicitly names both the map type you need and the family of your active base checkpoint. This page does not mirror binaries or infer compatibility from filenames.

Do not assume so. Treat the base checkpoint family and control-model family as a required match unless the control-model publisher explicitly documents a different supported pairing.

The current repository contains ControlNet infrastructure, but old tutorials and community requests do not prove that any arbitrary FLUX control model works in the integrated unit. Follow the exact model publisher’s requirements and prove one fixed-seed A/B run; otherwise mark compatibility unknown.

Use a clear source image, select an OpenPose-family preprocessor and a matching control model for the base family, preview the skeleton, correct detection failures, then compare one enabled and disabled generation with the same seed.

Select Canny, choose a Canny control model for the active base family, preview the edge map, and check that important contours are present without excessive noise. Start at weight 1, full range, Balanced, and one unit.

Select a depth preprocessor and matching depth control model, preview near-to-far separation, and verify that the broad spatial layout is correct. Depth control is not a promise to preserve fine contours or identity.

Choose Preprocessor: None when the image is already in the exact map format expected by the selected control model. Running another detector can replace or corrupt the signal you prepared.

It computes preprocessor resolution from the control image, output dimensions, and Resize Mode. It does not make a mismatched control model compatible and does not guarantee that the detected content is correct.

It is a useful first baseline for image-derived maps. Disable it only when you have a reason to set preprocessor resolution manually and can compare the resulting detected maps.

Turn on Allow Preview, choose the preprocessor, and run its preview action. Inspect the detected map before Generate. If it encodes the wrong structure, repair the input or preprocessor settings first.

Start from the inspected default of 1. Compare the same seed before changing it. Higher values can force the map more strongly but cannot fix a wrong preprocessor, incompatible model, or bad resize choice.

It sets the portion of denoising during which the control applies. The inspected default is 0–1. Keep the full range until the pair works, then change one endpoint at a time for an attributable comparison.

Start with Balanced. “My prompt is more important” and “ControlNet is more important” adjust priority between prompt and control; neither fixes model-family or map-type mismatches.

Use Just Resize when aspect ratios match; Crop and Resize when filling the canvas matters and center-cropping is safe; Resize and Fill when preserving the whole control map matters more than avoiding padded bands.

Just Resize changes the map to the output width and height directly. If ratios differ, use a deliberate crop/fill mode or make the first test at a matching ratio. Record the final map and output dimensions.

Yes. The integrated script handles Txt2img and Img2img, but Img2img adds its own source-image and denoising influence. Prove ControlNet in one simple route first, then add the second influence without changing the seed.

Yes. The inspected setting defaults to three units and can be configured from one to ten with a restart. Test each unit alone before combining them so you know which map causes a change or failure.

When Hires. fix is active, Forge exposes Both, Low res only, and High res only. Leave Hires. fix off for the first proof; then restore it and compare the same seed and target size with one option at a time.

The inspected UI includes Single Image, Batch Folder, and Batch Upload modes. Batch adds ordering, seeds, dimensions, and file matching to the receipt, so prove a single image before scaling the workflow.

The common gates are Enable, a valid input, a useful detected map, a non-None compatible model, weight above zero, an active timestep range, and an A/B comparison that changes only the unit. If all pass, preserve the console and exact files for a reproducible report.

Include original Forge commit, environment, base and control model filenames and sources, preprocessor, input and detected map, all unit controls, generation receipt, console trace, and same-seed enabled/disabled outputs. State whether a clean install reproduces it.

No. Current code filters model names using preprocessor filename tags. A filtered dropdown is a convenience; only the model publisher’s contract and a completed controlled test establish support.

Not by itself. Pose controls body arrangement, Canny controls edges, and depth controls spatial layering. Identity and styling require the base model, prompt, and possibly separate compatible methods, each proved independently.