Fix ControlNet in Forge by testing the signal path

Find the first failed gate: base generation, unit input, detected map, model load, patch, or same-seed output proof. Test one unit with Hires. fix, LoRAs, extra units, batches, postprocessors, and third-party extensions disabled.

A correct preview proves only the preprocessor stage. Confirm Enable, a non-None publisher-documented model for the same map and base family, weight above zero, an active Timestep Range, and a same-seed disabled/enabled comparison.

The integrated interface does not include every control weight. Put the intended weight in the effective models/ControlNet path or configured extra path, use the ControlNet model refresh action, and inspect the scanner or first loader error.

Current Control Type filtering uses filename filters attached to preprocessors. A filename that does not match those filters may not appear in a narrower view. This filtering behavior is navigation, not compatibility proof.

Forge found the file but the registered control-model loaders did not build a patcher from its state dictionary. Verify that it is actually a supported control model for the exact original-Forge build, not another file role or fork-specific package.

Current UI code disables the action and labels it “No ControlNet input image available” when the unit has no input image. Upload or reselect the source for that unit before testing the detector.

Verify the actual unit input and detector first. Change exposed resolution or thresholds one at a time and inspect the new map. For an already processed control map, compare Preprocessor: None rather than preprocessing it again.

The failure is at the detected-map gate. Use a clearer input, verify the intended person is visible, test the relevant OpenPose preprocessor controls one at a time, and do not proceed until the preview contains the joints you need.

Prove the same seed with ControlNet disabled. If that passes, restore one unit and remove Hires. fix, other units, LoRAs, postprocessors, and third-party extensions. Preserve the preview and console to locate the first bad stage.

Input and output aspect ratios differ, and the selected Resize Mode applies a tradeoff. Just Resize can distort; Crop and Resize can remove outer areas; Resize and Fill preserves the full map but adds bands.

No. Pixel Perfect calculates preprocessor resolution from input, output, and resize mode. It does not make aspect ratios equal, undo a crop, remove padding, or make an incompatible model work.

No. Weight changes the strength of a later control stage. Repair the input or preprocessor until the map is correct, then compare weight from a passing full-range baseline.

No. Balanced, My prompt is more important, and ControlNet is more important change priority. They do not change the control model’s architecture, base family, or expected map type.

Current Forge can apply the unit to Both, Low res only, or High res only and logs when a pass is skipped. Prove base resolution first, then record the selected Hires-Fix Option and compare one option at a time.

Img2img adds its own source image, mask, resize, and denoising influence. Prove the same control pair in a simple route, then restore Img2img settings. A dated open TypeError report is evidence for one environment, not a universal cause.

Disable all but one. Prove every unit separately with the same base receipt, then combine two passing units without changing prompt or seed. The first failing combination is the useful boundary.

Do not make a universal claim. The inspected generic loader code does not certify every named FLUX or Union package, while open repository reports include unanswered questions and recognition failures. Require publisher documentation and a controlled test in the exact build.

Do not add a second ControlNet implementation during diagnosis. Original Forge already contains an integrated extension. Advice that edits Forge to force-enable another implementation creates an unsupported comparison and may introduce conflicts.

Community reports describe extension-specific conflicts, but that does not identify yours. Reproduce with all third-party extensions disabled, then enable one at a time while preserving repository and commit for the first extension that restores the symptom.

Record the last passing and first failing original-Forge commits and exact extension commits. Reproduce the same minimal receipt on a separate clean install. That distinguishes a core regression from local accumulated state or an extension conflict.

Control weights use models/ControlNet by default. Downloaded detector assets use the separate models/ControlNetPreprocessor tree. Forge can also use configured extra paths; record the effective path shown by your installation.

It proves the inspected code reached the patch stage for that method. It does not prove the preview is correct, the model is useful for your base family, or the enabled output follows the intended structure.

Include original repository and commit, environment, exact base and control files with sources, unit settings, source and detected map, prompt and fixed seed, full console, disabled/enabled outputs, extension-disabled result, and clean-install result.

Use the workflow page when you have not yet built a first passing pose, Canny, depth, or prepared-map pair. This page starts from a symptom and localizes the first failed diagnostic gate.