Description
The PiD Color Bias Correction node is a post‑processing corrector for models using a 4‑step decoder (e.g., the Flux2 backbone). It addresses a systematic drift observed in such models: dark scenes tend to be brightened and a blue cast is introduced. The node works by subtracting a per‑channel bias from x0_pred at the first sampling step, using a small linear model that was calibrated against this systematic drift.
Inputs
| Name | Type | Default | Range | Description |
|---|---|---|---|---|
model | MODEL | – | – | The diffusion model to be patched. |
strength | FLOAT | 1.0 | -20.0 … 20.0 (step 0.01) | Controls the intensity of the correction. Positive values reduce the brightening and blue cast (recommended for typical drift). Negative values invert the correction (may be used for creative purposes). A value of 0 disables the correction entirely. |
backbone | COMBO | "flux2" | ["flux2"] | The target backbone architecture. Currently only "flux2" is supported. This may be extended in future releases. |
Outputs
| Name | Type | Description |
|---|---|---|
model | MODEL | The input model patched with the color bias correction. The model behaves identically in all other respects; only the initial sampling step is affected. |
Usage Notes
- The node is intended for use with 4‑step decoders (e.g., Flux2). It is unlikely to produce meaningful results with other decoder architectures.
- The default
strengthof 1.0 is calibrated for the typical drift observed in Flux2. You may adjust it slightly (e.g., 0.5–1.5) depending on the prompt or desired output. - Because the correction only modifies the first sampling step, it has negligible impact on inference speed.
- Set
strengthto 0 if you do not want any correction, or if you are using a model variant that does not exhibit the brightening/blue‑cast drift. - This node is part of the KJNodes/experimental category — behavior may change in future versions.
Comments
Sign in with GitHub to join the discussion.