Donut Image Adjust

If you've poked at this pack at all, you've seen the graveyard: DonutGammaCorrection, DonutAutoGamma, DonutAutoWhiteBalance, DonutHistogramStretch, DonutHiRaLoAm, DonutCAS - six image nodes, all marked DEPRECATED. DonutImageAdjust is why. It folds all six operations into a single node behind one operation dropdown, with each operation's parameters namespaced and tucked into the optional block so unused ones never get in your way. This is the node to reach for; the standalone six are kept alive purely so old workflows keep loading.

It's a post-processing workhorse, and the nice thing is how deliberately it's built: every operation reuses the original implementation verbatim - no math was re-implemented, so a local_contrast here is byte-for-byte the old DonutHiRaLoAm.

The operations

• auto_gamma - detects the exposure and corrects it. Methods: simple_mean (fastest), iagcwd, gslf (histogram-based, "most sophisticated"), and percentile (robust to outliers). Add ag_strength, ag_target, and ag_direction (both / brighten_only / darken_only) to steer it.
• gamma - the power-curve midtone fix. gm_gamma from -100 to +100 (positive lifts shadows), plus gm_dither to break up banding.
• white_balance - color-cast removal via wb_method: gray_world, white_patch, combined, auto_levels, or photoshop (per-channel plus midtone snap). wb_strength and wb_clip_percent control the aggressiveness.
• histogram_stretch - auto-levels, with hs_black_point / hs_white_point percentiles and hs_mode (global / per_channel / luminance).
• local_contrast - the Clarity-style boost (lc_radius, lc_amount, lc_protect_tones, lc_fast_blur).
• cas - Contrast Adaptive Sharpening (cas_sharpness, cas_contrast).

The namespacing means you can pre-tune every operation's widgets and switch between them by flipping one dropdown, which is genuinely handy for A/B comparisons on a fixed seed.

The outputs

Two outputs. image is the processed frame. The gamma FLOAT is the diagnostic from auto_gamma - the average gamma it measured, which is genuinely useful feedback when you're dialing in ag_target. For every other operation it's a flat 1.0, so don't build logic on it unless you're in auto_gamma.

Install

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/DonutsDelivery/ComfyUI-DonutNodes.git donutnodes
cd donutnodes
python -m pip install -r requirements.txt

or ComfyUI Manager → "DonutNodes," then restart. No extra models to download - it's pure tensor math, so it runs on the CPU if it has to.

The trap to avoid

The classic post-processing mistake the KB calls out: reaching for img2img to fix contrast when a deterministic pass does it in milliseconds without a seed lottery. These operations are the "proper" fix - gamma, not brightness, for exposure (additive brightness flattens and clips), luminance-mode histogram stretch for flatness, and low-amount local contrast for depth. Chain them: stretch, then gamma, then a touch of CAS, and you've got the whole finishing kit in one node.