RescaleCFG [Multival] 🎭🅐🅓

Displayed as RescaleCFG [Multival], this patches RescaleCFG directly onto your model - the sibling to ADE_CFGExtrasRescaleCFG, which does the same job but chained through this pack's Sample Settings system instead.

What RescaleCFG fixes

Classifier-free guidance extrapolates away from the unconditional prediction toward the conditional one, and the further you push that extrapolation with higher CFG, the more the result's variance drifts beyond what the model was actually trained to produce - the mechanical cause of the overexposed, oversaturated look that shows up when you crank CFG to fight weak prompt adherence. RescaleCFG measures that drift and pulls the final prediction back toward the unguided prediction's own statistical range, keeping the extra push CFG buys you without the blown-out contrast. It's a general SD technique, not AnimateDiff-specific - ComfyUI has its own native RescaleCFG node, and other ecosystems reach for the same fix when pairing high CFG with newer sampler families.

This node vs. the CFG Extras version

ADE_CFGExtrasRescaleCFG returns a CFG_EXTRAS object meant to chain into Sample Settings, composable with other guidance adjustments (PAG included) via prev_extras. This node skips that plumbing entirely: it takes MODEL directly and hands back a patched MODEL, no Sample Settings object required - the fastest path if RescaleCFG is the only guidance adjustment you're making. The one real difference from a plain, stock RescaleCFG node is that mult here is a MULTIVAL rather than a bare float, so it can be masked or given spatial variation, though most workflows will just plug in one flat value.

Use this node for a straightforward, standalone fix. Use the CFG_EXTRAS version once you're already stacking PAG, custom CFG keyframes, or other Sample Settings features and want RescaleCFG to combine with them in one chain.

Inputs and output

Both required:

• model (MODEL) - the model to patch.
• mult_multival (MULTIVAL) - how strongly the rescale pulls toward the unguided prediction's range. Build it with ADE_MultivalDynamic; 0 is effectively off, higher values pull harder.

Output is MODEL, the patched version - wire it into your KSampler.

How to install it

ComfyUI Manager: search AnimateDiff Evolved, confirm the author is Kosinkadink, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/Kosinkadink/ComfyUI-AnimateDiff-Evolved

No model download for this node - it's a sampling-time adjustment, not a weights file. The rest of your AnimateDiff graph still needs its own motion module, of course.

Common issues & troubleshooting

Output goes flat or muted instead of just less blown-out. Rescale too aggressively and you pull the result too far toward the unguided prediction's dynamic range, losing the punch you raised CFG for in the first place. Start moderate and adjust from there rather than maxing it out.

You're not actually running high CFG. RescaleCFG solves a specific artifact - overexposure from a strong CFG push. If your CFG is already modest and your images look fine, this node has nothing to fix; adding it won't generally "improve" output.

Using this and ADE_CFGExtrasRescaleCFG together. Redundant - you're applying the same correction twice. Pick one path rather than stacking both.

No visible change. Confirm the patched MODEL output actually reaches your KSampler rather than an unpatched model wired in by mistake elsewhere.