DARASK Anima Sampling Tuner

If you've come over from Forge (or from one of the half-dozen Anima tutorials that quietly assume you're on Forge), you've hit this wall: Forge Classic Neo has a "Shift" slider that Anima wants set to 3.0, Wan 2.2 wants 5.0, and ComfyUI has the same machinery buried in ModelSamplingSD3, ModelSamplingAuraFlow and ModelSamplingFlux nodes you have to know to go looking for. The DARASK Anima Sampling Tuner exists to end that scavenger hunt. It bundles the four sampling tweaks people actually use on Anima/Wan/SD3-flow - shift, temporal score rescaling, epsilon scaling, and CFG rescaling - into one node with one set of sliders, using the Forge preset values as defaults.

How it works

All of these are patches applied to your model before it hits the sampler. Shift calls ModelSamplingDiscreteFlow.set_parameters() with a shift value - the exact same call Forge Classic Neo makes under the hood for its Anima slider, which is why the "shift" concept transfers cleanly (see the haoming02 panel for the Forge line). TSR (Temporal Score Rescaling) rescales the noise prediction for video models, which reads as better motion coherence on Anima. Epsilon scaling nudges exposure bias with a subtle post-CFG rescale, 1.0–1.05, where 1.0 is a no-op. The CFG section is a radio: Rescale CFG or CFG Zero Star, because both replace the sampler's sampler_cfg_function and only one can sit in that slot. The other three are post-CFG hooks, so they stack fine with each other and with other patchers.

The architecture-aware multiplier is the subtle bit. Anima, Wan and LTX use a native multiplier of 1.0; SD3 uses 1000. shift_multiplier defaults to 0, meaning auto-detect from the source model - leave it alone.

The inputs that matter

• model - your MODEL. Wire from the checkpoint/UNET loader.
• shift_enable + shift + shift_style - the main event. Anima default is 3.0; bump to 5–8 for stronger motion consistency, drop toward 3 for sharper stills. Style selects which flow-family sampling to patch (SD3/Anima/Wan, AuraFlow, or Flux - Flux uses the resolution-aware max/base shifts instead).
• tsr_enable - turn this on for Anima/Wan; tsr_k and tsr_sigma are usually fine at defaults.
• cfg_mode - off until you know you want Rescale CFG or CFG Zero Star.

Outputs: a patched MODEL (straight into your KSampler) and an info string that tells you exactly what got activated - handy for a Show Text log.

Install

This ships in the DARASK pack: ComfyUI Manager → search DARASK, or install via Git URL https://github.com/daraskme/comfyui_darask_node, then restart. No extra models required for this node - it's pure sampler math.

Gotchas

The big one is the sandstorm. If your Anima output comes out as colored static/mosaic, check that shift_multiplier is 0 (auto). A saved workflow from an early version of the pack may carry 1000 - the SD3 value - which compresses sigma to near-zero on Anima and produces exactly that noise. If you're stacking this with the DARASK Anima Step Cache, both work together (this one is a CFG hook + model sampling patch, the cache is a UNet wrapper); just make sure your Lora Loader's model input is actually connected, or you'll see the classic 'NoneType' object has no attribute 'clone' error downstream.