PIA Input [Paper Presets] 🎭🅐🅓②

The sibling node in this pair, PIA Input [Multival], makes you build your own strength value from scratch. This one skips that: pick one of the canned presets the PIA paper's authors themselves defined, and get a working pia_input without having to guess at numbers. If you're new to PIA (Personalized Image Animator) and just want to see it work before you start hand-tuning, this is the node to start on.

What it's doing

PIA's condition module was trained around specific affinity-hint magnitudes - how strongly a frame should stick to the reference image versus depart from it. The paper's own preset values encode a few of those tested settings directly, so picking one here gives you a starting point the original authors validated, rather than a number you picked cold. It's the same relationship "recommended settings" has to "advanced settings" anywhere else: use the preset until you have a specific reason to deviate from it.

The inputs and outputs that matter

Required: preset, a dropdown of the paper's named presets, and batch_index (default 0) - which entry in your batch this preset value applies to, relevant if you're generating more than one PIA animation at once from the same graph.

Optional: mult_multival (MULTIVAL) - a multiplier on top of the chosen preset, letting you push the preset's strength up or down without abandoning it entirely for a fully manual value, and print_values (default false) - flip it on and the node prints the actual resolved numbers the preset maps to, useful the first time you're trying to understand what a given preset actually does before committing to a long render.

Output is a single PIA_INPUT, which feeds the pia_input slot on Apply AnimateDiff-PIA Model, alongside your motion model, reference image, and VAE.

How to install it

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

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

then restart. No model download tied to this node specifically - the PIA checkpoint (pia.ckpt, from Leoxing's HuggingFace repo) is injected earlier in the chain by Inject PIA into AnimateDiff Model, not built here.

Common issues & troubleshooting

All the presets look and behave about the same. Turn on print_values and actually look at the numbers - it's easy to assume presets are more different than they are without checking, and the printout tells you exactly what you're comparing rather than guessing from output alone.

A preset felt close but not quite right. That's what mult_multival is for - nudge the chosen preset up or down rather than throwing it out and switching to fully manual values on PIA Input [Multival]. It's a smaller, more targeted adjustment than rebuilding the strength value from scratch.

Motion looks broken regardless of which preset you pick. Before blaming the preset choice, check the more fundamental PIA requirements: the autoselect or sqrt_linear (AnimateDiff) beta_schedule the README calls out as mandatory, and that PIA was actually injected into your motion model upstream. A preset can't fix a pipeline that's missing those.

Barely any community discussion of what these presets actually look like in practice. Correct, and worth expecting going in - PIA is one of the more obscure corners of an already-niche pack, so you're mostly working from the README, the paper it's based on, and your own test renders rather than a well-worn community recipe.