SCAIL-2 Reference Pack

SCAIL-2's zero-shot multi-reference is one of the reasons it displaced Wan Animate: no skeleton, no per-character LoRA, just reference images and a driving video, and it can handle several characters in one pass. But "several reference images" turns into a pile of inputs fast, and that's the entire job of SCAIL2ReferencePack. It's the organizer node - it collects your references, arranges them, and hands the model a single tidy package.

You almost never use this node alone. Its one output, reference_pack, plugs straight into SCAIL2SimpleVideo's reference_image socket (that socket accepts either a plain IMAGE or this pack type). The whole thing is optional: single-reference users can skip it and just wire a LoadImage directly. Reach for it when you want more than one subject, or one subject with multiple views.

How it works

The interesting part is what it does with multi-subject packs. It doesn't just bundle images - for multiple subjects it stitches each subject's main image into one primary reference collage, so SCAIL-2 sees all the characters in a single reference frame. That's the "pack" in the name, and it's the bit that would otherwise be fiddly hand-assembly on your end. Extra reference images are passed along separately to provide angles, clothing detail, or story frames, and scene_image optionally guides the background in animation mode.

Internally it builds a versioned reference-pack object (the source calls it version 3) with a subject count, and it only counts subjects that actually have an image connected. Set subject_count higher than the images you wired and it quietly uses what's there.

The inputs that matter

Only two inputs are required, and they're both just counters:

• subject_count - 1 to 6. How many characters you're working with.
• reference_count - 0 to 5. Extra reference images beyond the main ones.

Everything else is optional sockets that appear as you need them: subject_1_image through subject_6_image (one clear, full-body main image per character), reference_1 through reference_5, and scene_image. There's also a set of legacy subject_N_image_M inputs for backward compatibility with the pack's earlier layout - ignore them unless you're opening an old workflow.

Outputs

• reference_pack - a SCAIL2_REFERENCE_PACK object. Wire it into SCAIL2SimpleVideo.reference_image, or into SCAIL-2 Reference SAM Builder first if you're doing the multi-reference workflow properly.

A tip from the README

Order matters. The author suggests making subject 1, 2, 3 correspond to the order the target people appear in the driving video - correspondence stays more stable that way. And start with few images: clear, complete main images beat a pile of noisy angles. The bundled 2. SCAIL2_multi_ref.json workflow shows the whole chain.

Installing and gotchas

It's part of ComfyUI-SCAIL2-Easy, so install the pack once:

cd ComfyUI/custom_nodes
git clone https://github.com/jieg9341-lab/ComfyUI-SCAIL2-Easy

or find it in ComfyUI Manager, then restart. This node itself needs nothing beyond the pack's own light dependencies - no model download, no extra packs. The gotchas live downstream: multi-subject packs require SAM3_VideoTrack driving track data feeding SCAIL2SimpleVideo, and if that track data doesn't contain enough objects the generation node will error with a message telling you to raise max_objects. That's a SAM3_VideoTrack setting, not this node's fault, but it's where people blame the wrong node first.