Character animation: Wan-Animate-2 vs H3 vs Viggle-Animate

Same four official demo pairs (character image + driving clip), RTX 5090, ComfyUI with Sage attention, seed 42. The models have different contracts, so read the columns accordingly:

• Wan-Animate-2 — distilled int8 · 10 steps · animates the reference in its OWN scene
• H3 ref2va (vanilla) — pruned int8 + TaoMate 3-step · replaces the person in the clip
• Viggle-Animate (3 passes) — H3 ref2va finetune, pruned int8 + DMD r64 · 3 passes (upstream's operating point) · ref = repainted frame 0
• Viggle-Animate (8 steps) — same, full r128 DMD LoRA · 8 steps (as in drbaph's ComfyUI workflow) — sharper on hard inputs

Audio: the driving clip's original soundtrack is muxed onto every output (the robot clip has none). Times are per clip on one RTX 5090, generation only; a clip marked "incl. model load" was the first after loading that model.

Viggle caveat: Viggle expects frame 0 of the driving clip with the character painted in. Open editors (Flux.2 Klein 4B, Qwen-Image-Edit Plus2) only managed that for the werewolf; the other three repaints keep the reference's own background/pose (shown under each pair), which Viggle then carries into the video (kat's jungle, the robot's ghosting). Viggle's own examples use gpt-image-2 for this step.

Werewolf → talk-show host

▶ Play all (synced)❚❚ ⟲ muted

Gothic anime girl → dancer

▶ Play all (synced)❚❚ ⟲ muted

Mech-arm girl → street dancer

▶ Play all (synced)❚❚ ⟲ muted

Toy robot → costume portrait

▶ Play all (synced)❚❚ ⟲ muted