WanImageMotion Inductive

SVI long-video continuity, but smart about who stays on screen

SVI (Stable Video Infinity) is the error-recycling LoRA that took Wan past the five-second ceiling - the whole community's long-video default. Its weakness was always control: you got length, but identity and continuity across a minute-plus take were a dice roll. The earlier WanImageMotion node in this pack attacked the "slow-motion" problem by boosting motion amplitude. This one is the identity-first sibling: it re-anchors each continuation pass on the anchor image and injects motion residuals from the previous pass, rather than just carrying raw tail latents forward. The result is that the subject stays recognizable while the motion still flows.

How the induction modes differ

induction_mode is the heart of it, and the tooltip is worth quoting:

• inductive_hybrid (default) - anchor image plus motion residuals from prev_samples, mixed with a small raw-tail carry-over. The safe default: continuity without freezing the shot.
• inductive_only - anchor + residuals, no raw tail. Tighter identity lock, but a motion-smoothness tradeoff.
• raw_tail - the classic SVI carry-over. This is the "go back to how everyone else does it" switch when induction feels too stiff.

induction_strength (0–2, default 1) scales how much motion signal gets injected; induction_blend (default 0.85) is the mix ratio between the residual motion and the anchor. Think of strength as "how much movement" and blend as "how much anchor vs motion."

The frame bookkeeping knobs

A stack of slot-based controls manage continuity without hard-coding numbers: motion_latent_count (how many prev frames supply the motion reference), lock_start_slots, end_lock_slots, end_overshoot_slots, and prev_trim_locked_slots (how many previous frames to lock in place rather than regenerate). auto_trim_prev trims the previous pass's tail so you don't double-generate the overlap. use_end_frame plus end_lock_slots gives you an FLF-style end-lock - lock the end of the shot, which is what keeps a looping or matched-clip project from wandering. delta_limit caps how far motion residuals can push a frame, which is your artifact safety valve.

Outputs: updated positive/negative conditioning (with the motion-boosted latents baked in), a latent for sampling, plus trim_slots and prev_trimmed_slots so you can see exactly how much got trimmed this pass.

Using it in practice

Wire it like the other Wan SVI Pro motion nodes: positive/negative conditioning in, anchor_samples (your start frame latent), optional prev_samples and end_samples, length (default 81 frames), and out to the sampler. It lives in the IAMCCS/Wan category of the IAMCCS-nodes pack - ComfyUI Manager ("IAMCCS") or git clone https://github.com/IAMCCS/IAMCCS-nodes.git into custom_nodes, restart.

The honest starting recipe: leave induction_mode on hybrid, induction_strength at 1.0, and only start nudging when you see either face-drift (raise prev_trim_locked_slots or go inductive_only) or stutter (lower induction_strength, add a little raw tail). Where people get burned is pushing induction_strength past ~1.3 and getting warping at the seams - delta_limit exists for exactly that. It's the node you reach for when SVI gives you length but the character stops being the character.