WanImageMotionPro Plus

Plain WanImageMotion fixes floaty SVI output. This is the Pro version: everything the plain node does, plus an end_samples end-lock so your clip converges onto a specific final frame - which is what turns a loop or a continuation into an actual A→B transition. The KB's SVI panel credits IAMCCS-nodes with adding exactly this first-frame/last-frame (FLF) support to the SVI ecosystem back in March 2026; this node is where that lives in its current form.

The core controls

The first ten or so inputs are shared with WanImageMotion - positive/negative, length, anchor_samples, motion (1.0–2.0), motion_mode, vram_profile, safety_preset. Then it diverges:

• end_samples (optional LATENT) - encode your target final image and wire it here. use_end_frame defaults true; flip it false to safely bypass the end-frame encoder without unplugging, which prevents ping-pong artifacts.
• end_lock_slots (default 1) - how many final latent slots are hard-locked to the end frame. 1 ≈ the last ~4 video frames. Set 16 for 1:1 parity with the original FLF node.
• end_overshoot_slots (0–8, recommended 1) - extends the internal generation window so the clip converges toward the end frame instead of hard-freezing on it. If the end frame looks stamped-on, this is the dial. Known trap from the tooltip: with KJNodes WAN attention optimizations (FETA / wrapped attention) enabled, overshoot > 0 can crash with an einops shape mismatch - set it to 0 or disable that optimization.
• clip_vision_mode - auto concatenates start+end tokens (correct only for true A→B); end_only is recommended for SVI Pro continuation where anchor=end=same image, because it stops crossfade bleed.
• latent_refresh / delta_max - the moment-stabilization pair for long chains. Start latent_refresh at 0.5 for 3+ segments; cap over-correction with delta_max 0.5–1.0.
• continuity_profile - raw-like paths for SVI chains that keep FLF semantics without motion shaping. Most people can leave it off and let the explicit controls do the work.

There's also a preset widget (default [custom]) with prebuilt combos like parity 1:1, flf overshoot, and the SVI continuity family - useful as starting points, then you tune individual widgets.

Outputs and where it goes

Out comes edited positive/negative, the latent to sample, and - new compared to the plain node - trim_slots. That's the number of latent slots the overshoot added, which you subtract after sampling so your visible clip lands on the intended length. Wire trim_slots into a trim/crop node and the math stays honest.

Honest note: this is a lot of dials, and end_transition_frames is explicitly deprecated - keep it at 0, because blending padding latents toward the end frame in latent space produces static/frozen output (VAE latent interpolation isn't linear). If you want the exact original FLF+SVI behavior to benchmark against, raw_mode gives you a 1:1 reconstruction of it.

Install

Part of IAMCCS-nodes:

• ComfyUI Manager → search "IAMCCS" → install → restart.
• Or:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

Restart ComfyUI. Requirements: ComfyUI ≥ 0.3.0, Python ≥ 3.12, Torch ≥ 2.8. Runs on top of your existing WAN 2.2 + SVI LoRA setup.