Vantage I2V Dual Model Looper

Wan 2.2 is a mixture-of-experts model, and the split is physical: a high-noise expert handles motion and composition, a low-noise expert handles detail. People who actually run 2.2 seriously feed those two experts as separate models - a high checkpoint and a low checkpoint chained in sequence - and it's a wiring headache every time you rebuild a graph. The Vantage I2V Dual Model Looper is the single looper restructured around exactly that setup: same overlapping-window chaining, same disk-based resume, but two model slots and a three-stage denoise schedule instead of one.

What it is

The niche variant of the pack's core idea. Where the single looper takes one model, this one takes model_high, model_low, and an optional model_init, then carves the denoise pass between them. The pack's own example workflows show what it's for: both "Vantage + Wan 2.2 + Lightning LoRA 8 Steps" and "Vantage + Wan 2.2 GGUF + 20 Steps" feed it two GGUFs (one high, one low) through a LoRA loader. That matches the community consensus the KB records: keep the high-noise pass clean and apply speed LoRAs to the low-noise pass only - this node is a graph that encodes that habit for you.

How it works

Everything about windows, overlap, and resume is shared with the single looper: one prompt line per five-second clip, 81-frame windows at 16 fps (129 at 25 fps), frames saved to ComfyUI/vantage/<project_id>/<n>/, and regeneration from any start_prompt. The difference is the schedule. Total steps are steps_init + steps_high + steps_low. With a model_init hooked up, the init model runs the first steps_init steps at its own cfg_init, the high model continues through steps_high, and the low model finishes the last steps_low and forces full denoise. Without model_init, it's just high then low. Defaults of 1 / 9 / 10 land at 20 steps total, which is why the Lightning LoRA example sits at 8 - you're expected to tune these down with distillation LoRAs.

The inputs that matter

• model_high, model_low - the two Wan 2.2 experts. Which one is which matters; get them swapped and the motion/detail split goes weird.
• model_init (optional) - a third model for the opening steps; skip it if you don't have one.
• project_data - from the Vantage Project Loader; your prompts live there, one line per clip.
• steps_init / steps_high / steps_low, cfg_init, cfg - the split schedule. This is the tuning surface for the whole node.
• clip, vae, start_image plus the usual seed, sampler_name, scheduler, width, height, fps, overlap, crop, denoise, batch_size. Notably clip_vision is optional here, unlike the single looper - but for image-to-video you generally still want it.

The single output, images, feeds a VideoCombine node, same as the single looper.

Installing it

Identical to the rest of the pack: ComfyUI Manager → search VantageLongWanVideo → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/vantagewithai/VantageLongWanVideo.git

then restart. No pip dependencies, no bundled models - bring your own pair of Wan 2.2 checkpoints (GGUF or full).

Where people get burned

This is the least-trafficked node in the pack - effectively zero Google impressions - so there's no body of community write-ups to fall back on; the two example workflow JSONs in the repo are your best reference, and they're both GGUF-based. Model order is the classic mistake: high vs low isn't cosmetic. And if you only have one Wan checkpoint, this node is the wrong tool - the single-model looper is the one you want. Watch the step math too: the defaults assume you're using speed/distillation LoRAs, so at full 20-step quality on a big GGUF, budget accordingly. The disk-and-resume behavior (PNGs piling up in ComfyUI/vantage/, start_prompt clamped to 0 if out of range) is inherited from the single looper, so the same gotchas apply.