Wan 2.2 LoRA · Text to Video

What it is

The top of the Wan 2.2 tree as far as this pack is concerned: text-to-video from the A14B mixture-of-experts model, per-second billing, and three LoRA slots split across the two expert passes. No opening frame to lean on, which means the prompt is doing all the structural work, and no cheap flat rate to hide behind.

It's also the node that lets you replicate the single most useful piece of Wan 2.2 folklore without owning the weights. Wan 2.2 runs a high-noise expert for motion and composition and a low-noise expert for detail. The community discovered that speed LoRAs - the ones everyone reaches for because a 720p clip otherwise takes minutes even on a 4090 - wreck the things that make Wan good: composition, lighting, motion, expressions, and skin that starts looking plastic. The compromise that stuck was to keep the high-noise pass clean and run the adapter on the low-noise pass. That's what high_noise_loras and low_noise_loras are for.

How it works

Queue the node and PopNow quotes the job, runs both expert passes on their hardware, downloads the result, and hands you a VIDEO value. Point it at Save Video. The task id string is what identifies the run in PopNow's console.

Adapters arrive via PopNow LoRA nodes: source can be a .safetensors in ComfyUI/models/loras (uploaded once, reused twelve hours), a Hugging Face repo id or page, a Civitai page, or a plain https link - the remote options are imported server-side once. Up to three per slot.

The inputs that matter

prompt - subject, action, camera. With no first frame, everything structural is decided here, so write a shot rather than a still: verbs, a camera move, a sense of pacing. This is the node where a lazy prompt hurts most.

aspect_ratio is 16:9 or 9:16 only. resolution is 480p or 720p, and 720p is close to double the per-second rate - this is the node's real cost lever, along with duration_seconds at 5 or 8.

negative_prompt matters more here than on the image-to-video sibling: nothing is anchoring the frame, so exclusion is one of your few controls.

The three LoRA inputs:

• loras - both experts, the sane default for a style or character adapter.
• high_noise_loras - motion and composition only.
• low_noise_loras - detail only. A speed or texture adapter goes here if you must use one.

Outputs: video, task id. And yes: seeds default to fixed, so re-queueing an unchanged graph returns the same clip instead of quietly charging you again.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui   # or Manager > search "PopNow"

Create a key at popnow.ai/api/console/keys, restart ComfyUI, paste it into Settings > PopNow, restart once more - the node list comes from your account's catalog at startup. No model files, no GPU, no extra pip packages. On a shared or headless machine, use POPNOW_API_KEY instead of the settings field.

Where it goes wrong

The bill, first. Per-second pricing plus 720p plus eight seconds is the most expensive run in this whole Wan lineup, and it's easy to fire off five of them while "just trying something". Set Settings > PopNow > Max credits per run to a number you're comfortable losing, and do your experimenting at 480p and five seconds.

Second: overloaded LoRA stacks. Nine adapters is a technical maximum, not a target. One adapter, fixed seed, then add the next and watch what changes.

Third: Wan 2.2's native context is 81 frames, which is why five seconds is the default. Pushing duration_seconds to 8 gets you a longer clip, but the model is extrapolating past its window - expect the last second or two to be the weakest.

And if the characters' faces keep changing, drop the speed adapters, raise the quality tier, and lower your ambitions per shot. That's the honest limit of the model, not a misconfiguration.