Wan 2.2 LoRA · Image to Video

What it is

Wan 2.2 image-to-video with your adapters on board, and - this is the interesting part - with a slot for each of the model's two experts. Wan 2.2's A14B is a mixture of experts: a high-noise pass that decides motion and composition, then a low-noise pass that refines detail. Locally that split has been the source of a year of confusion and one very good piece of community advice. Here it's three fields.

The advice, for context, because it's the reason this node exists in the shape it does: when people loaded speed LoRAs into Wan 2.2 they found composition, lighting, motion and character expression collapsing, along with a distinctly plastic skin. The workable compromise that came out of those threads was to leave the high-noise pass alone and put the adapter on the low-noise pass, where it accelerates detail work without touching the structure. This node lets you do exactly that.

How it works

You give it an opening frame and a motion prompt; PopNow runs both passes on their hardware and returns an MP4 as a VIDEO value for Save Video. The frame uploads once and is cached against your key for twelve hours.

Adapters come from PopNow LoRA nodes - source plus scale, chained into whichever slot you want. A .safetensors sitting in ComfyUI/models/loras is uploaded once and reused; a Hugging Face repo id or page, or a Civitai page, gets imported by PopNow once, which is how you use an adapter you never downloaded.

The inputs that matter

image_url is the opening frame. The clip keeps its aspect ratio, so framing is decided upstream.

prompt - the tooltip here is just "describe the motion to generate," and that's the correct reading. This endpoint wants motion, not a scene description. The frame already establishes everything else.

duration_seconds is 5 or 8 and resolution is 480p or 720p. Note that both of these affect the bill directly, because this node prices per second of output - and 720p is double the 480p rate. That also makes short clips here cheaper than the flat-rate plain Wan 2.2 node, which charges the same per run regardless.

Then the three LoRA inputs, which are the whole point:

• loras - both experts. The default choice, and the right one for a character or style adapter.
• high_noise_loras - high-noise only: motion and composition. Put a camera-move or motion-style adapter here.
• low_noise_loras - low-noise only: detail. This is where a speed or texture adapter belongs if you're using one, and where a grain or skin adapter should go if it's upsetting the layout.

Outputs: video, task id.

Installing it

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

Then a key from popnow.ai/api/console/keys into Settings > PopNow (or POPNOW_API_KEY), and restart so the catalog loads. Nothing to download. Gated adapters on Hugging Face or Civitai need HF_TOKEN or CIVITAI_API_KEY in the environment - the pack reads tokens from the environment only, because widget values get saved into workflows and image metadata.

Where it goes wrong

Putting the same adapter in all three slots is the classic mistake: you don't get three times the style, you get an overloaded stack and often mush. One slot, one adapter, then add.

Wrong base model is the second: these are Wan 2.2 A14B adapters. A 5B adapter from the light model, or a Wan 2.1 one, will be accepted and then underperform quietly.

And a genuine limitation to plan around: 81 frames is Wan 2.2's native context, which is where the "five seconds" comes from. Eight seconds is the endpoint extending itself, not a longer native window, so don't expect the tail of a long clip to hold a face as well as the head does. For anything longer, chain into the Video Extend node with the same adapters so the look carries over.