Wan 2.2 5B LoRA · Image to Video

What it is

The light one. Wan 2.2's 5B dense model - the TI2V-5B branch, the one that fits on an 8GB card locally with ComfyUI's native offloading - hosted, with a LoRA socket and almost nothing else on the node. Prompt, opening frame, seed, adapters. That's the list.

The stripped-down widget set is not laziness. The endpoint is a fixed spec, so there's no resolution dropdown to get wrong and no duration to agonise over; the node's own one-liner calls them "light 720p Wan 2.2 clips". You get the cheap tier of Wan 2.2 for a flat per-run price, and you iterate fast because there's nothing to tune.

Reach for it when you're blocking out a sequence - ten quick shots to find the one worth spending on - or when 480p/720p on a full A14B model is overkill for a background element.

How it works

Your image gets uploaded (PNG, or a high-quality JPEG if it's over the 10MB ceiling), the request goes out, PopNow quotes it, runs it, and the finished MP4 comes back as a VIDEO value. Wire that to Save Video. Uploads are cached per key for twelve hours, so re-running with a new prompt doesn't re-send the frame.

Because the price is per request rather than per second, the economics are simple in a way most video nodes in this pack aren't: the run costs what it costs, whether you like the result or not. So the seed matters. It defaults to fixed, and re-queueing an unchanged graph hands you the same clip back instead of a new charge - flip control-after-generate to randomize only when you want a genuinely different take.

The inputs that matter

image_url is the opening frame, and the clip starts from it and keeps its aspect ratio. If you want a vertical short, feed a vertical still; there is no aspect widget here to override that.

prompt describes what happens after the frame, and how the camera moves. This is where Wan rewards specificity: "she lifts the cup, steam curls, the camera drifts left and settles on the window" gives it something to animate. "Beautiful woman drinking coffee" gives it a slow zoom.

seed is the only other required widget, which makes the whole node feel almost suspiciously simple.

loras is the optional and genuinely valuable input: up to three adapters trained for Wan 2.2 TI2V-5B, each from a PopNow LoRA node with its own scale. Note the size class - 5B adapters, not the A14B two-expert ones people train for the big model. Point it at a file in ComfyUI/models/loras, a Hugging Face repo id or page, a Civitai page, or an https:// link, and PopNow either uploads it once or imports it once. Outputs are video and task id.

Installing it

# Manager > Custom Nodes Manager > search "PopNow" > Install > restart
cd ComfyUI/custom_nodes && git clone https://github.com/popnow-ai/popnow-comfyui

Then the API key: popnow.ai/api/console/keys, pasted into Settings > PopNow, or set POPNOW_API_KEY in the environment. Restart after the key is saved so the node list builds from your account's catalog. No model files, no CUDA, no VRAM - the reason a 5B model is interesting at all is that it's cheap to run, and the reason this node is interesting is that you don't have to run it.

Where it goes wrong

LoRA mismatch is the trap that costs the most time: A14B adapters are not 5B adapters, and the wrong one will be accepted and then do nothing or degrade the clip. Check the adapter's base model before wiring it up.

If the clip drifts away from your opening frame - identity changing, wardrobe mutating - shorten the prompt to motion only, and trim ambition. A 5B model is not going to hold a complex action and a complex camera move at once.

Flat price means no cost surprise, but also no partial refund for a bad take; that's what 40201 (empty wallet) and content_rejected (model declined, refunded) are for, and both are Pack-level, not node-level, problems. If a video looks soft, that's the model tier, not your settings - go up to the full Wan 2.2 nodes when detail matters.