Wan 2.2 5B LoRA · Text to Video

What it is

Wan 2.2's 5B dense model doing text-to-video, hosted, with a place to hang your adapters. Four widgets and a LoRA socket. No resolution dropdown, no duration slider, no guidance scale - the endpoint is a fixed 720p-class spec and the node tells you so by simply not offering the choice.

That sounds like a downgrade until you look at what it's for. The 5B is the Wan that fits on small cards; locally it's the model people run at 8GB with offloading, and it's the one that made "actually, I can do video" possible for a lot of rigs. As a hosted node with a flat per-run price, it's the cheapest honest way to answer "does this shot work at all?" before you spend real money on A14B.

How it works

Nothing renders on your machine. The node builds a request from your widgets, PopNow quotes it, runs it, and the MP4 comes back decoded as a ComfyUI VIDEO value - into Save Video, ideally, and never into Save Image, which doesn't know what a VIDEO socket is.

The flat per-request price is the thing that changes how you use it. There's no per-second meter quietly ticking up when you pick a longer clip, because you don't pick a length. Every run costs the same, so a bad take costs exactly what a good one does - which argues for drafting in a batch of seeds rather than nursing one prompt.

Uploads and LoRA imports are cached per API key for twelve hours, so iterating doesn't re-send files.

The inputs that matter

prompt is nearly the entire node. With no opening frame and no knobs to compensate, the text has to carry subject, action, camera and light. Describe a shot, not a photograph: what moves, what the camera does, how it ends. Wan's own fans would tell you it's a prose model - sentences with verbs, not a tag list.

aspect_ratio is 16:9 or 9:16, full stop. Landscape and vertical shorts are the supported shapes; don't fight it.

seed - and specifically the control-after-generate setting next to it. It ships fixed, which means re-queueing an unchanged graph returns the first clip instead of buying a new one. Turn it to randomize when you want a different take, and remember that every randomised run is a billed run.

loras accepts up to three adapters, each from a PopNow LoRA node with a source and a scale. Sources can be a .safetensors in ComfyUI/models/loras (uploaded once, reused for twelve hours), a Hugging Face repo id or page, a Civitai page, or any https:// link. These need to be Wan 2.2 TI2V-5B adapters - the 5B size class, not the A14B ones.

Outputs: video and task id.

Installing it

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

Then create a key at popnow.ai/api/console/keys, restart, paste it into Settings > PopNow (or POPNOW_API_KEY in the environment), and restart once more - the node list is pulled from your account's catalog at startup, so a missing node usually means "you haven't restarted since adding the key". There are no weights and no pip dependencies; the pack deliberately imports only what ComfyUI already ships.

Where it goes wrong

Two failure modes account for most of the frustration. The first is asking a 5B model for a ballet: complex action plus complex camera plus a specific character is more than this tier can hold, and the result is a soft, drifting clip. Simplify, then simplify again. The second is a LoRA trained for the big model - separate weight structure, separate adapters, and the wrong one will pass through without an error and change nothing useful.

On the money side: because the price is flat, "it's cheap so I'll try ten" is the actual risk. Set Settings > PopNow > Max credits per run if you're the type to queue and walk away, and remember a cancelled run keeps going server-side and is charged if it succeeds - cancelling stops your wait, not the job.