Wan 2.1 LoRA · Text to Video

What it is

Wan 2.1 text-to-video with your own adapters bolted on, and unlike its image-to-video sibling there's no first frame to save you. Everything the clip becomes - subject, lens, light, motion - comes out of what you typed. That makes it the hardest of the Wan nodes to get good results from, and the most impressive when it lands, because there's nothing to hide behind.

The practical reason to pick it over a local Wan setup is the same as everywhere in this pack: no 14B checkpoint, no two-expert sampling rig, no VRAM conversation. The practical reason to pick it over Wan 2.2 text-to-video is that 2.1 still has the friendliest LoRA library - the adapters and finetunes the community spent a year building were built on these weights.

How it works

The node is generated from the endpoint's schema and runs entirely on PopNow's hardware. Queue it and the sequence is: quote the job, submit with an idempotency key so a network hiccup can't double-bill, poll, download the MP4, hand control back with a VIDEO value. Feed that into Save Video. The task id output is a plain string, useful when you want to find the run in PopNow's console.

A pleasant side effect of the schema-driven design: because the sampler settings are fields on the endpoint, you get steps, guidance_scale and flow_shift as real widgets rather than a black box.

The inputs that matter

prompt - subject, action, camera. Wan responds to plain prose, not tags, and the single most common failure is describing a still instead of a shot: no verbs, no camera, nothing to animate. Give it a motion arc.

aspect_ratio is only 16:9 or 9:16. That's not a limitation to work around, it's the model telling you it's built for landscape and vertical shorts. Portrait social clips are the second case.

steps (default 30, max 40) and guidance_scale (default 5) trade time for polish. Under about 15 steps things get mushy. flow_shift should stay at -1 unless you're deliberately testing - that hands the model its own default of 3 at 480p and 5 at 720p, and guessing a different value is a hobby, not a workflow.

resolution is 480p or 720p, duration_seconds is 5 or 10, negative_prompt is where you list what must not appear - with no source frame, that field does more work than usual. loras takes up to three Wan 2.1 14B adapters, each from its own PopNow LoRA node with its own scale.

Outputs: video, task id.

Installing it

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

Then a key from popnow.ai/api/console/keys pasted into Settings > PopNow - or POPNOW_API_KEY exported before launching ComfyUI - and one more restart, because the node list is fetched from your account's catalog at startup. Zero model downloads, zero extra pip packages. If the old single-file popnow_api.py is still in custom_nodes, remove it first.

Where it goes wrong

Pricing is by request, not by second: 5s at 480p is the floor, and 10s or 720p each push it up, so 10s 720p is roughly twice the entry rate. Draft cheap, upscale the take you like. And remember the seed is fixed by default - randomising it is a deliberate choice and every randomised run is a billed run.

LoRA trouble is nearly always architecture: these slots want Wan 2.1 14B weights. A 2.2 adapter will load and underperform; an SDXL or Flux adapter will load and do nothing. If a LoRA has trigger words, they go at the front of the prompt or the adapter is wasted.

Finally, the failure codes are boring but they're specific: No PopNow API key is set (no key), PopNow 401 (dead key), 40201 (empty wallet), 40310 (unverified email), content_rejected (the model declined; you're refunded). A 400 naming a field means the service rejected that value, so narrow it one widget at a time.