Wan 2.1 Image LoRA · Edit

What it is

Wan is Alibaba's video model, but the 14B transformer does stills too, and that's what this node is: the Wan 2.1 image model, hosted, with a strength control and a LoRA socket. The tooltip on image_url is the important bit - "the result keeps its aspect ratio" - because this is closer to a guided re-render than to an instruction editor. You hand it a picture, you describe what you want back, and strength decides how far the model is allowed to walk away from what you gave it.

Reach for it when you want a restyle that stays compositionally honest: a photo turned illustration, a rough sketch turned photoreal, a product shot re-lit into a different mood. If you want "move this person's arm and nothing else", that's an editor's job, not this.

How it works

Like every model node in the pack, the widgets are generated from the endpoint's own schema, so there's nothing here you can't send. Behind the queue button: your image is uploaded (PNG if it fits under 10MB, high-quality JPEG if not - the pack re-encodes rather than refusing), the request goes out with your settings, PopNow quotes it, runs it, and the result is downloaded and decoded into an IMAGE batch. Uploads are cached per key for twelve hours, so iterating on the prompt doesn't re-upload the same photo every run.

LoRAs come in sideways. Add a PopNow LoRA node, set source and scale, plug it into loras. Up to three, each with its own strength. A local .safetensors from ComfyUI/models/loras is uploaded once and reused; a Hugging Face repo id or Civitai page is imported by PopNow once.

The inputs that matter

prompt - describe the picture you want returned, not the delta. "An oil painting of the same woman on the same balcony, thick impasto strokes, warm dusk light" works; "make it a painting" leaves the model with too little to aim at.

strength defaults to 0.6 on a 0–1 scale, and it's the only slider worth fiddling with first. Low values keep the source: identity, layout, texture. High values rebuild the scene from your prompt and treat the input as a vibe. If output looks untouched, raise it; if it looks like a different photo, lower it. Changing anything else while leaving this alone usually doesn't fix anything.

resolution is 1k (1024px) or 1.5k (1536px) on the width, with height following your source's shape - you don't pick a height, which is a nice guard against the stretched faces people get from free-form aspect ratios. output_format is jpeg, png or webp; use png when the result feeds another stage.

loras is where the personality lives - a style LoRA does more here than five extra sentences of prompt. Outputs are images and task id.

Installing it

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

Restart, then paste an API key from popnow.ai/api/console/keys into Settings > PopNow - or set POPNOW_API_KEY if this box is remote - and restart once more so the node list comes from your account. No weights, no pip installs, nothing in models/.

Where it goes wrong

The LoRA mismatch is the classic. Wan 2.1 14B LoRAs share a family with Wan 2.2, but "shares a family" is not "works": cross-revision LoRAs are reported to transfer only partially and sometimes make results worse than no LoRA at all. An SDXL or Flux LoRA will load and simply do nothing. Add one adapter at a time at a fixed seed until you know what it does.

If the result is a different person, strength is too high or the prompt describes someone the model has no reason to preserve. If the image comes back soft, it's the 1k tier - pay for 1.5k. And the pack's usual suspects apply: 401 is a bad key, 40201 is an empty wallet, 402-adjacent field errors name the offending value.

One last thing worth internalising: seeds default to fixed. Switch control-after-generate to randomize when you actually want a new roll, otherwise you'll re-queue and get the same image back wondering why nothing changed.