Wan 2.1 Image LoRA · Text to Image

What it is

Wan 2.1 is a video model that also does stills, and its fans are loud about one property in particular: it looks like photography rather than like illustration. Skin has texture, light behaves, nothing has that glossy AI sheen. That's the pitch of the family, and it's why a Wan image node with a LoRA socket is more interesting than it sounds - the adapters people train for Wan are overwhelmingly photographic ones: a look, a film stock, a body, a specific face.

The trade is control. There's no instruction-following editor here, no negative prompt, no seed tricks beyond the one seed field. Just prompt, framing, size, format, and up to three adapters.

How it works

The pack generates one node per model from the model's published input schema, and runs it on PopNow's servers - the node itself only uploads your prompt, waits, and downloads the result. Each run is quoted before anything is held, so the number on the node badge is what you're agreeing to; failed runs get refunded.

The LoRA mechanics are the part people fumble. You don't attach a file to this node. You add a separate PopNow LoRA node, set source to a .safetensors in ComfyUI/models/loras, a Hugging Face repo id or page, a Civitai page, or a plain https:// link, set a scale, and wire its loras output into this node's loras input. Chain LoRA nodes to stack up to three. Local files upload once and are reused for twelve hours; HF and Civitai sources are imported server-side once, which is how you use a 300MB Civitai adapter without ever owning it.

The inputs that matter

prompt - subject, setting, light. And if your LoRA has trigger tokens, lead with them. That single habit fixes most "the LoRA isn't doing anything" complaints before they happen.

aspect_ratio defaults to 1:1 and runs 1:1 through 21:9 plus 9:21 - a portrait-leaning set, which tells you what the model is good at. resolution is 1k (about a megapixel) or 1.5k (1536 on the long edge). On a per-image node this is the difference between a keepable frame and a thumbnail, and it's cheap.

output_format is jpeg, png or webp. Default jpeg is fine for viewing; use png if you're going to upscale or mask, because jpeg artefacts compound.

loras is optional and, honestly, the reason you'd pick this node over a dozen free local models. Everything else is a plain T2I call. Outputs are images (a batch) and task id - the latter is the string PopNow's API console and support use, so route it into a Note node during a debugging session.

Installing it

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

Either way, then a key: create one at popnow.ai/api/console/keys, restart ComfyUI, paste it under Settings > PopNow (or set POPNOW_API_KEY), restart again. There is nothing else to install - no CUDA, no model files, and the pack's dependency list is empty on purpose. If you still have the old single-file custom_nodes/popnow_api.py, delete it; its node type is the same one the pack ships.

Where it goes wrong

Architecture mismatch is the big one. A Wan 2.1 14B LoRA in a Wan 2.2 context, or an SDXL LoRA anywhere near this, loads without complaint and changes nothing - LoRAs are bound to the architecture they were trained on, and the pack can only pass the file along. Check the adapter's base model before you blame the node.

Stacking is the second: two style LoRAs at full strength produce the usual mush, so add one at a time and drop scale when you combine. The general community guidance of 0.5–0.8 comes from the SDXL era, so treat it as a starting point, not a rule.

Then the plumbing errors, which are all boring and all specific: No PopNow API key is set means exactly that, PopNow 401 is a revoked or mistyped key, 40201 is an empty wallet, and a 400 that names a field is the model rejecting that value - narrow it by changing one widget at a time.