Wan 2.2 Image LoRA · Text to Image

What it is

The one node in this batch that gives you somewhere real to put your adapters. Wan 2.2's A14B model is a mixture of experts - a high-noise expert that lays out the frame and a low-noise expert that finishes it - and this text-to-image endpoint exposes that split as three separate LoRA slots: loras for both experts, high_noise_loras for layout only, low_noise_loras for detail only. Three adapters per slot, so up to nine total.

If that sounds like over-engineering for a still image, it isn't, and the reason is the most useful thing the local Wan community learned the hard way. People dumped speed LoRAs into Wan 2.2, found composition, lighting and motion falling apart, and landed on a rule: keep the high-noise pass clean and put the adapter on the low-noise pass only. The same logic applies here. A style LoRA that eats your composition is a low-noise LoRA on this node; a LoRA that changes how the picture is built goes high.

How it works

Nothing runs locally. The node mirrors the endpoint's schema, so the three LoRA slots are real fields, not conveniences. You supply adapters through PopNow LoRA nodes - one per adapter, each with a source (a file in ComfyUI/models/loras, a Hugging Face repo id or page, a Civitai page, or an https:// link) and a scale - and wire them into whichever slot you want. Local files upload once and are reused for twelve hours; HF and Civitai sources get imported server-side once.

When you queue, the request is quoted, run on PopNow's hardware, and the images come back as an IMAGE batch.

The inputs that matter

prompt - subject, place, light, lens. If any adapter has trigger words, lead with them; that one habit prevents most "this LoRA does nothing" outcomes.

aspect_ratio runs 1:1 through 21:9 and back down through 9:21, a tolerant set. output_format is jpeg, png or webp - png if you're feeding a mask or an upscale afterwards. seed is your sameness lever, and it defaults to fixed so that re-queueing doesn't silently buy a second image.

Then the three LoRA inputs, in ascending order of how clever you're being:

• loras - applied to both experts. This is the normal choice, and it's what you want 90% of the time.
• high_noise_loras - high-noise expert only, i.e. layout and composition. Use it when a style adapter keeps wrecking your framing.
• low_noise_loras - detail and texture only. Use it for a grain, a skin or a material adapter that should sit on top rather than restructure the shot.

Outputs: images, task id.

Installing it

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

Key at popnow.ai/api/console/keys → Settings > PopNow (or POPNOW_API_KEY), then restart so the node list rebuilds from your account. Nothing to download, no pip installs; the pack's dependencies list is empty by design. Gated Hugging Face or Civitai adapters need HF_TOKEN or CIVITAI_API_KEY in the environment - the pack never accepts tokens as widgets, because widget values get written into every workflow you share.

Where it goes wrong

The expert slots are a trap for anyone who hasn't met the A14B architecture: putting the same adapter in all three slots triples its influence and usually muddies the output. Pick one slot.

LoRA base mismatch stays the top time sink. Wan 2.1 adapters and Wan 2.2 A14B adapters are different weight structures, and an SDXL adapter isn't in the conversation at all. When in doubt, run the node with no adapters, confirm the base output is good, then add one at a time.

Also worth knowing: stacking is not additive. Community guidance of 0.5–0.8 per LoRA is SDXL-era advice and doesn't transfer cleanly - start at 1.0 on a single adapter and only then combine, lowering each one as you add. And if an image comes back black or noisy, that's the classic signature of an overloaded LoRA stack, not a broken node.