P-Image LoRA · Text to Image

Nine tenths of a credit. That's the headline on P-Image LoRA · Text to Image - just under a cent a render, which is the cheapest thing in the entire PopNow pack and cheap enough to treat renders as free. It's also the pack's only text-to-image node with real width and height fields, which matters more than the price.

What it is

pruna-ai/p-image-lora/text-to-image, one node in the PopNow pack, running on PopNow's hardware. No GPU, no checkpoint, no sampling on your side - prompt in, IMAGE out. P-Image is Pruna's fast image model; the node is sold as fast renders, and at this price the intended use is obvious: draft, then draft again, then draft a fifth time.

That's the honest case for it. A local Flux or SDXL render is "free" after the download but it costs you forty seconds and a chunk of VRAM each; at under a cent, this costs you less than the electricity did.

The inputs

prompt - sentences, not tags. Subject, clothing, place, light, lens. And if you're using a LoRA, the trigger word goes first, exactly like the tooltip says. Trigger words buried in sentence four get ignored.

aspect_ratio defaults to 1:1 with the usual set of ratios, plus one value you don't see on most nodes in this pack: custom. Pick that and the two pixel fields come alive.

width and height are the interesting pair. Both default to -1, which means "leave it to the model", and both accept 256 to 1440 in steps of 16. Steps of 16 isn't arbitrary - it's the VAE's downscale factor, and values that aren't multiples of it produce off-by-a-few crops and seams. If you leave them at -1 with custom selected, the tooltip says 1024 is used.

The trap: width and height only apply when aspect_ratio is custom. Set 1200x800 and leave the ratio on 1:1 and you'll get squares and blame the nodes. This is the single most common complaint about pixel fields generally, and it's a schema behaviour, not a bug.

output_format defaults to jpeg - fine for drafts, switch to png for anything that gets another pass.

seed starts fixed. At 0.9 credits it barely matters, but the default is there so that queueing twice doesn't produce two different pictures.

loras takes one LoRA. Wire a PopNow LoRA node in: source can be a .safetensors in ComfyUI/models/loras (uploaded once and reused for 12 hours), a Hugging Face repo id or page, a Civitai page, or a direct link, with scale for strength. Gated downloads read HF_TOKEN/CIVITAI_API_KEY from the environment only - the pack deliberately refuses to store tokens in widgets, because widget values get baked into workflows and image metadata.

Outputs: images and task id.

Install

Manager → search PopNow → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Then add a key from popnow.ai/api/console/keys in Settings > PopNow, or on a server:

POPNOW_API_KEY=sk-popnow-... python main.py

The pack declares no dependencies and needs no model downloads. Restart once more after the key is in.

Where people get burned

Custom size, wrong ratio. Covered above and worth repeating because it wastes an evening: aspect_ratio has to be custom before width/height mean anything.

It's fast, which means it's not the detail king. For a hero image you'll probably want a bigger model - the pack has plenty, and the Qwen and FLUX nodes in it are the ones people reach for when the draft is approved. Use this for the twenty thumbnails and the layout tests.

A LoRA trained on a different base won't transfer. Prompt-embedded LoRAs are architecture-specific; there's a reason the tooltip says "trained for P-Image text-to-image". The LoRA training doc explains what a LoRA captures and why that doesn't survive a base change.

Cheap still means metered. 0.9 credits a picture adds up to 90 credits over a hundred pictures - nine cents, sure, but the per-run quote and the Max credits per run setting in Settings > PopNow exist for exactly the workflows that queue a hundred.