Neta Lumina · Text to Image

Four inputs, and there's nothing to tune. This is the leanest node in the whole pack, and it's also the cheapest per image at 1.8 credits - a fifth of a cent. If you're generating anime and you don't want to babysit a local checkpoint, its LoRAs and its sampler settings, this is the shortest path from a tag list to a finished picture.

What it is

Neta Lumina is an anime fine-tune of Lumina-Image-2.0 by Neta.art Lab, released mid-2025. Worth knowing what it was built for, because the prompt style follows from it: it's tuned across the anime-adjacent territory - furry, guofeng (traditional Chinese aesthetics), pets, and a wide spread of characters and styles from mainstream to obscure - while keeping danbooru tag support and strong natural-language adherence. Its own release notes push native multilingual prompts, with Chinese, English and Japanese first.

The node wraps the hosted endpoint in the PopNow pack, so nothing is downloaded and no GPU is involved. That last part is the actual argument: anime models are the one place where local still has a deep bag of tricks (LoRAs, ControlNets, inpainting), and if what you want is a clean 1024px illustration rather than a whole toolkit, paying 1.8 credits beats building a checkpoint folder.

How it works

You send a prompt, an aspect ratio, a seed and an output format. PopNow renders and the node returns an IMAGE batch plus a task id. That's genuinely all - there's no negative prompt, no steps, no CFG, no sampler. The model's defaults are its own, and the pack deliberately omits fields that were set to "model default" rather than inventing sliders for them.

Seeds behave the way they should: control after generate is fixed, so queueing the same graph twice returns the same image instead of charging you for a near-identical one. Set it to randomize when you're rolling for a good composition.

The four inputs

• prompt - danbooru-style tags, natural language, or both. This model handles a sentence better than most tag-trained anime checkpoints do, so "a girl in a rain-soaked Tokyo alley, looking up at a neon sign, cinematic lighting" works as well as 1girl, solo, rain, neon, night, looking up.
• aspect_ratio - 1:1 by default, with 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9 and 9:21. Vertical and ultrawide both work; pick before generating, cropping anime art afterwards tends to cut faces.
• output_format - jpeg by default; png and webp are available. Use png if you're going to upscale or composite, because JPEG artefacts on line art are unforgiving.
• seed - as above.

Wire images to Save Image or Preview Image. Nothing else to connect.

Prompting it well

Two habits get you a lot out of this model. First, front-load subject and framing: who, doing what, shot type, then style and lighting. Second, remember there's no negative prompt - "no text" has to become "clean background, no signage" phrased positively, or you'll get the sign.

For style control, a tag pile still works and is still the fastest route to a specific look. Mixing registers is fine - the model was trained on both - but keep the tag block and the sentence about different things. A natural-language paragraph describing the pose plus a tag block for style quality usually beats either alone.

Install

Manager → search PopNow → Install, or:

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

Restart, then add a key from popnow.ai/api/console/keys in Settings > PopNow and restart once more. No Python dependencies - the pack declares none, and it needs only ComfyUI's V3 node API (Python 3.10+).

Troubleshooting

Because there are so few knobs, most failures are on the account side. No PopNow API key is set means the key didn't save. PopNow 40201 means no credits - at 1.8 per image you'd need to queue a lot. PopNow content_rejected means the model declined and the run is refunded. If an aspect ratio won't take, it's not in the nine this model supports; the PopNow 400 message names the field.

The one workflow trap: this is a text-to-image endpoint, full stop. No img2img, no inpainting, no reference image, no LoRA. If you need to edit a Neta Lumina output, take the image to an edit node - there are several in the same pack - rather than looking for a strength slider here.