Anime to Real · Edit

You've got an anime character you like - a render from an Illustrious or Pony checkpoint, or a frame someone drew - and you want to see what that character looks like as an actual person. Not a cosplay filter, not an img2img pass that keeps the cel shading. A photograph.

That's the whole job of Anime to Real · Edit. It's one of nine one-tap image tools in the PopNow pack, it costs 4.5 credits (about four and a half cents) per image, and it has exactly four things you can set that matter.

What this node actually is

PopNow is a cloud studio - video, image and audio models behind one developer API - and this pack wraps it as ComfyUI nodes. Nothing here runs on your GPU: your image goes up, the result comes back as a normal IMAGE, and the graph carries on. Same category as ComfyUI's own API nodes, same tradeoffs - no weights, no VRAM, per-call pricing, and your picture leaves your machine.

Locally, an anime-to-real pass was always fiddly: a realistic checkpoint whose latent space agreed with your anime one, a denoise high enough to kill the cel shading and low enough to keep the costume. A hosted instruction editor collapses that into one call. The cost of the collapse: the model is redrawing, so it makes judgement calls you don't get to veto.

How it works

The node is a thin client over popnow/anime-to-real-v1/edit. It encodes your image_url input to PNG (or a high-quality JPEG if the PNG would blow past the 10 MiB upload ceiling), sends it with your other fields, polls the task until it succeeds, downloads the returned file and decodes it into a ComfyUI image batch. Plus a task id you can look up in the API console.

The inputs worth touching

Only image_url is mandatory, and everything else has a sane default.

prompt is optional and the tooltip undersells it: a short description of the people and the place helps skin, fabric and light look real. It is not a scene description - the composition comes from your source. Use it to say "a young woman in a school uniform on a city street, overcast daylight" and you'll get better skin than leaving it blank.

aspect_ratio defaults to model default, which keeps the shape of your source. Leave it there. Changing it on an edit model means the model recomposes the character to fill a new frame, which is the fastest way to lose the resemblance you were trying to preserve.

output_format is JPEG by default. If you're going to feed the result into another pass, switch it to something lossless.

seed starts fixed on purpose, because every run is billed - queue the same graph twice and you get the same picture rather than a second charge. Flip control-after-generate to randomize when you want a different take, and accept that each run costs another 4.5 credits.

Outputs are images (a batch, in case the model returns several) and task id. Wire images into Save Image, or into a second pass.

Install

It's on the Comfy Registry as popnow-comfyui. Manager → search PopNow → Install → restart, or:

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

Then set a key from popnow.ai/api/console/keys, either in Settings > PopNow or in the environment before launch (required on a headless box):

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

No model downloads, nothing in requirements - the pack installs with zero dependencies because it only uses what ComfyUI already ships. If you used the old single-file popnow_api.py, delete it; the same node type is in the pack now.

Where people get burned

It is a restyle, not a restoration. The face will land close, not identical. Anime faces aren't real faces - the proportions don't survive the trip, and the model resolves eyes, nose and jaw into plausible anatomy. If you need the character to stay recognisable across many images, generate the real version once, then use it as the reference for everything else rather than re-running this node per shot.

One image per run. The socket takes a single image; if you connect a batch, the pack uploads the first picture and logs that it ignored the rest. Batch your anime-to-real passes by queueing per image, not by throwing a batch at the socket.

Two restarts, not one. The node list is rebuilt from your account's catalog at startup. Install, add the key, restart - and restart again if a model you expect is missing from the menu.

Watch the spend ceiling. Settings > PopNow > Max credits per run stops any run whose quote goes over your limit, before anything is held. Set it to something small while you're learning what the model does with your prompts. Failed runs are refunded, but the ones that succeed because you queued five randomizes are not.