Anima Turbo LoRA · Edit

What you're actually looking at

Anima is CircleStone Labs' anime model - a 2B base built on NVIDIA's Cosmos-Predict2 rather than the usual SDXL lineage, from tdrussell of diffusion-pipe fame. "Turbo" is the distilled build: same model, far fewer steps. On the open-weights side you'd run it in ComfyUI with your own character or style LoRA loaded, and on a modest card that's a wait measured in minutes per image. This node is the same job done on PopNow's servers: hand it a picture, a prompt, and up to three Anima LoRAs, and get the edit back as a normal IMAGE in your graph.

That last part is the reason to care. If you own a laptop, a Mac, or a card that simpers at anything past SD1.5, a Turbo-LoRA stack on a 2B anime model isn't slow, it's absent. This node turns it into one queue click.

The mechanism, briefly

There's no model download, no sampler, no VAE. The pack is a thin HTTP client over api.popnow.ai (it uses only what ComfyUI already ships - aiohttp, torch, Pillow - so dependencies = [] in its pyproject.toml is not a lie). When you queue, the node encodes your image_url tensor to PNG, or to a high-quality JPEG if it's over 10 MiB, uploads it, and swaps an opaque popnow:// reference into the request. That upload is cached for twelve hours keyed on the file's hash, so re-running the same picture with a different prompt doesn't send it again.

Then it asks for a quote, creates the task with an Idempotency-Key so a network hiccup can't bill you twice, polls until the task settles, downloads the result and hands you a tensor.

The inputs worth setting

• prompt - the author's own tooltip is the right instruction: "Describe the image you want back, not just the change." Write the finished picture, not a diff. For a LoRA, put its trigger words at the front.
• image_url - the source image. The result keeps its aspect ratio, so don't bother pre-cropping to a "supported" shape.
• strength - 0 to 1, default 0.8. This is the "how far from the source" dial. 0.8 is a real restyle; 0.3 is close to a filter. Turning it down while pushing the resolution tier up is how you get mush.
• resolution - 1k (about a megapixel) or 1.5k (about 2.4). Aspect ratio survives either way.
• seed - fixed on purpose. Queue the same graph twice and you get the first result back rather than a fresh bill. Switch control-after-generate to randomize when you actually want a new take.
• loras (optional) - a POPNOW_LORAS socket fed by a PopNow LoRA node. Chain up to three. The tooltip is explicit that these must be trained for Anima: a Pony or SDXL LoRA is the wrong foundation and will produce mud.

Outputs are images (IMAGE - wire it to Save Image or a preview) and task id (STRING - the handle you quote to support or find the run in the API console).

Install

ComfyUI Manager → search PopNow → Install, or:

comfy node install popnow-comfyui
# or, by hand:
cd ComfyUI/custom_nodes && git clone https://github.com/popnow-ai/popnow-comfyui

Restart, then create a key at popnow.ai/api/console/keys and paste it into Settings > PopNow (or set POPNOW_API_KEY). Restart once more - the node list is built from your account's catalog at startup, so a missing node usually means you skipped that second restart. If you installed the old single-file custom_nodes/popnow_api.py, delete it.

Nothing is downloaded and nothing is quantized, so your VRAM is irrelevant. Your credit balance is not: this node is priced from about 2.7 credits per image, quoted before the run, with a Max credits per run ceiling in settings that stops a run before anything is held.

Where people get burned

A PopNow 401 means the key is wrong or revoked, not that the model is down. 40201 means the wallet is empty. 40004 means that combination of fields isn't servable - change the one the message names. A content_rejected or generation_failed ends the run and the credits come back.

The one non-obvious trap: cancelling does not stop the job. It stops you waiting. The task finishes on PopNow, is charged if it succeeds, and the result sits in your console logs. If you're mid-thought and about to hit cancel, hit nothing instead.

And a fair warning about the vendor: PopNow has essentially no footprint in the community discussion archives - I found zero threads mentioning it - so this is a new paid API pack, not a battle-tested one. It's MIT and the source is readable, but there's no crowd yet telling you where the edges are.