Anima Turbo · Text to Image

Try before you download

Anima has a fanbase for reasons that have nothing to do with benchmarks: it's a small, fast anime model from a two-person-scale lab, built on NVIDIA's Cosmos-Predict2 rather than the SDXL family everyone else kept refining, and its licence was quietly rewritten to let individual people sell their own finetunes and LoRAs. If you've been circling it and haven't committed to a 12 GB download, this node is the cheap test. Bare Anima Turbo, no LoRA socket, from about 1.8 credits an image - the least expensive thing in this corner of the pack.

The trade is that you get a hosted, distilled model with a vendor's moderation layer, not weights. That's fine for answering "do I like this look?" and much less fine if you need the model's unrestricted character.

Under the hood

This is an API node, full stop. Your prompt goes up, a quote comes back, the task is submitted with an Idempotency-Key so retries don't double-charge, and the node polls the task until it settles and downloads the result into an IMAGE tensor your graph can use like any local output. The pack itself installs nothing - dependencies = [], and it leans on aiohttp, torch, Pillow and NumPy that ComfyUI already ships - so "install" is really "clone and add a key".

Turbo is a distillation, which sets expectations: fewer denoising steps means it can't spend the time a full model spends on fine detail. At anime-model resolutions that's usually invisible. On dense, hand-drawn linework at 1.5k you'll occasionally see it.

The four inputs

• prompt - Danbooru tags, natural language, or both. This is the one field that decides whether the image is good.
• aspect_ratio - default 1:1. Anime models are opinionated about composition; go portrait for characters, landscape for scenery, and don't ask for 21:9 and then complain about the hands.
• resolution - 1k is about one megapixel, 1.5k about 2.4. Bump it only once the composition is right.
• seed - fixed by default. Re-queueing the exact same graph gives you the same billed image back rather than a new one, which is a nice touch from a pack that charges per run. Flip control-after-generate to randomize when you're hunting.

Outputs are images (IMAGE) and task id (STRING). Wire images wherever you'd wire a local checkpoint's output - Save Image, a detail pass, an upscaler - because it's just a tensor.

Getting it installed

ComfyUI Manager → search PopNow → Install and restart. Or:

comfy node install popnow-comfyui

Then key setup, which is where most first runs die: create one at popnow.ai/api/console/keys, open Settings > PopNow, paste, restart. On a headless box, POPNOW_API_KEY=sk-popnow-... in the environment works and takes precedence. If you ever ran the standalone popnow_api.py from popnow.ai's docs, delete it - same node type, and keeping both is asking for confusion.

No checkpoint to download. No VRAM math. The resource you're spending is credits, at roughly 1.8 per image, quoted before the run and refunded if generation fails.

When it goes sideways

No PopNow API key is set - you didn't restart after pasting. PopNow 401 - the key is mistyped, missing, or dead. PopNow 40201 - out of credits; lower the resolution tier or top up. PopNow 40004 - the model can't serve that combination of settings, and the error names which one. PopNow 40310 - verify your email.

A model you expected isn't in the node list? Restart ComfyUI; the catalog is fetched from your account at startup, not baked into the install.

The one thing that isn't a bug: hitting cancel doesn't cancel the job. It stops you waiting while the task keeps running server-side and is charged if it succeeds. The result still lands in your console logs.

Worth knowing before you invest: as of writing, PopNow has effectively zero presence in the community corpora - I couldn't find a single thread discussing it. That doesn't make it bad; it makes it unproven. Start small, keep the key out of shared workflows, and treat it like any other API-key node from a young vendor.