FLUX 3 · Text to Video

Most text-to-video prompts are written by someone picturing a video. The FLUX 3 model actually generates a soundtrack alongside the picture - the generate_audio switch defaults on, and per the tooltip the price is identical either way. So write the sound into the prompt. Half the people who try this node don't, get a clip whose audio is generic wind, and come away underwhelmed by a feature they never asked for.

This is the purest form of the API-node trade: there's no FLUX 3 video to download, so if you want it, you call it. Billed per second of output, on PopNow's credits, from a graph that runs fine on a laptop.

How it works

The node builds the request from your widgets and nothing else. Any field left at model default is dropped from the payload so the model decides; the run is quoted before anything is held, submitted under an idempotency key (so a retry can't double-bill you), then polled while the node's status line shows the estimate, the queue and finally the charge. The finished MP4 is downloaded and returned as a ComfyUI VIDEO, which Save Video accepts without conversion.

The inputs that matter

prompt - the tooltip's list is the template: subject, action, camera, light, and sound. "A cyclist carves a wet switchback at dawn, tracking shot from a car window, tyres hissing, distant birdsong" gives the audio pass something to work with. Leave sound out and it improvises.

quality - standard or draft. Draft is the fast, lower-cost preview that renders at 720p only, and in the shipped price table it's roughly a third of the standard 720p rate. Iterate in draft. Only promote to standard once the motion is right; motion is what you're actually re-rolling, and you shouldn't pay 1080p money to find out the subject walked backwards.

resolution - 720p or 1080p. It interacts with quality: draft ignores your 1080p choice.

aspect_ratio - 16:9 default, plus 9:16, 1:1, 4:3, 3:4, 21:9, 2:1. Set it here; there's no reframe pass downstream.

duration_seconds - a slider from 5 to 20 in whole seconds, and the thing that multiplies your bill. Five seconds is plenty for a test.

generate_audio - on by default, same price either way. Keep it. Even if you'll dub over it, a synced track tells you whether the action reads correctly.

Outputs

video is the clip, and it has no last-frame companion on this node - so if you're planning to keep a shot going, go through FLUX 3 · Video Extend rather than trying to pull a frame out of a node that never exposes one. task id is the job's identity in the API console, and what support asks for when a render dies.

Installing it

Manager → search PopNow → Install → restart. Manually:

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

Then a key from popnow.ai/api/console/keys, entered under Settings > PopNow or exported as POPNOW_API_KEY, and a second restart so the catalog loads. There is nothing else to install - no weights, no torch surgery. If you ever used the single-file popnow_api.py node, delete it first.

Where people get burned

Cost, almost always. A rate on the badge (credits per second) is not the total, and standard 1080p at twenty seconds is a serious number. Set Settings > PopNow > Max credits per run to something you'd shrug at, and raise it deliberately.

Audio mismatch is the second one: if you asked for a conversation, the model may invent speech with garbled words. Describe ambient sound and physical noise rather than dialogue, and if you need voices, generate them separately and mux later.

And remember the semantics of stopping: hitting cancel ends your wait, not the job. The render continues on PopNow, is charged if it succeeds, and the file shows up in your API console logs - so "I cancelled it" is not a refund.