Skip to content
Checking the API
  • New renderOpen the composerN
  • Go to StudioG S
  • Go to JobsQueue, progress, cancelG J
  • Go to GalleryG G
  • Go to ModelsStatus, profiles, licences, pricesG M
  • Go to SettingsTeam · devices · providers · budget · healthG ,
  • Connect my GPUSettings › My devices
  • Licences & creditsEvery model licence, limits and required credits
  • HelpWhich model when · CPU vs GPU vs API · costsG H
  • Switch to light themeT
  • Keyboard shortcuts?
  • Sign out

Keyboard shortcuts

Press ? anywhere to open this sheet.

  • Command palette⌘K
  • Keyboard shortcuts?
  • Go to StudioG then S
  • Go to JobsG then J
  • Go to GalleryG then G
  • Go to ModelsG then M
  • New renderN
  • Submit the composer⌘↵
  • Toggle themeT
  • Close dialogsEsc
Sign in

Guides

How to get the clip you want — and know what it costs.

Short, honest answers. The composer shows the same facts live: device, ETA, cost, licence.

Which model when

Product shots, abstract loops, drafts

Wan 2.2

Open, no audio, the quickest CPU draft. Upgrade to GPU for 720p.

Clips with sound

LTX-2.5 or Kandinsky 6.0

Both produce audio. LTX is the most CPU-friendly large model; Kandinsky adds lip-sync.

Image-to-video with character consistency

MiniMax H3

#1 image-to-video on the leaderboard. Self-hosted needs a GPU worker (credit shown); the API option gives 2K.

Highest quality, paid

Kling 3.0 · Seedance 2.0 · Wan 3.0 · Gemini Omni Flash 1.1

API only, priced per second. Veo 3.1 is no longer offered: Google shut it down in the Gemini API on 2026-10-22.

Still images and edits

Nano Banana 2 · Nano Banana Pro · FLUX.3

API only, priced per image; count 1–4 per job. (Nano Banana is deprecated by Google.)

Cinematic 720p without audio

HunyuanVideo 1.5

Open; 1080p super-resolution only on GPU.

CPU vs GPU vs API

Server · CPU

Always available, included. Smallest sensible variant, ≤ ~640×384, 2–3 s clips, few steps. Tens of minutes to an hour per clip — the ETA is measured from past jobs and refreshes live. One job at a time; cancel any time. MiniMax H3 cannot run here and says so.

GPU (yours or a node)

Full variants at 720p–1080p, 5–10 s clips, full steps. 8–24 GB cards get the low-VRAM profile (quantised/offloaded); 24 GB+ the full one. Your own machine runs the Studio Agent in pull mode — nothing to open on your router. Still included.

API · pay per use

Runs in the provider's cloud through an admin-held key. Priced per second (video) or per image, in USD and SAR, estimated before and final after. Progress is approximate. Your prompt and reference leave the server — self-hosted renders never do.

Costs explained

  • Included means no per-clip fee: self-hosted compute on our server, a rented GPU, or your own machine. The ETA is the “cost” you see.
  • API prices are seconds × USD per second (video) or images × USD per image, then × 3.75 for SAR (the riyal is pegged).
  • Verified prices carry the date they were checked and a link to the provider's pricing page. A price older than 30 days asks to be re-verified.
  • “Price to be verified” is a reference figure from an earlier date. It is shown, never hidden, and never presented as current.
  • Budgets (monthly team cap, per-user cap) warn before they block; a paid job that would exceed them is refused with the reason. Self-hosted jobs are never budget-gated.

Connect your GPU

  1. Open Settings › My devices and choose Connect a device. A one-time token is created for the agent.
  2. Run the install command for your platform (Windows PowerShell, macOS, Linux). The agent installs ComfyUI and the dispatcher, detects your GPU (NVIDIA cuda or Apple Silicon mps) and registers.
  3. Enable the models you want on that machine; only those weights are downloaded, after a disk and VRAM fit check.
  4. In the composer, pick My GPU · your machine from the “Run on” list — or leave Auto, which prefers the fastest included option.

The agent polls for work (no inbound ports) and is marked offline after two minutes without a heartbeat.

Licences & credits

Each model card links its licence. Wan 2.2 (Apache-2.0) and Kandinsky 6.0 (MIT) are permissive. LTX-2.5 is free for commercial use under $10M annual revenue (AlGraphy qualifies). HunyuanVideo 1.5's community licence allows commercial use with no revenue cap, but not in the EU, the UK or South Korea (no use, outputs or GPU workers there). MiniMax H3's community licence requires the credit “Powered by MiniMax H3” on every self-hosted output — the portal shows it wherever the output or the model appears — and does not cover hardware in the US, EU, UK or South Korea, so a self-hosted H3 worker must run elsewhere (KSA is fine); outputs may not be used in those places either. API models fall under each provider's terms, and API calls leave the server.

Read every licence in one place

Keyboard shortcuts

  • Command palette⌘K
  • Shortcut sheet?
  • New renderN
  • Go to Studio / Jobs / Gallery / ModelsG then S / J / G / M
  • Submit the composer⌘↵
  • Toggle themeT
  • Close dialogs and drawersEsc

Status & support

Workers, providers, queue depth and disk live in Settings › Health. To report a problem, open the job's drawer and use Report a problem — it attaches the job id. Provider keys are entered by admins only and are never shown again.

Changelog: v0.1 (2026-10-25) — portal scaffold with composer, queue, gallery, models, settings, mock mode.