Bring your own keys
Connect your own provider accounts. Usage is billed by each provider directly to you; SeriousPixel never resells model access. Keys live in the OS keychain.
Engines & hardware
Your engines. Your accounts. Your call.
SP Studio routes each job to an engine that’s strong at it, behind one provider interface, so any engine can be swapped without touching your project. On set you choose quality and run location; engine names live in Settings and the shot log.
Engine policy
Open weights: free, never bundled
Open-weights engines are downloaded from their publishers at your request, under your own acceptance of their license. SP Studio records the license you accepted.
Video engines must do audio
The default registry lists only video engines with native synchronized audio, so scratch dialogue and room tone arrive with every take.
License-aware routing
Every registry entry carries license flags. SP Studio checks your organization profile — country, revenue, headcount, use — and warns or blocks before a job runs.
Constraints first
Require Local only, a cost cap, a region or no-training API settings per project. The router drops any engine that breaks a constraint.
Failover with a record
If a provider is down or retires a model, SP Studio offers the next engine and records the change in the take’s provenance.
Registry defaults · as of November 2026
What runs each job by default.
The engine market moves monthly, so defaults live in an updatable registry, not in the app. Power users can override any choice in Pro mode.
| Job | Default | Alternatives | Runs |
|---|---|---|---|
| Writing | |||
| Spark, logline, treatment, adaptation | Your selected language model — Claude recommended | OpenAI flagship; local Qwen, Gemma or Llama-class models | Cloud or local |
| Screenplay, dialogue polish, continuity notes | Your selected language model | A second model for “writers’ room” A/B drafts | Cloud or local |
| Shot direction for image and video engines | Your selected language model, tuned per engine | — | Cloud or local |
| Design & boards | |||
| Concept art, key art, boards with text | OpenAI gpt-image-2 | Grok Imagine image, FLUX.2 | Hosted |
| Character sheets, recurring identity | Google Nano Banana 2 / Pro | gpt-image-2, Seedream 5.0, FLUX.2 | Hosted |
| Fast draft boards and thumbnails | FLUX.2 [klein] 4B (Apache 2.0) | Other open image models | Local |
| Shoot | |||
| Shots — local default | LTX-2.5 (open weights, video and audio together) | LTX-2.3 | Local or API |
| Shots — hosted engines | LTX-2.5 Pro API, Grok Imagine Video 1.5, MiniMax H3 API, Wan 3.0 API | Seedance 2.5, Veo 3.1, Kling 3.0 | Hosted |
| Take review and continuity QA | Claude vision on sampled frames | Gemini vision | Hosted |
| Sound & delivery | |||
| Dialogue and narration | ElevenLabs Eleven v4 | Native audio from the take, for scratch | Hosted |
| Voice consistency across shots | ElevenLabs Voice Changer | — | Hosted |
| Sound effects | ElevenLabs Sound Effects | Native audio from the take | Hosted |
| Music | Your own tracks — WAV, MP3 or stems | ElevenLabs Music (optional, off by default) | Imported |
| Transcription and captions | ElevenLabs Speech to Text | Local Whisper-class model | Hosted or local |
Language models
Cloud or local. Tested on your own pages.
Pick a cloud model with your own key or the best local model your machine can run — globally or per project.
- Writing Check. A blind A/B that runs the same scene or treatment through two or more models, cloud and local. You pick the better page; the result sets the default for that task.
- Capability routing. Long adaptations and whole-bible continuity default to cloud; short rewrites, loglines and captions can stay local.
- GPU memory manager. A local language model and a local video engine share memory, so the queue swaps them automatically on smaller machines.
- Runtimes managed for you. Ollama, llama.cpp or MLX as optional components, fetched on request.
Running locally
Qwen3.6 27B and Qwen3.8-27B (about 18 GB of weights at 4-bit), Gemma 4 31B, and Llama 3.3 70B on 48 GB-class machines and above.
Local quality for long-form screenwriting varies. Treat local models as drafting and editing tools until they pass the Writing Check on your own material. A Local-only project shows a note wherever long-form quality may drop.
License notes
The fine print, surfaced before you shoot.
SP Studio checks these for you, but you should know them. This is a summary, not legal advice — confirm terms with each provider and your counsel.
LTX-2.5 open weights are free for commercial use by organizations under $10M in annual revenue, measured across the entity and its affiliates. Larger studios need their own agreement with Lightricks. Distributing fine-tunes may require a paid license.
MiniMax H3 open weights list the United States as an excluded territory. US organizations use H3 through MiniMax’s hosted API only.
ElevenLabs Music self-serve plans exclude film, TV, radio and games; only the Enterprise music tier covers that work. That’s why music is user-supplied by default and hosted music is off.
Imported music — from a composer, a library or another tool — is logged in the License Ledger with its source and terms, and checked against each delivery. SP Studio imports files; it doesn’t call unofficial APIs.
Compute profiles
Recommended at first run, changeable any time.
The first-run doctor checks your GPU, drivers, memory and disk, benchmarks the machine, and walks you through engine downloads — including gated models that need a publisher sign-in and license acceptance.
Local
Video, local images and local audio run on this machine. Language models follow your Language Models setting. Voice services remain cloud.
Hybrid
Drafts run locally; finals and long or high-resolution shots go to hosted engines.
Cloud
Everything runs through your API accounts. Suits laptops, editing seats and trial use.
Remote GPU Roadmap
A headless SP Studio Core on a workstation or rented GPU, with the desktop app as the client.
Hardware
Local generation wants a workstation. Editing doesn’t.
Apple Silicon Macs and Windows or Linux with NVIDIA GPUs are both supported from the first release. Minimums follow published engine requirements; Recommended and Studio-class are planning figures confirmed during the beta.
Windows & Linux · NVIDIA
| Component | Minimum | Recommended | Studio-class |
|---|---|---|---|
| OS | Windows 10/11 64-bit, or Ubuntu 22.04+ | Windows 11 or Ubuntu 22.04+ | Windows 11 Pro, or Linux for long unattended jobs |
| GPU | NVIDIA, 24 GB VRAM (e.g. RTX 3090 / 4090) | RTX 5090, 32 GB | RTX PRO 6000 Blackwell, 96 GB ECC |
| Memory | 32 GB | 64 GB | 128 GB+ |
| CPU | 8 cores, recent generation | 12–16 cores | 24+ cores |
| Storage | 1 TB NVMe | 2 TB NVMe plus a fast media drive | 4 TB NVMe plus RAID or NAS |
| Runs | LTX-2.5 at reduced speed; language model swaps in and out | LTX-2.5 with a 27B language model (swapped), 4K editing | LTX-2.5 and a 70B-class language model loaded together |
macOS · Apple Silicon
| Component | Minimum | Recommended | Studio-class |
|---|---|---|---|
| OS | macOS 26 or later for local generation | macOS 26 or later | macOS 26 or later |
| Chip | Apple Silicon (e.g. M5 Pro) | M5 Max, 40-core GPU | M5 Ultra |
| Unified memory | 48 GB | 128 GB | 256 GB, or 512 GB for the largest local models |
| Storage | 1 TB SSD | 2 TB SSD plus a fast external drive | 4–8 TB SSD plus Thunderbolt 5 storage |
| Runs | LTX-2.5 at full local settings; language model swaps | LTX-2.5 and a 27B language model together, 4K editing | LTX-2.5 and a 70B-class model, with room for parallel jobs |
Below the minimum? A 16 GB card or a smaller Mac runs SP Studio in the Hybrid or Cloud profile. On a 32 GB Mac, local video is limited to short, lower-resolution takes.
Editing and review seats need far less — the target floor is in line with professional editors: 16 GB RAM and a modest GPU, more for 4K timelines.
Disk. Local video engines need 100 GB or more (up to 160 GB with LoRAs). A local language model adds 18–37 GB. Project media comes on top.
Not sure what your machines can run?
Tell us what’s in your edit bays and render boxes when you apply, and we’ll recommend a profile.