Reference · Models
The models behind PV30
One agent, several providers — each model keeps its real capabilities.
PV30 does not wrap every video model behind one generic generate button. It orchestrates several providers and keeps each model’s real capability set visible: which ones take multiple image references, which support first and last frame, what resolution and clip length they actually produce.
You pick the model per job. Defaults are set for a reason — Seedream 5.0 Lite for element and storyboard images, Seedance 2.0 Mini for first-pass video — but nothing stops you from switching a single shot to Pro, or to MiniMax H3 when a shot needs more reference images.
| Kind | Models | Used for |
|---|---|---|
| Images | Seedream 5.0 Lite (default) · Seedream 5.0 Pro · FLUX | Character, environment and prop images; storyboard frames |
| Video | Seedance 2.0 Mini (default) · Seedance 2.0 Pro | Multi-image reference, first frame, one-click film, reference-video replication |
| Video | MiniMax H3 | Text, first / last / first-and-last frame, up to 9 references; 768P or 2K; 4–15 s per clip; durable jobs |
| Video | LTX-2 | First-frame generation on a self-hosted service |
| Voice | MiniMax TTS | Voice-over tracks |
Images — Seedream 5.0 Lite and Pro
Seedream 5.0 Lite is the default for character, environment and prop images and for storyboard frames. Pro is the higher-fidelity option when a shot carries more detail than the first pass needs.
FLUX is also available for element and storyboard images, for workspaces that prefer it.
Video — Seedance 2.0 Mini and Pro
The default video models. They support multi-image reference, first-frame generation, one-click film, and reference-video replication — borrowing a clip’s motion, camera movement or effects for a shot, anchored to your own character images. It is not face swap and not a one-to-one copy.
Pro is the step up when a shot needs more fidelity than the Mini first pass delivers.
Video — MiniMax H3
Generates from text, a first frame, a last frame, or both, and takes up to 9 reference images — the widest reference set of the models listed here. Clips run 4–15 seconds at 768P or 2K.
Jobs are durable: they keep running in the background and stay visible in the Task Center, so closing the tab does not lose the work.
Video — LTX-2
First-frame generation on a self-hosted service, for workspaces that run their own inference rather than calling a hosted provider.
Voice — MiniMax TTS
Voice-over tracks, kept as a separate audio track from source audio on the timeline so narration can be re-timed without re-rendering picture.
Choosing a model
Start on the defaults — Seedream 5.0 Lite for images, Seedance 2.0 Mini for video — and switch only when a shot asks for something specific. Reach for MiniMax H3 when a shot needs more than a couple of reference images or a 2K clip; reach for Seedance 2.0 Pro when fidelity matters more than speed.
Availability depends on the provider accounts configured for your workspace, and generation tier and final render resolution are set separately from model choice.
Questions
Does PV30 have its own video model?
No. PV30 is an agent that orchestrates several providers — Seedream and FLUX for images, Seedance 2.0, MiniMax H3 and LTX-2 for video, MiniMax TTS for voice — and keeps each model’s real capabilities visible instead of pretending they are the same.
Can I switch model per shot?
Yes. Models are chosen per job, so one shot can run on Seedance 2.0 Mini while another runs on MiniMax H3. Defaults exist to get you started, not to lock you in.
Which model should I use for a consistent character?
Start with Seedance 2.0 Mini or Pro, which support multi-image reference and first-frame generation. Consistency across shots comes from the element reference images PV30 reuses, whichever video model renders the clip.
What resolution does MiniMax H3 produce?
768P or 2K, with clips of 4–15 seconds. It accepts text, a first frame, a last frame, or both, plus up to 9 reference images.
Is LTX-2 available by default?
No — LTX-2 runs as first-frame generation on a self-hosted service, so it is for workspaces that provide their own inference. Availability overall depends on how your workspace is configured.