LongLive-2.0 vs LTX-2.5: inputs, controls and API cost
Compare LongLive-2.0 and LTX-2.5 APIs: inputs, outputs, streaming delivery and ready-second cost. Choose a Windflow model for your product.
LongLive-2.0 and LTX-2.5 are both realtime video generation options in Windflow's public catalog. Choose by the inputs and interaction you need, then evaluate quality and responsiveness with your actual application. This page compares the published serving profiles; it does not present a head-to-head performance experiment.
Quick decision
- Evaluate LongLive-2.0 for: Multi-shot narrative clips directed live, one text prompt per shot.
- Evaluate LTX-2.5 for: Audio-visual scenes where sound is generated with the picture rather than dubbed after.
Specifications side by side
| Published specification | LongLive-2.0 | LTX-2.5 |
|---|---|---|
| Architecture | Autoregressive multi-shot DiT (Wan2.2-TI2V backbone) | 22B joint audio+video DiT, two-stage, generated one window at a time |
| Resolution | 1280 × 704 | 832 × 448 measured; the canvas is configurable in multiples of 64 |
| Chunk | 29 frames | 241 frames per window (10.04 s) |
| Scene budget | 48 chunks per scene; scene_cut opens a fresh budget | Not published |
| Inputs | Text shot prompts only — no reference image in this release | A text prompt and an optional reference still |
| Outputs | A multi-shot clip on main_video; chunk_index resets per scene while session_chunk keeps counting | Video on main_video and the audio generated with it on main_audio, aligned per window |
| Delivery | WebSocket (H.264) · WebRTC | WebRTC |
| Session protocol | sdk.realtime.v2 | sdk.realtime.v2 |
| Price | $0.0021 per ready second | $0.0040 per ready second |
| Playout | Not published | 24 |
| Control latency | Not published | A whole window: the pipeline does not yield progressively, so time-to-first-frame equals one window's generation time. The reference still is published while it runs. |
| Implementation status | production | production |
Implementation status describes runtime maturity. Live GPU readiness is reported separately; Windflow is in private beta and brings capacity up on demand.
Input and control fit
Both models accept text prompts in the published Windflow profile. LTX-2.5 accepts reference or seed images; LongLive-2.0 does not in its published profile. Both models do not accept uploaded source clips in the published Windflow profile. Both models do not accept live webcam input in the published Windflow profile. Both models do not accept movement and look controls in the published Windflow profile.
A shared session API does not make commands interchangeable. Use each model's supported inputs and control timing. In particular, prompt changes may land at a chunk or generation-window boundary; movement input can represent persistent state until replaced.
What LongLive-2.0 is designed for
LongLive-2.0 is an autoregressive, multi-shot video diffusion model from NVIDIA Research (NVlabs), built on a Wan2.2-TI2V backbone and designed to keep a long video coherent as the prompt changes. On Windflow you direct it with text shot prompts: switch the shot softly while it runs, or cut to a new scene.
- Multi-shot narrative clips directed live, one text prompt per shot.
- Interactive long-form video where a viewer or an agent decides what happens next.
- Storyboard-to-motion experiments that need scene-to-scene continuity.
- Streams that run scene after scene without a fixed end point.
What LTX-2.5 is designed for
LTX-2.5 is Lightricks' 22-billion-parameter joint audio-and-video diffusion model. On Windflow it generates a stream window by window from a text prompt and an optional reference still, delivering video on main_video and the audio generated with it on main_audio, aligned per window.
- Audio-visual scenes where sound is generated with the picture rather than dubbed after.
- Continuous AI television or channel formats built from successive windows.
- Reference-still-anchored characters and settings across windows.
- Comparing a joint audio-video model with video-only alternatives on one protocol.
Compare a realistic session budget
- LongLive-2.0: $0.0021 per ready second, or $1.26 for ten ready minutes.
- LTX-2.5: $0.0040 per ready second, or $2.40 for ten ready minutes.
Queueing and worker startup are free. A ready owner session can remain billable while delivery is paused or unwatched; release it when done. Independent sessions carry independent meters. Use the cost calculator to estimate your duration and daily sessions.
Evaluate before choosing
Use one repeatable input sequence, the same client and network, and equivalent output expectations. Record first visible frame, control-to-visible delay, frame gaps, scene continuity and rejected controls. Run several starts and a sustained session. If serving profiles differ, label the differences instead of treating displayed FPS as directly comparable native generation speed.
See the latency measurement guide and the model selector for a reproducible decision process.
Sources
- NVlabs/LongLive on GitHub — Apache-2.0
- Lightricks/LTX-2 on GitHub — LTX-2 Community License
- Lightricks/LTX-2.5 on Hugging Face
FAQ
Is LongLive-2.0 better than LTX-2.5?
The answer depends on the inputs, control behavior, quality and output your product requires. The published specifications identify capability differences; an application-level evaluation establishes performance for your scenario.
Can I switch between these models through the same API?
Windflow keeps the session lifecycle common. Your application must adapt to each model's supported inputs, commands, output tracks and generation timing.
Do the published prices guarantee available GPU capacity?
No. Rates and implementation status are catalog facts. Live readiness is a separate operational signal, and private-beta access is required to run sessions.