Self-Hosted LTX-2.3 vs. Kling API
Kicked this one off. The question is simple: does self-hosting video generation actually save money once time and quality get factored in, or does it just end up costing about the same as paying for Kling's API and dealing with none of the setup.
Setup
Running on an RTX 6000 Ada rented through RunPod, $0.77/hr. Generation is LTX-2.3 through ComfyUI, full fp16 checkpoint — 48GB of VRAM means no need to quantize down to fp8.
What's being tested
Same four prompts through both LTX and Kling, same length, same resolution, so it's an actual apples-to-apples comparison instead of a vibes-based one:
- A human subject talking or moving — usually where cheaper models fall apart
- A product shot with a camera move
- A wider scene/environment
- Something abstract or stylized
What's getting logged
| Metric | How |
|---|---|
| Generation time | Cold-start/model-load timed separately from actual generation |
| Cost | Self-host: pod uptime × $0.77/hr. API: actual Kling charge |
| Usable as-is | Yes/no — would this get posted without a redo |
| Quality | Plain 1–5, artifacts and prompt match |
Going in
Rough expectation based on published LTX-2.3 benchmarks: something around $0.0026/sec self-hosted versus Kling's $0.04–0.07/sec — call it 15–25x cheaper on paper. Whether that holds up once the human-subject clip gets judged is the actual point of running this.