EdenLive: Video Rendering, RTX 5090 Setup and Costs
EdenLive renders a reference video into a new visual style using a target first frame and a text prompt. It offers downloadable 5B and 14B models, including four-step streaming variants. For interactive scenes, an engine or application maintains the world state while EdenLive renders the frames. The published RTX 5090 path uses a specific quantized LightX2V configuration.
Content reviewed:
Open the EdenLive setup and sample recordings ↗
Use this model
See access details in the guide below.
Where the reference-video workflow fits
- Render a greybox game recording with a target appearance while retaining its scene and camera guidance.
- Try a supplied proxy recording before building a live application.
- Compare offline video processing with a streaming deployment using the same reference and first frame.
Prepare a reference video, first frame and prompt
The reference stream carries layout, motion and timing. The first frame and text describe the intended appearance. For a game, keep rules and interaction state in the upstream program; plan the renderer and application as separate parts of the system.
The repository includes cryo-crossing and pulse-skirmish recordings with a manifest, proxy video, target cover image and prompt files. Start with these samples so input preparation and model installation can be checked separately. A text-to-game workbench adds game generation and first-frame services beyond the video renderer.
Choose the model and inference path
Both model sizes have bidirectional and few-step chunk-autoregressive variants. The official examples use 5B at 720p and 14B at 480p. Keep each DiT paired with its matching VAE, scheduler and decoder rather than mixing folders between sizes.
The complete model repository is approximately 210GB across variants and components. That is a download inventory, not a requirement to load every variant into GPU memory. The current LightX2V example downloads the complete release and converts it with --all. Reserve disk space for both the full download and converted files, including quantizations. If downloading selectively, first confirm every component dependency of the chosen recipe; the guide does not supply a reduced download manifest.
| Path | Example output | Starting point |
|---|---|---|
| 5B bidirectional | 720p | Offline reference rendering |
| 5B four-step AR | 720p | Single-GPU streaming recipe |
| 14B bidirectional / four-step AR | 480p | Size-matched official configuration |
Make the first offline render
The upstream installation targets Linux and NVIDIA GPUs with a CUDA 13 driver. The upstream README’s verified environment uses 80GB NVIDIA GPUs; that describes the tested setup rather than a minimum requirement. Follow its pinned environment instructions before downloading the selected weights. The default 5B bidirectional attention profile has SM90-specific kernel requirements; use the LightX2V RTX 5090 recipe for that consumer GPU rather than assuming every upstream profile runs unchanged.
From the installed EdenLive repository, the command below selects its supplied 5B four-step profile and sample manifest. Inspect the output video for alignment with the reference and target appearance. Check a short recording before changing the dataset manifest to your own video.
python -m edenlive.scripts.infer \
--config edenlive/configs/wan22_5b_720p_few_step.yamlUse the complete RTX 5090 recipe
LightX2V prepares INT8 ConvRot DiT and proxy-encoder files plus an FP8 T5 text encoder. Its guide includes conversion commands, model and input paths, a LightTAE decoder and scripts/edenlive/run_5090.sh. Complete conversion and path setup before launching that preset.
Its streaming preset uses compilation, CUDA Graphs and warmup. Keep those costs in a cold-session budget even when evaluating warmed inference separately. The guide also has an H100/H200 configuration; do not transfer its FP8 recipe or memory figures to the RTX 5090.
What the 24.69 FPS result measures
LightX2V reports 24.69 inference FPS on an RTX 5090 for EdenLive 5B at 1280×704, batch one and four denoising steps, with INT8 ConvRot DiT/proxy and resident FP8 T5. It reports 28.38GiB peak allocated and 28.95GiB reserved GPU memory. Keep both memory measures and the exact quantizations with the result.
Inference FPS includes request input preparation, encoding and decoded chunks. It excludes model loading, compilation, warmup, final video concatenation and saving. Peak memory covers the full request, including concatenation and saving. For an interactive application, also measure input delivery, buffering, decoding, transport and display delay. A throughput above 24 FPS does not establish the time from a user action to its displayed result.
Budget the whole session and usable video
Use the GPU rental’s actual hourly rate and the full billed duration. As a hypothetical example, 30 minutes at $1/hour costs $0.50. If that session produces two minutes of accepted video, compute cost is $0.25 per accepted minute, before storage or network charges. Retries and setup time change that denominator.
For a live service, include warm capacity waiting for users, the application server and any relay traffic. Benchmark offline generation and end-to-end streaming separately. The reported FPS alone is insufficient to price a hosted service or predict how many concurrent users one GPU can serve.
Check the components used by your workflow
First-party code and the listed DiT, VAE, UMT5, scheduler and decoder components use Apache 2.0. CLIP uses MIT. The proxy vision encoder and its derivatives carry the separate DINOv3 License. Review the component notices for the files actually used.
The optional workbench first-frame service uses FLUX.2 [klein] 9B with its non-commercial license and content-filtering conditions. The upstream documentation says playing registered games with start.sh does not load FLUX. Keep the workbench’s first-frame licensing separate from the video-only pipeline, while retaining the other component licenses in both.
Before you use it
- The upstream application supplies scene state and reference frames for interactive use.
- The complete workbench adds services and licensing conditions beyond video inference.
How to get started
Start with a supplied recording and the model-matched offline profile, then move to the hardware-specific LightX2V preset when testing streaming throughput.