# Choosing Models ## Detection model The detection model finds the mosaics in each frame. - **`rfdetr-v6`** (default, included): fast and good for most videos. Use this one if you are not sure. - **`rfdetr-v6-large`**: better detection, but slower. Can be better for 4K. It is a separate download. Take the file for your GPU and put it in `model_weights/`. Jasna finds it automatically. - NVIDIA: [rfdetr-v6-large.onnx](https://github.com/Kruk2/jasna/releases/download/0.1/rfdetr-v6-large.onnx) - AMD: [rfdetr-v6-large.pt](https://github.com/Kruk2/jasna/releases/download/0.1/rfdetr-v6-large.pt) - **`rfdetr-vr-v1`** (included): best for VR180 videos. - **`zelefans-vr-yolo-v2`** (separate download): another VR180 model. - **Lada YOLO** models: can be better for 2D animation. - The old `rfdetr-v5` still works. On AMD, RF-DETR is slower than on NVIDIA. If speed is more important than detection quality, use `lada-yolo-v4`. Each model has its own default detection threshold (`rfdetr-v6`: 0.35, `rfdetr-v6-large` and `rfdetr-vr-v1`: 0.40). Change it with **Detection Threshold** in **Basic Processing**, or with `--detection-score-threshold`. ```bash jasna --input input.mp4 --output output.mkv --detection-model rfdetr-v6 ``` In the [Segment Editor](segments.md) you can choose another detection model for each video. ## Restoration model | | **BasicVSR++** (default) | **LTX** | | --- | --- | --- | | Speed | Fast | Much slower (minutes per minute of video) | | Quality | Good | More detail, more natural | | GPU | Every supported GPU | NVIDIA or supported AMD GPUs (Windows AMD LTX: beta) | | Cost | Free | For supporters ([Supporting the project](../../README.md#supporting-the-project)) | Use LTX for the videos or scenes you care about most. Before you choose it, look at these samples: [1](https://discord.com/channels/1196376491815092265/1199059436199759943/1556346988486856826), [2](https://discord.com/channels/1196376491815092265/1199059436199759943/1556337300814958684), [3](https://discord.com/channels/1196376491815092265/1199059436199759943/1555612574643388598), [4](https://discord.com/channels/1196376491815092265/1556960822578909244/1559341521575084063), [5](https://discord.com/channels/1196376491815092265/1556960822578909244/1559295317529202780), [6](https://discord.com/channels/1196376491815092265/1556960822578909244/1559205894959403121), [7](https://discord.com/channels/1196376491815092265/1556960822578909244/1559197341473574973), [8](https://discord.com/channels/1196376491815092265/1556960822578909244/1558798127195226112) (explore the [new channel](https://discord.com/channels/1196376491815092265/1556960822578909244) and the [old channel (archived)](https://discord.com/channels/1196376491815092265/1199059436199759943) on SLS Discord for more samples), and run **Trial mode** to see if your PC can run it and how long it takes. Pick it in the **Restoration Model** section, or: ```bash jasna --input input.mp4 --output output.mkv --restoration-model-name ltx ``` ### What LTX needs - **License**: a License code that includes LTX ($25 USD, also gives UNet 4x and SD 1.5, see [Supporting the project](../../README.md#supporting-the-project)). Without one, the app offers **Trial mode** instead (see below); the CLI stops with an error. - **GPU**: NVIDIA RTX 30, 40 or 50-series. RTX 20-series should work, but only the RTX 2080 Ti was tested. GTX 16-series is untested. Linux also supports Radeon RX 7000/9000 and Radeon AI PRO R9700 cards. AMD integrated GPUs are untested for LTX. Windows AMD LTX is in **beta** and still needs GPU testing. - **VRAM**: 8 GB minimum. More VRAM is faster: on smaller cards parts of the model are kept in system RAM. - **System RAM**: 16 GB minimum (LTX uses about 10 GB). - **Disk**: the model is not bundled. Jasna asks before downloading it the first time you start an LTX job, and downloads only the model you picked: | Model | Download | | --- | ---: | | Standard Quality (default) | about 15 GB | | Fast mode | about 9 GB | | Good, High or Top Quality, on top | about 13 GB more each | If Hugging Face can't be reached (for example in China), Jasna uses hf-mirror.com automatically. Download still fails? See [Downloading LTX yourself](#downloading-ltx-yourself). Jasna 0.12 needs updated LTX model files, including on NVIDIA. If you copy `model_weights` from 0.11.3, press ▶ and accept the model update; your existing License code and shared VAE files still work. LTX does not work with VR180 side-by-side videos or streaming. Secondary restoration and denoising are turned off with LTX. Using ESET? If LTX won't start and the message names `ebehmonl.dll` or `ebehmoni.dll`, add the Jasna folder to ESET's Deep Behavioral Inspection exclusions. Add it again when you put a new Jasna version in a new folder. ### Pause and continue LTX saves progress automatically, including jobs with ranges or **LTX only for big mosaics**. Use **Ⅱ** to pause and **▶** to continue. After closing Jasna, add the same video again with the same output and settings. In the command line, run the same command again. The current unfinished track may run again, and saving the output starts over; completed parts of a range job are kept. A long single shot can be one track. Jasna asks before replacing saved work if the input, the detection or LTX settings or the app version changed. Changing only output settings (codec, quality, color LUT, sharpening) keeps the saved LTX work. Work saved with sharper large mosaics (768 px) needs 12 GB of free GPU memory again to continue. BasicVSR++ whole-video jobs, previews and streaming do not resume. ### Downloading LTX yourself If the download in Jasna keeps failing, get the files in your browser from [huggingface.co/Kruk2/ltx-restore](https://huggingface.co/Kruk2/ltx-restore/tree/0.12.0) (or [hf-mirror.com](https://hf-mirror.com/Kruk2/ltx-restore/tree/0.12.0)). These links are for Jasna 0.12 and the protected AMD beta. For stable 0.11.3 or older, use the repository's `main` branch. You need 3 files: one for the model you picked, plus the two VAE files. | You picked | Files | | --- | --- | | Standard Quality | `ltx-restore-alpha1-distill2-int8.safetensors.enc` | | Standard Quality + Fast mode | `ltx-restore-alpha1-distill2-nvfp4.safetensors.enc` | | Good Quality | `ltx-restore-alpha1-distill4-int8.safetensors.enc` | | Good Quality + Fast mode | `ltx-restore-alpha1-distill4-nvfp4.safetensors.enc` | | High Quality | `ltx-restore-alpha1-distill8-int8.safetensors.enc` | | High Quality + Fast mode | `ltx-restore-alpha1-distill8-nvfp4.safetensors.enc` | | Top Quality | `ltx-restore-alpha1-int8.safetensors.enc` | | Top Quality + Fast mode | `ltx-restore-alpha1-nvfp4.safetensors.enc` | | Always | `vae.safetensors.enc` and `vae-decoder.safetensors.enc` | Put all 3 in `model_weights/ltx-restore` inside your Jasna folder (create the `ltx-restore` folder if it is missing). Keep the file names exactly as they are, including `.enc`. Start the LTX job again; Jasna will not ask to download. ### LTX options - **Model**: **Standard Quality** (default, recommended), **Good Quality**, **High Quality** or **Top Quality**. Standard Quality is the fastest. Good Quality has a bit more detail and is about 1.5x slower. High Quality has more detail and is about 2.5x slower. Top Quality is slightly better than High Quality and about 5x slower; it is rarely worth it. On an RTX 5090, Standard Quality takes about 1-2x the video's length. CLI: `--restoration-model-name ltx` (Standard), `ltx-4step` (Good), `ltx-8step` (High) or `ltx-undistilled` (Top). - **Fast mode** (RTX 50-series only): about 1.4x faster, slightly less detail, smaller download. CLI: `--ltx-fast`. - **Sharper large mosaics**: more detail on big mosaics (close-ups, 4K), about 3x slower on those. Needs about 12 GB of free VRAM (16 GB cards for Segment Editor ranges); otherwise Jasna restores them normally. CLI: `--ltx-large-canvas`. - **Less overlap (experimental, faster)**: on by default, about 13% faster when a mosaic stays on screen for more than 4 seconds. Turn it off for the older mode with more overlap. CLI: `--no-experimental-ltx-less-overlap`. - **Seed**: a different seed gives a slightly different result; the same seed gives the same result. CLI: `--ltx-seed 7`. - **Trial mode**: runs LTX at full speed without the model or a license, so you can see how fast it is on your PC. The video is **not** restored. CLI: `--ltx-trial`. - **LTX only for big mosaics**: big mosaics get LTX, the rest the faster BasicVSR++. Saves a lot of time on long videos. See [Restoring only parts of a video](segments.md#cli---ltx-auto-split). To preview LTX, open the [Segment Editor](segments.md), select a range, choose LTX and turn on **Restore preview**. The Video Player always uses BasicVSR++. ### How long it takes Rough time for **1 minute of 1080p video with mosaics on screen the whole time** (Standard Quality model, the default). Scenes without mosaics are much quicker. Good Quality is about 1.5x slower, High Quality about 2.5x slower and Top Quality about 5x slower. With **Sharper large mosaics** on, large mosaics take about 2.5x as long. | GPU | Time | With Fast mode | | --- | ---: | ---: | | RTX 5090 | about 1 min | under 1 min | | RTX 4090 | about 1.5 min | not available | | RTX 5080 | about 2 min | about 1.5 min | | RTX 3090 | about 3 min | not available | | RTX 2080 Ti | about 10 min | not available | - The first LTX run takes longer while Jasna prepares it for your GPU. - Progress shows four steps: finding mosaics, preparing, restoring, saving. The first LTX video on your GPU, and the first after each update, starts with a short speed test. After that, the time left is accurate from the end of step 1. - LTX slows down when other apps use the GPU. Close games and browsers. - **Windows**: LTX fails with "Access is denied" while Windhawk runs inside Jasna. Exit Windhawk or add `jasna.exe` to its excluded processes. Jasna warns about this when it starts. ## Secondary restoration Jasna restores each mosaic as a 256x256 picture. On big mosaics, close-ups and 4K video this can look blurry. A secondary model makes this picture bigger (512x512 or 1024x1024) before it goes back into the frame, so the result is sharper. NVIDIA only. - **UNet 4x** (for supporters): quality close to TVAI `iris-2`, and faster than TVAI. See the [examples on SLS Discord](https://discord.com/channels/1196376491815092265/1199059436199759943/1516497879684874260) and [Supporting the project](../../README.md#supporting-the-project). - **RTX Super Res** (free): very fast, nothing to install. Quality is OK. Some videos flicker, so test a short clip first. - **Topaz TVAI**: similar to UNet 4x, but very slow. Needs [Topaz Video](https://www.topazlabs.com/topaz-video), which is paid and Windows only. Use the `iris-2` model. Not used with LTX. ```bash jasna --input input.mp4 --output output.mkv --secondary-restoration unet-4x ``` For TVAI, set the environment variables `TVAI_MODEL_DATA_DIR` and `TVAI_MODEL_DIR` to your Topaz Video model folders, like below. `--tvai-args` changes other Topaz settings. Topaz Video environment variables ### Speed and VRAM comparison Time and VRAM for two 1080p videos: | Secondary type | CAWD 1080p | KV-109 1080p | | ------------------------ | -----------------:| -----------------:| | No secondary | 22s / 10.0 GB VRAM | 11s / 10.7 GB VRAM | | UNet 4x | 29s / 12.5 GB VRAM | 14s / 12.6 GB VRAM | | RTX Super Res | 25s / 11.7 GB VRAM | 13s / 11.4 GB VRAM | | TVAI (2 workers, Iris-2) | 52s / 12.1 GB VRAM | 24s / 12.4 GB VRAM | ## Still-image restoration (SD 1.5) For images, Jasna uses a different model: a trained Stable Diffusion 1.5 model. Add an image to the queue (or give it to the CLI) and Jasna uses this model automatically: ```bash jasna --input photo.png --output restored.png ``` - For supporters. One License code can include SD 1.5, UNet 4x and LTX. See [Supporting the project](../../README.md#supporting-the-project). - Not included in the download. It is about **6.9 GB**. Jasna asks before it downloads it from [huggingface.co/Kruk2/sd-15-jav](https://huggingface.co/Kruk2/sd-15-jav). - Needs about **7 GB of VRAM**, a bit more for big 4K images. This model is experimental. Results change a lot from image to image, but some are very good. Make a few versions and keep the best one: ```bash jasna --input photo.png --output restored.png --sd15-variants 4 ``` To try seeds one by one in the GUI, use [Process images interactively](gui.md#images-one-by-one). All options (`--sd15-steps`, `--sd15-strength`, `--sd15-seed`, ...) are in the [CLI reference](cli.md#sd-15-image-restoration). Examples: [SD 1.5 examples on SLS Discord](https://discord.com/channels/1196376491815092265/1199059436199759943/1492139124348420106) and [more](https://discord.com/channels/1196376491815092265/1199059436199759943/1516571355317800990).