Configuration reference

The settings page is organised into six tabs. Every control here maps 1:1 to a key in PluginConfiguration.cs.

Settings tab

AI Service connection

Field	Default	Notes
`AI Service URL`	—	Full URL to the Docker AI service. Typically `http://localhost:5000` or `http://<host>:5000`.
`AI Service API Token`	empty	Shared secret. Plugin attaches it as `X-Api-Token` to every call. Must match the container's `API_TOKEN` env var.
`Test Connection`	—	Fires `GET /health` — shows online/offline + auth posture.

Library scan

`Enabled Libraries`	all	Chip picker. Empty = scan all; picks limit to specific virtual folders (added in v1.6.1.14 for issue #64).
`Min Width / Min Height`	1920 / 1080	Threshold for what counts as "low-resolution". OR-logic: a video qualifies if either dimension is below.
`Output Codec`	libx264	12 options: libx264/libx265/libsvtav1/libaom-av1/libvpx-vp9 (software), h264/hevc/av1_nvenc (NVIDIA), h264/hevc/av1_qsv (Intel), plus stream-copy.
`Target Scale`	2x	1x–4x. 3x is the sweet spot for 480p→1440p source material.

Auto Model Selection

New in v1.6.1.15. When enabled, the plugin calls /Upscaler/recommend-model at playback start and picks both model and color-filter preset per video based on genre tags.

Anime → animesr-v2-x4 / anime-compact-x4 + Vivid filter
Horror/thriller → realesrgan-x4plus + Drama
Sci-fi → hat-x4 + Cyberpunk
Documentary/low-res → swinir-m-x4 + Sharp HD
Multi-frame batch → edvr-m-x4 / realbasicvsr-x4

AI Models tab

Live model catalog pulled from /models. Each row shows:

Status dot — available / downloaded / loaded / unavailable
Scale factor — 2×, 3×, 4×
VRAM estimate — based on benchmark
Actions — Download, Load/Unload, Benchmark

Filter bar: search-by-name + status filter (all / available / downloaded / loaded / unavailable). See the model catalog for the full list.

Filters tab

15 camera-style presets that can be applied either as live CSS overlay (no re-transcode, 60fps) or as server-side FFmpeg filter chain (persists in the output file).

Cinematic · vintage film look
Vintage · low saturation + warm
Vivid · saturated + contrast boost
Sepia · monochrome warm
Cyberpunk · neon teal/magenta
Teal & Orange · Hollywood colour-grade
Drama · high contrast desaturate
Soft Glow · bloom on highlights

Sharp HD · edge enhance + clarity
Retro Game · slight scanline
Pastel · low saturation pastel
Noir · grayscale high-contrast
Night Vision · green-channel
Thermal · false-colour
Custom · 6 FFmpeg params exposed

Advanced filter params

Under Advanced you get 6 FFmpeg filter knobs that persist across sessions via POST /Upscaler/filter-config:

Param	Range	FFmpeg filter
Gamma	0.5 – 2.0	`eq=gamma=...`
Sharpness	0 – 5	`unsharp=5:5:...`
Colour Temperature	-100 – +100 K	`colortemperature`
Vignette	0 – 1	`vignette`
Film Grain	0 – 30	`noise=alls=...`
Denoise	0 – 10	`hqdn3d=...`

Face Restoration tab

New in v1.6.1.7. GFPGAN / CodeFormer passes for restoring faces in low-res footage.

`Enable Face Restoration`	off
`Model`	`gfpgan-v1.4` or `codeformer`
`Max Faces Per Frame`	5 (guard rail)
`Min Face Size (px)`	48
`Alpha Feather (px)`	12

Compute cost: face restoration is a second model pass on top of upscaling. Budget ~2× the frame time per face.

Hardware tab

Read-only view of what the AI service detected on the container host — useful for debugging why a model isn't loading.

Execution providers (CUDA / OpenVINO / DirectML / ROCm / CPU) with version strings
GPU list with VRAM, driver, CUDA/compute capability
Last benchmark timestamp + FPS per model
Onnxruntime version + build flags

Advanced tab

Field	Purpose
`Max Concurrent Streams`	Cap on parallel AI jobs. 1 for single-GPU, 2+ only if you have VRAM headroom.
`Enable Pre-Processing Cache`	Stores upscaled frames on disk so a replay doesn't re-run inference.
`Enable Remote Transcoding`	Lets Jellyfin hand off transcoding to the AI service host over SSH (see Deployment).
`Enable Performance Metrics`	Collect per-frame timing; visible on Hardware tab.
`Log Level`	Information / Debug / Trace. Affects plugin logs in `jellyfin.log`.