AI model catalog

32 models shipped, 5 marked [self-host required]. Weights are pulled on demand — none of them are bundled in the Docker image.

General-purpose upscalers

Model	Scale	Type	VRAM (est.)	Best for
`realesrgan-x4plus`	4×	ESRGAN	~4 GB	General live-action, default
`realesrgan-x2plus`	2×	ESRGAN	~3 GB	Fast 2× pass
`realesrnet-x4plus`	4×	ESRGAN	~4 GB	Noise-heavy sources
`swinir-m-x4`	4×	SwinIR	~5 GB	Film grain preservation
`swinir-l-x4`	4×	SwinIR Large	~8 GB	Best quality, slow
`hat-s-x4`	4×	HAT Small	~4 GB	Balanced modern
`hat-m-x4`	4×	HAT Medium	~6 GB	Balanced modern
`hat-l-x4`	4×	HAT Large	~9 GB	Sharpest general

Anime upscalers

Model	Scale	Type	VRAM (est.)	Best for
`realesrgan-anime-x4`	4×	ESRGAN-anime	~3 GB	Classic anime, default
`anime-compact-x4`	4×	Compact	~2 GB	Fast anime, low-VRAM GPUs
`waifu2x-cunet-x2`	2×	CUNet	~2 GB	Line-art preservation
`waifu2x-upconv-x2`	2×	UpConv	~2 GB	Fast waifu2x variant
`animesr-v2-x4 [self-host required]`	4×	AnimeSR v2	~5 GB	Modern anime, best quality
`apisr-x3 [self-host required]`	3×	APISR	~4 GB	Anime at 3× scale

Fast / low-VRAM

Model	Scale	Type	VRAM (est.)	Best for
`fsrcnn-x2`	2×	FSRCNN	~200 MB	Real-time on iGPU
`fsrcnn-x3`	3×	FSRCNN	~250 MB	Fast 3× pass
`fsrcnn-x4`	4×	FSRCNN	~300 MB	CPU-only fallback
`espcn-x2`	2×	ESPCN	~150 MB	Lightest model
`espcn-x4`	4×	ESPCN	~200 MB	CPU real-time

Temporal / multi-frame

Model	Scale	Type	VRAM (est.)	Best for
`edvr-m-x4 [self-host required]`	4×	EDVR Medium	~8 GB	Temporal consistency
`realbasicvsr-x4 [self-host required]`	4×	BasicVSR	~7 GB	Real-world low-res video
`nomos8k-hat-x4 [self-host required]`	4×	HAT-Nomos	~9 GB	Nomos8k-trained variant

Frame interpolation (RIFE)

Model	Purpose	VRAM (est.)	Speed tier
`rife-v4.9`	Best quality	~2 GB	Slow
`rife-v4.8`	Balanced (new in v1.6.1.12)	~2 GB	Medium
`rife-v4.7`	Fast	~1.5 GB	Fast

Migration: old rife-v4.6 / rife-v4.6-lite references in saved configs are transparently mapped to the new versions via MODEL_ALIASES.

Face restoration

`gfpgan-v1.4`	Forgiving restoration, works on very blurry faces, less sharp	~1 GB
`codeformer`	Sharper restoration, needs reasonably recognisable face input	~1.5 GB

About [self-host required]

Five models do not have a public ONNX mirror available at a URL the plugin can hit directly. You have to export them yourself.

Why: the authors either published only PyTorch .pth weights, or published them behind a Google Drive / registration gate.
Recipe: see docs/MODEL-HOSTING.md in the repo — PyTorch → ONNX export script.
Once exported, drop the .onnx file into /app/models/<model-id>.onnx on the service host and it'll appear as available.

Model endpoints (REST)

GET  /models                       # full catalog
GET  /models/{id}                  # single model details
POST /models/{id}/download         # fetch ONNX weights
POST /models/{id}/load             # warm into GPU memory
POST /models/{id}/unload           # free VRAM
GET  /Upscaler/recommend-model     # plugin → service: pick per-video