AI model catalog
76 models registered, 17 marked [self-host required]. Weights are pulled on demand - none of them are bundled in the Docker image.
Looking for more? 660+ importable community models from OpenModelDB, with license and checksum info.
Last regenerated: 2026-07-03 from docker-ai-service/app/main.py:AVAILABLE_MODELS
Real-ESRGAN (2)
| Model | Scale | Type | Status | Description |
|---|---|---|---|---|
realesrgan-x4 | x4 | onnx | ✓ available | Best quality 4x for photos & anime (67MB ONNX) |
realesrgan-x4-256 | x4 | onnx | ✓ available | Optimized for 256px tiles, better for low VRAM |
Fast (OpenCV) (6)
| Model | Scale | Type | Status | Description |
|---|---|---|---|---|
fsrcnn-x2 | x2 | pb | ✓ available | Very fast 2x upscaling, good for real-time |
fsrcnn-x3 | x3 | pb | ✓ available | Fast 3x upscaling |
fsrcnn-x4 | x4 | pb | ✓ available | Fast 4x upscaling, lower quality but quick |
espcn-x2 | x2 | pb | ✓ available | Fastest model, minimal quality improvement |
espcn-x3 | x3 | pb | ✓ available | Fastest 3x model |
espcn-x4 | x4 | pb | ✓ available | Fastest 4x model |
Quality (OpenCV) (6)
| Model | Scale | Type | Status | Description |
|---|---|---|---|---|
lapsrn-x2 | x2 | pb | ✓ available | Good quality 2x upscaling |
lapsrn-x4 | x4 | pb | ✓ available | Good quality 4x upscaling |
lapsrn-x8 | x8 | pb | ✓ available | Extreme 8x upscaling |
edsr-x2 | x2 | pb | ✓ available | Best quality 2x with OpenCV, requires more compute |
edsr-x3 | x3 | pb | ✓ available | Best quality 3x with OpenCV |
edsr-x4 | x4 | pb | ✓ available | Best quality 4x with OpenCV, slowest but best |
Next-Gen (SPAN/SwinIR/APISR/OmniSR/DAT-light/MAN/CRAFT/RGT) (16)
| Model | Scale | Type | Status | Description |
|---|---|---|---|---|
span-x2 | x2 | onnx | ✓ available | SPAN — fastest quality model, NTIRE 2023 winner. Best for real-time video. |
span-x4 | x4 | onnx | ✓ available | SPAN 4x — fast quality upscaling, great speed/quality balance. |
realesrgan-x2-plus | x2 | onnx | ✓ available | Real-ESRGAN x2 Plus — high quality 2x for photos and live-action. |
realesrgan-animevideo-x4 | x4 | onnx | ✓ available | Real-ESRGAN trained specifically for anime video. Optimized for temporal consistency. |
swinir-x4 | x4 | onnx | ✓ available | SwinIR — Swin Transformer for image restoration. Best quality for photos & live-action. |
apisr-x3 | x3 | onnx | self-host | CVPR 2024 — general 3x for photos & video. Ideal for 720p to 1080p. ~25MB. [Upstream Xenova repo gat… |
purephoto-realplksr-x4 | x4 | onnx | ✓ available | RealPLKSR tuned for realistic photos and portraits. 30MB, 67ms/64px-tile CPU. ('RealPLSKR' in the UR… |
omnisr-x2 | x2 | onnx | ✓ available | OmniSR x2 - CVPR 2023, omni-axis self-attention. HFA2k training set (anime-leaning), ~4.6MB. Faster… |
omnisr-x4 | x4 | onnx | ✓ available | OmniSR x4 - CVPR 2023 official weights (epoch895 export). ~5.6MB. Sweet spot between SwinIR and FSRC… |
dat-light-x2 | x2 | onnx | self-host | DAT-light x2 - smaller/faster sibling of DAT2. Upstream HF repo removed; no ONNX mirror found. See d… |
dat-light-x4 | x4 | onnx | self-host | DAT-light x4 - smaller/faster sibling of DAT2. Upstream HF repo removed; no ONNX mirror found. See d… |
man-x2 | x2 | onnx | self-host | MAN x2 - Multi-scale Attention Network, ICME 2023. Upstream HF repo removed; no ONNX mirror found. S… |
man-x4 | x4 | onnx | self-host | MAN x4 - Multi-scale Attention Network, ICME 2023. Upstream HF repo removed; no ONNX mirror found. S… |
craft-x2 | x2 | onnx | self-host | CRAFT x2 - Compositional Refinement texture-aware SR, 2023. Upstream ships only .pth via Google Driv… |
craft-x4 | x4 | onnx | self-host | CRAFT x4 - Compositional Refinement, 2023. Upstream ships only .pth via Google Drive; no public ONNX… |
textures-rgt-s-x4 | x4 | onnx | ✓ available | RGT-S (Recursive Generalization Transformer, small) - new architecture not previously in catalog, st… |
Video Fast (Compact / Speed) (9)
| Model | Scale | Type | Status | Description |
|---|---|---|---|---|
clearreality-x4 | x4 | onnx | ✓ available | SPAN architecture, only 1.7MB. Real-time 4x for clean video. Best for faces, nature, hair. Measured… |
nomosuni-compact-x2 | x2 | onnx | ✓ available | Compact 2x with medium degradation handling. 2.4MB, ideal for real-time video playback. |
lsdir-compact-x4 | x4 | onnx | ✓ available | Compact 4x trained on 85k images. 2.5MB, fast enough for near-real-time video. |
swinir-small-x2 | x2 | onnx | ✓ available | Lightweight SwinIR 2x, only 7.9MB. Good quality/speed balance for video. |
swinir-small-x4 | x4 | onnx | ✓ available | Lightweight SwinIR 4x, only 8MB. Best quality in the fast category. |
bhi-realplksr-x4 | x4 | onnx | self-host | BHI-RealPLKSR x4 - 2x faster than DAT2 at comparable quality. Release asset renamed upstream; no sta… |
realesr-general-x4v3 | x4 | onnx | ✓ available | Real-ESRGAN general-purpose v3 - tiny (~5MB), fast, dynamic-shape all-rounder. Best modern default f… |
lsdir-compact-v2-x4 | x4 | onnx | ✓ available | v2 upgrade of LSDIR Compact - tiny (~2.5MB), fast real-time 4x for low-power/low-VRAM devices. |
spanx2-ch48 | x2 | onnx | ✓ available | SPAN 2x with 48 channels - very fast (~1.7MB), good for real-time on modest GPUs (T600/Arc A380). |
Video Quality (DAT2 / DRCT-L / RealPLKSR / ESRGAN / HAT-L) (9)
| Model | Scale | Type | Status | Description |
|---|---|---|---|---|
ultrasharp-v2-x4 | x4 | onnx | ✓ available | DAT2 Transformer — best overall quality for photos and video. 49MB. License CC-BY-NC-SA-4.0: persona… |
nomos2-dat2-x4 | x4 | onnx | ✓ available | DAT2 trained on Nomos v2 dataset. Fixes noise, compression, blur. 53MB. |
nomos2-realplksr-x4 | x4 | onnx | ✓ available | Modern RealPLKSR architecture. 30MB — best quality-to-size ratio for video. |
drct-l-x4 | x4 | onnx | self-host | DRCT-L x4 - Dense-Residual Connected Transformer. The only verified public ONNX (Phhofm RealWebPhoto… |
realwebphoto-v4-dat2-x4 | x4 | onnx | ✓ available | DAT2 trained specifically on degraded web/compressed images - best match for h264/h265 streaming fra… |
nomoswebphoto-realplksr-x4 | x4 | onnx | ✓ available | RealPLKSR trained on web photos - efficient quality restore for compressed sources. ~30MB. |
foolhardy-remacri-x4 | x4 | onnx | ✓ available | 4x_foolhardy_Remacri - legendary general-purpose ESRGAN upscaler, sharp detail without over-smoothin… |
nmkd-siax-x4 | x4 | onnx | ✓ available | 4x_NMKD-Siax_200k - top-rated ESRGAN for detail/general content, 200k iterations. ~67MB. |
nomos8k-hat-l-x4 | x4 | onnx | ✓ available | Full HAT-L (vs HAT-S in catalog) - highest photo quality, very heavy (~162MB, high VRAM). Poster/bac… |
Film Restoration (7)
| Model | Scale | Type | Status | Description |
|---|---|---|---|---|
fsdedither-x4 | x4 | onnx | ✓ available | Removes dithering artifacts from old DVDs and digitized VHS tapes. 67MB. |
nomos8k-hat-x4 | x4 | onnx | self-host | HAT Transformer trained on 8k dataset. Handles JPG compression + blur. 57MB. |
nomos8kdat-x4 | x4 | onnx | ✓ available | DAT trained on Nomos8k — restores heavily JPEG-compressed sources (old rips). 86MB, heavy on CPU (48… |
nafnet-denoise | - | onnx | ✓ available | NAFNet (SIDD width64) - ECCV 2022 image denoiser. Use as pre-pass before upscaling on noisy source m… |
realesr-general-wdn-x4v3 | x4 | onnx | ✓ available | Real-ESRGAN general v3 with denoise weighting (wdn) - ideal for noisy/compressed sources. ~5MB. |
dejpg-realplksr-1x | - | onnx | ✓ available | 1x DeJPEG restoration - removes JPEG/block compression artifacts before upscaling. Pair with any 4x… |
denoise-realplksr-1x | - | onnx | ✓ available | 1x denoise restoration pre-pass - complements NAFNet with the faster RealPLKSR arch. ~30MB. |
Anime (Real-CUGAN / APISR / Compact) (5)
| Model | Scale | Type | Status | Description |
|---|---|---|---|---|
anime-compact-x4 | x4 | onnx | ✓ available | Ultra-lightweight anime 4x, only 5MB. Perfect for anime real-time playback. |
apisr-anime-x2 | x2 | onnx | ✓ available | CVPR 2024 — trained on anime production pipeline. Best anime 2x quality. 18MB. |
fallin-soft-x2 | x2 | onnx | ✓ available | Real-CUGAN-arch anime 2x by the Adore author, permissively licensed. 5.7MB, 17ms/64px-tile CPU — rea… |
real-cugan-x2 | x2 | onnx | self-host | Real-CUGAN x2 - Bilibili AI Lab. Original HF mirror is gone; the only public ONNX exports (styler00d… |
real-cugan-x4 | x4 | onnx | self-host | Real-CUGAN x4 - Bilibili AI Lab. Original HF mirror is gone; the only public ONNX exports (styler00d… |
Multi-Frame VSR (self-host) (3)
| Model | Scale | Type | Status | Description |
|---|---|---|---|---|
edvr-m-x4 | x4 | onnx | self-host | EDVR-M — Multi-frame video super-resolution. Uses 5 frames for temporal consistency. Best batch qual… |
realbasicvsr-x4 | x4 | onnx | self-host | RealBasicVSR — Recurrent VSR with optical flow (CVPR 2022). Best for real-world degraded video (VHS,… |
animesr-v2-x4 | x4 | onnx | self-host | AnimeSR v2 — Anime-specialized multi-frame VSR (NeurIPS 2022). Preserves line art and flat colors. ~… |
Vulkan / ncnn (3)
| Model | Scale | Type | Status | Description |
|---|---|---|---|---|
ncnn-realesrgan-x4 | x4 | ncnn | ✓ available | Real-ESRGAN x4 via ncnn-Vulkan. Works on any Vulkan GPU (AMD RX 5700, Intel Arc, etc.). Bundled with… |
ncnn-realesrgan-anime-x4 | x4 | ncnn | ✓ available | Real-ESRGAN Anime optimized via ncnn-Vulkan. Best for anime content on Vulkan GPUs. |
ncnn-realsr-x4 | x4 | ncnn | ✓ available | RealSR DF2K x4 via ncnn-Vulkan. Photo-realistic super-resolution on Vulkan GPUs. |
Frame Interpolation (6)
| Model | Scale | Type | Status | Description |
|---|---|---|---|---|
rife-v4.7 | - | onnx | ✓ available | RIFE v4.7 — Faster Frame Interpolation (2x FPS). Ensemble enabled, scale=1. Lighter model for real-t… |
rife-v4.8 | - | onnx | ✓ available | RIFE v4.8 — Balanced Frame Interpolation (2x FPS). Middle ground between v4.7 (fast) and v4.9 (quali… |
rife-v4.9 | - | onnx | ✓ available | RIFE v4.9 — Best-quality Real-Time Frame Interpolation (2x FPS). Ensemble enabled, scale=1. Recommen… |
ifrnet | - | onnx | self-host | IFRNet intermediate-flow interpolation (2x FPS). Second interpolation architecture beside RIFE — arb… |
cain | - | onnx | self-host | CAIN channel-attention interpolation (2x FPS, fixed midpoint). Second interpolation architecture bes… |
rife-v4.25 | - | onnx | ✓ available | RIFE v4.25 - current SOTA frame interpolation, better scene-bleeding handling than v4.7-4.9. Recomme… |
Face Restoration (4)
| Model | Scale | Type | Status | Description |
|---|---|---|---|---|
gfpgan-v1.4 | - | onnx | ✓ available | GFPGAN v1.4 — Tencent ARC's face restoration GAN. Restores heavily degraded faces. 512x512 crops. Ap… |
codeformer | - | onnx | ✓ available | CodeFormer - Robust face restoration with transformer codebook. Good for severely degraded faces. 51… |
restoreformer-plus-plus | - | onnx | ✓ available | RestoreFormer++ - state-of-the-art face restoration, better than GFPGAN/CodeFormer for severely degr… |
gpen-512 | - | onnx | ✓ available | GPEN-512 - face restoration with GAN prior. Different visual style than GFPGAN (more conservative).… |
About [self-host required]
5 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
.pthweights, or the upstream repo is gated (e.g. Xenova HF returns 401), or the model uses ops (HAT LayerNorm) that fail on CPUExecutionProvider. - Recipe: see
docs/MODEL-HOSTING.mdin the repo - PyTorch → ONNX export script. - Once exported, drop the
.onnxfile into/app/models/<model-id>.onnxon 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