Super-resolution models

Tips

Super-resolution requires the Professional DLC. All SR models are provided in the Professional version. Beta in the table means the model is available only on the public-beta branch.

Anime IRL Anti-Subtitle Decompress Turbo-Only

Recommendation is product positioning (★★★★★ = current default pick). 1× restore models are detailed under Image restoration models.

Algorithm overview

Algorithm	Edition	Tags	Rec.	AMD
realCUGAN	Pro	Anime	★★★★☆	×
ncnnCugan	Pro	Anime	★★★★☆	√
realESR	Pro	IRL	★★★★☆	×
ncnnRealESR	Pro	IRL	★★★☆☆	√
Anime4K	Pro	Anime	★★★☆☆	√
AnimeSR	Pro	Anime	★★★☆☆	×
waifu2x	Pro	Anime	★★★☆☆	√
waifuCuda	Pro	Anime	★★★☆☆	×
RTXSR	Pro	IRL	★★★☆☆	×
TensorRT (ONNX)	Pro	Anime IRL	★★★★★	×
Compact	Pro Beta	Anime IRL	★★★☆☆	×
SPAN	Pro Beta	Anime IRL	★★★☆☆	×

Tips

SVFI’s genre split:

Anime is mostly flat layers with clear layer boundaries (hand-drawn 2D, most 3D-to-2D). 3D backgrounds + 2D characters still count as anime.

IRL is single-camera live-action or CG where layers cannot be told apart (live-action film, 3D CG, 3D games).

realCUGAN / ncnnCugan

Anime-first, excellent results. up2x / 3x / 4x are scale factors; pro is the enhanced line, see the official notes. conservative is conservative; no-denoise skips denoise; denoise_N is denoise strength.

Model	Edition	Tags	Rec.	Strengths	Weaknesses
Full realCUGAN pth set	Pro	Anime	★★★★★	Main anime SR	CUDA only
Matching ncnnCugan set	Pro	Anime	★★★★☆	AMD / Intel / NVIDIA	A bit slower than CUDA

realESR / ncnnRealESR

Usable on 3D anime; still anime-leaning. RealESRGAN hallucinates more (sharper, punchier); RealESRNet smears more and keeps color. Models with anime in the name are faster; anime is official. RealESR_RFDN is fast for anime.

Model	Edition	Tags	Rec.	Strengths	Weaknesses
RealESRGAN_x2plus / x4plus and anime variants	Pro	IRL	★★★★☆	Clear and vivid	Easy to oversharpen
RealESRNet_x4plus	Pro	IRL	★★★☆☆	Color-preserving smear	Less detail
RealESR_RFDN_x2plus_anime110k	Pro	Anime	★★★★☆	Fast	Locked to 2×
RealESR_x2_anime_APISR_RRDB_GAN	Pro Beta	Anime	★★★☆☆	APISR anime line	Beta only
ncnn: animevideov3 x2/x3/x4, x4plus, x4plus-anime	Pro	IRL	★★★☆☆	Cross-vendor	Slightly below CUDA
ncnn: AnimeJaNai / AniScale / LSDIR / nomo8ksc	Pro Beta	Anime IRL	★★★★☆	Extra beta ncnn models	Available only on the public-beta branch

realesr-animevideov3: conservative anime-video SR, fast and stable; avoid TTA.

AnimeSR

AnimeSR was developed by Tencent ARC Lab. Only AnimeSR_v2_x4 is provided, and its look is more conservative than CUGAN.

Model	Edition	Tags	Rec.	Strengths	Weaknesses
AnimeSR_v2_x4	Pro	Anime	★★★☆☆	Tencent ARC; more conservative than CUGAN	4× only

Anime4K

Anime4K is a very fast, conservative real-time anime SR algorithm with six presets: Anime4K_Upscale_x2_A/B/C/D are 2× presets (A is the default), followed by x3 and x4 presets.

Model	Edition	Tags	Rec.	Strengths	Weaknesses
Anime4K_Upscale_x2 A/B/C/D, x3, x4	Pro	Anime	★★★☆☆	Very fast, realtime-ish, conservative	Low detail ceiling

Custom Anime4K

JSON chains live under models\sr\Anime4K\models. Example Anime4K_Upscale_x2_A.json:

{
  "shaders": [
    { "path": "Restore/Anime4K_Clamp_Highlights.glsl", "args": [] },
    { "path": "Restore/Anime4K_Restore_CNN_VL.glsl", "args": [] },
    { "path": "Upscale/Anime4K_Upscale_CNN_x2_VL.glsl", "args": ["upscale"] }
  ]
}

• 1× restore shaders (Clamp / Restore): empty args
• 2× upscale shaders: upscale
• Anime4K_AutoDownscalePre_x2.glsl-style: downscale
• List order is execution order; edit or add JSON files as needed

waifu2x

waifu2x is a classic conservative SR algorithm: cunet and anime are mainly for anime, while photo can be used for live-action footage.

Model	Edition	Tags	Rec.	Strengths	Weaknesses
waifu2x cunet / anime / photo	Pro	Anime	★★★☆☆	Classic conservative; photo for IRL	Old generation

waifuCuda

waifuCuda is the CUDA implementation of waifu2x; it is mainly for anime and has a speed and look somewhat similar to CUGAN.

Model	Edition	Tags	Rec.	Strengths	Weaknesses
waifuCuda nunif-cunet2x	Pro	Anime	★★★☆☆	CUDA port, CUGAN-like	CUDA only

RTXSR

Model	Edition	Tags	Rec.	Strengths	Weaknesses
rtxsr_q1–q4	Pro	IRL	★★★★☆	NVIDIA renderer SR, quality 1–4	Needs the NVIDIA encode path

Compact / SPAN

Compact and SPAN are available only in the public-beta Professional version. In Steam, open the app properties → Betas and opt in first. Compact commonly hosts AnimeJaNai and AniScale; SPAN hosts Nomos and similar.

Model	Edition	Tags	Rec.	Strengths	Weaknesses
AnimeJaNai HD V3 Compact / Ultra / SuperUltra	Pro Beta	Anime	★★★★☆	Fast; SuperUltra > Ultra > Compact	Weak DoF, easy to sharpen backgrounds
AnimeJaNai V2 three tiers	Pro Beta	Anime	★★★☆☆	Previous JaNai	Behind V3
2x-AniScale-compact / AniScale2S	Pro Beta	Anime IRL	★★★★☆	Keeps detail, low smear/sharpen	Slower
SPAN Nomos / ClearReality / PurePhoto	Pro Beta	Anime IRL	★★★☆☆	Many IRL/general models	SVFI loads nf=48 only

AnimeJaNai works on 3D anime and some live-action footage, but is better suited to anime. It is a lighter CUGAN-like model with weak depth-of-field recognition, so it can sharpen backgrounds. Speed is generally SuperUltra > Ultra > Compact.

TensorRT (ONNX)

NVIDIA acceleration. qa_fte is tagged IRL Turbo-Only and only runs on the Turbo path.

Model	Edition	Tags	Rec.	Strengths	Weaknesses
CUGAN onnx set	Pro	Anime	★★★★☆	TensorRT CUGAN	First compile is slow
qa_fte v0/v1 (1× / 2× / 4×)	Pro	IRL Turbo-Only	★★★★★	Good live-action results	Turbo required
realesr-animevideov3-4x, RealESRGANv2-animevideo xsx2/xsx4	Pro	Anime	★★★★☆	Light anime video	Fixed scale
AnimeJaNai / AniScale / AniSD / Adore / Fallin / waifu2x cunet and similar	Pro Beta	Anime IRL	★★★★☆	Extra beta onnx	Beta only; some filenames say test_only

Warning

TRT must compile first: keep thread count at 1 on the first run. Retry five or six times on failure, then contact the developers. Quality should match the non-TRT sibling except on a few shots.

Visual comparisons

Aniscale Demo 1

2x-AniScale-compact, strong detail, low smear/sharpen (lower VRAM, slower)

Aniscale Demo 2

Same as above

realesr-animevideov3-x2

2× SR, good detail, slight smear, medium sharpen (lower VRAM, faster; avoid TTA)

Add OpenModelDB super-resolution models

SVFI can load extra model files that match its loaders. The Compact, SPAN, ATD, and ONNX (TensorRT) structures on OpenModelDB are compatible.

Example: add Compact

• Search Aniscale for AniScale-2-Compact

• Open the first-generation Aniscale page. Size: 64nf = features (channels), 16nc = convs (depth)

• Compact loading rules:
  • name contains super ultra (AnimeJaNai): nf=24, nc=8
  • name contains ultra: nf=64, nc=8
  • default nf=64, nc=16
• If Aniscale-2-Compact has no structure text, assume the default and put the pth in SVFI\models\sr\Compact\models (create the folder if needed)

SPAN is the same idea; only nf=48 is supported. Other forks are not.

Example: add a TensorRT model

You can also add onnx such as AnimeJaNai. Requirements:

• One input and one output, both [dynamic, 3, dynamic, dynamic]
• Input name input, output name output
• Place under SVFI\models\sr\TensorRT\models

Engine compile

Compile produces an .engine, e.g. realesrgan_2x.onnx.540x960_workspace128_fp16_io32_device0_8601.engine means a 540×960 tile. Tile size changes speed a lot; prefer no tiling.

Other rules

• ESRGAN defaults to nf=64, nb=23; if the name contains anime, nb is treated as 6

Terms

• nf → number of features
• nc → number of convs
• nb → number of blocks