Shader / Video Filter Selection Guide

Video-enhancement shaders and frame interpolation are disabled by default when mpv starts. They are enabled only when you explicitly select them.

Default mpv startup:
      ┌─────────────────┐
      │ Normal playback │
      │ No Anime4K      │
      │ No interpolation│
      │ No shader chain │
      └─────────────────┘
               │
               │ User selects an option from the right-click menu
               ↓
      ┌───────────────────┐
      │ Video enhancement │
      ├───────────────────┤
      │ Anime4K HQ/Fast   │
      │ FSRCNNX           │
      │ ArtCNN            │
      │ CuNNy             │
      │ NNEDI3            │
      │ CFL               │
      │ FSR               │
      │ Sharpen           │
      └───────────────────┘

Note: The recommendations below were compiled with AI based on the Shader combinations currently available in this project's input.conf. Results vary significantly with the source video, encoding quality, original resolution, and display device. Compare options during actual playback.

Analysis of MPV Video Enhancement Shader Combinations

TIP

• This guide primarily reflects the latest version; older versions may differ. 🍭 View the latest releases
• To check for updates in the player, right-click and select Tools > Check for updates. 🎯 How to update to the latest version
• Performance levels are Low / Medium / High / Very High, estimated as relative workloads when outputting 4K on a mid-range GPU (such as a GTX 1660 / RTX 3050).

---

1. Recommendations by content type

🎨 Anime

Source → Output	Recommended combination	Reason	Performance
720p → 4K (recommended)	Anime4K Mode A (HQ) or Mode B (HQ)	3x upscaling requires strong line reconstruction; HQ gives the best results	High
720p → 4K (limited performance)	Anime4K Mode A/B/C (Fast)	Keeps the main reconstruction logic while using smaller CNN models	Medium
1080p → 4K (recommended)	Anime4K Mode A (HQ) or Mode B (HQ)	2x upscaling; HQ is sufficient, though Fast can also work	High / Medium
1080p → 4K (limited performance)	Anime4K Mode A/B/C (Fast)	Keeps the main reconstruction logic while using smaller CNN models	Medium
1080p → 4K (efficient)	CuNNy/2x12 or CuNNy/fast	Low overhead; suitable for everyday use	Low
1080p → 4K (higher workload)	CuNNy/4x32 or CuNNy/8x32	High workload; newer AI upscaling with richer detail, for enthusiasts	High–Very High

Native 4K sources do not need upscaling.

🎬 Live-action

Source → Output	Recommended combination	Reason	Performance
720p → 4K	FSRCNNX/2x16+SSim+Krig	High-quality upscaling plus post-processing correction; natural results without a smeared look	High
720p → 4K (balanced)	FSRCNNX/2x8+Krig	Moderate upscaling with reasonable performance cost	Medium
720p → 4K (performance first)	FSRCNNX/2x2+Krig	Lightweight upscaling, suitable for weaker GPUs	Low
1080p → 4K	FSRCNNX/2x8+Krig or FSRCNNX/2x16+SSim+Krig	2x upscaling; 2x8 is often enough, while 2x16 is sharper	Medium / High
1080p → 4K (texture quality)	ArtCNN/32+RAVU-LITE-R4	Preserves natural texture with moderate denoising	High
1080p → 4K (efficient)	FSRCNNX/2x2+Krig or CuNNy/4x12	Low overhead and natural-looking results	Low–Medium

Native 4K sources do not need upscaling. Anime4K is not recommended for live-action: its line reconstruction and denoising are optimized for anime and can make live-action look painterly or remove skin texture.

---

2. Shader combination reference

1. Anime4K series

Menu name	Included shaders	Function	Best suited for	Performance
Mode A (HQ)	Clamp_Highlights + Restore_CNN_VL + Upscale_CNN_x2_VL + AutoDownscalePre_x2/x4 + Upscale_CNN_x2_M	Line reconstruction + high-quality upscaling	Anime	High
Mode B (HQ)	Same as above, with Restore_CNN_Soft_VL	Softer reconstruction + upscaling	Anime with softer lines	High
Mode C (HQ)	Clamp_Highlights + Upscale_Denoise_CNN_x2_VL + AutoDownscalePre_x2/x4 + Upscale_CNN_x2_M	Denoising and upscaling	Noisy anime	High
Mode A+A (HQ)	Mode A plus an additional Restore_CNN_M layer	Second reconstruction pass	Anime where line quality is the top priority	Very High
Mode B+B (HQ)	Mode B plus an additional Restore_CNN_Soft_M layer	Second, softer reconstruction pass	Anime	Very High
Mode C+A (HQ)	Denoising/upscaling + extra reconstruction	Denoising + reconstruction	Anime	Very High
Mode A (Fast)	Clamp_Highlights + Restore_CNN_M + Upscale_CNN_x2_M + AutoDownscalePre_x2/x4 + Upscale_CNN_x2_S	Lightweight line reconstruction + upscaling	Anime	Medium
Mode B (Fast)	Same as above, with Restore_CNN_Soft_M	Lightweight soft reconstruction	Anime	Medium
Mode C (Fast)	Clamp_Highlights + Upscale_Denoise_CNN_x2_M + AutoDownscalePre_x2/x4 + Upscale_CNN_x2_S	Denoising + lightweight upscaling	Anime	Medium
Mode A+A (Fast)	Clamp_Highlights + Restore_CNN_M + Upscale_CNN_x2_M + Restore_CNN_S + AutoDownscalePre_x2/x4 + Upscale_CNN_x2_S	Two lightweight reconstruction passes	Anime	Medium
Mode B+B (Fast)	Clamp_Highlights + Restore_CNN_Soft_M + Upscale_CNN_x2_M + AutoDownscalePre_x2/x4 + Restore_CNN_Soft_S + Upscale_CNN_x2_S	Two lightweight soft-reconstruction passes	Anime	Medium
Mode C+A (Fast)	Clamp_Highlights + Upscale_Denoise_CNN_x2_M + AutoDownscalePre_x2/x4 + Restore_CNN_S + Upscale_CNN_x2_S	Denoising + lightweight reconstruction	Anime	Medium
A4K-DTDD	Denoise_Bilateral_Mode + Deblur_DoG + Thin_HQ + Darken_HQ	Denoising, deblurring, line thinning, and darkening	Anime post-processing	Medium

Caution: Anime4K can make live-action look smeared or plastic-like. It is not recommended for live-action content.

2. FSRCNNX series

Menu name	Included shaders	Function	Best suited for	Performance
FSRCNNX/2x2+Krig	FSRCNNX_x2_2_12_0_4_1_v1_fastv2 + KrigBilateral	Lightweight luma upscaling + chroma upscaling	Live-action / general use	Low
FSRCNNX/2x8+Krig	FSRCNNX_x2_8_0_4_1 + KrigBilateral	Medium luma upscaling + chroma upscaling	Live-action / general use	Medium
FSRCNNX/2x16+SSim+Krig	FSRCNNX_x2_16_0_4_1 + SSimSuperRes_RT + KrigBilateral	High-quality upscaling + post-processing correction + chroma	Live-action / general use	High
FSRCNNX/2x56+SSim+Krig	FSRCNNX_x2_56_16_4_1 + SSimSuperRes_RT + KrigBilateral	Maximum-quality upscaling + post-processing + chroma	Live-action, high-end hardware	Very High

FSRCNNX is a general-purpose CNN that tends to look natural on live-action content, making it one of the top choices for live-action sources. It includes KrigBilateral chroma upscaling.

3. ArtCNN series

Menu name	Included shader	Function	Best suited for	Performance
ArtCNN/16	ArtCNN_C4F16.glsl	Medium-quality luma upscaling	Anime / live-action	Medium
ArtCNN/16DS	ArtCNN_C4F16_DS.glsl	Denoising + medium-quality luma upscaling	Noisy anime / live-action	Medium
ArtCNN/16DN	ArtCNN_C4F16_DN.glsl	Strong denoising + medium-quality luma upscaling	Very noisy anime / live-action	Medium
ArtCNN/32	ArtCNN_C4F32.glsl	High-quality luma upscaling	Anime / live-action	High
ArtCNN/32DS	ArtCNN_C4F32_DS.glsl	Denoising + high-quality luma upscaling	Noisy anime / live-action	High
ArtCNN/32DN	ArtCNN_C4F32_DN.glsl	Strong denoising + high-quality luma upscaling	Very noisy anime / live-action	High
ArtCNN/48	ArtCNN_v0_C4F8.glsl	Very high-quality luma upscaling (larger model)	Anime / live-action, high-end hardware	Very High
ArtCNN/48DS	ArtCNN_v0_C4F8_DS.glsl	Denoising + very high-quality luma upscaling	Anime / live-action, high-end hardware	Very High
ArtCNN/48DS_CMP	ArtCNN_v0_C4F8_DS_CMP.glsl	Denoising + very high-quality upscaling + compression-artifact repair	Anime / live-action, high-end hardware	Very High

Important: The newer ArtCNN menu items no longer include CfL_Prediction chroma upscaling. Enable CHROMA: CFL_RT or CHROMA: Krig manually through the individual shader toggles.

C4F8 is a larger model with the best quality but very high performance cost. C4F16 and C4F32 are common choices.

_DS means denoising + sharpening; _DN applies stronger denoising. Choose according to the source.

4. CuNNy series

Menu name	Included shader	Function	Best suited for	Performance
CuNNy/faster	CuNNy-faster-DS.glsl	Extremely lightweight upscaling	Anime	Low
CuNNy/fast	CuNNy-fast-DS.glsl	Lightweight upscaling	Anime	Low
CuNNy/2x12	CuNNy-2x12-DS.glsl	Entry-level upscaling	Anime	Low
CuNNy/3x12	CuNNy-3x12-DS.glsl	Low-end upscaling	Anime	Medium
CuNNy/4x12	CuNNy-4x12-DS.glsl	Low-to-mid-range upscaling	Anime	Medium
CuNNy/4x16	CuNNy-4x16-DS.glsl	Balanced upscaling	Anime	Medium
CuNNy/4x24	CuNNy-4x24-DS.glsl	Mid-to-high-end upscaling	Anime	High
CuNNy/4x32	CuNNy-4x32-DS.glsl	High-end upscaling	Anime	High
CuNNy/8x32	CuNNy-8x32-DS.glsl	Maximum-quality upscaling	Anime	Very High
dp4a variants	Same shaders under CuNNy/dp4a/, with -Q in the filename	Uses Vulkan dp4a instructions for acceleration	Anime	One level lower than the equivalent standard version

Important: Newer CuNNy menu items no longer include CfL_Prediction chroma upscaling, so a chroma shader must be paired manually.

CuNNy is mainly trained for visual novels and anime. Results on live-action may be inferior to FSRCNNX / ArtCNN.

The dp4a version requires vo=gpu-next and gpu-api=vulkan in mpv.conf, as well as hardware that supports dp4a.

5. Individual shader toggles

Menu name	Shader	Function	Performance
LUMA: NNEDI3/x4	nnedi3_nns128_win8x4.glsl	Edge-directed luma upscaling	Medium
LUMA: NNEDI3/x6	nnedi3_nns128_win8x6.glsl	Higher-quality edge-directed luma upscaling	High
LUMA: FSRCNNX/2x16	FSRCNNX_x2_16_0_4_1.glsl	Luma upscaling	High
LUMA: RAVU-ZOOM-R2	ravu_zoom_r2.glsl	RAVU luma scaling	Medium
LUMA: RAVU-ZOOM-R3	ravu-zoom-ar-r3.hook	Higher-quality RAVU luma scaling	Medium
LUMA: RAVU-LITE-R4	ravu-lite-ar-r4.hook	Lightweight RAVU luma scaling	Low
SSimSuperRes_RT	SSimSuperRes_RT.glsl	Post-processing sharpening/correction	Medium
SSimDownscaler_RT	SSimDownscaler_RT.glsl	Downscaling (high → low resolution)	Medium
CHROMA: CFL_RT	CfL_Prediction_RT.glsl	Chroma upscaling	Low
CHROMA: Krig	KrigBilateral.glsl	Chroma upscaling	Low
AMD: FSR_EASU_RT	AMD_FSR_EASU_rgb_RT.glsl	AMD FSR upscaling (EASU)	Low
AMD: FSR_RGB_RT	AMD_FSR_rgb_RT.glsl	AMD FSR upscaling (RGB)	Low
Anime4K: Restore	Anime4K_Restore_CNN_M.glsl	Line reconstruction	Medium
Anime4K: Upscale	Anime4K_Upscale_GAN_x2_M.glsl	GAN upscaling	High
OUTPUT: Adaptive Sharpen_RT	adaptive_sharpen_RT.glsl	Adaptive sharpening	Low

These toggles can be switched individually for fine-tuning or added to an existing combination. In particular, CHROMA: CFL_RT and CHROMA: Krig must now be enabled manually to get chroma upscaling.

3. Performance cost summary

Level	Typical combinations	Suggested GPU class
Low	CuNNy/faster, CuNNy/fast, CuNNy/2x12, FSRCNNX/2x2+Krig, CfL_Prediction, adaptive_sharpen	GTX 1050 / integrated GPU
Medium	Anime4K Fast series, CuNNy/3x12–4x16, FSRCNNX/2x8+Krig, ArtCNN/16+SSim	GTX 1660 / RTX 3050
High	Anime4K HQ series, CuNNy/4x24–4x32, FSRCNNX/2x16+SSim+Krig, ArtCNN/32+RAVU	RTX 3060 / RTX 4060
Very High	Anime4K Mode A+A/B+B/C+A (HQ), CuNNy/8x32, FSRCNNX/2x56+SSim+Krig, ArtCNN/32+RAVU-ZOOM-R3	RTX 4070 or higher

The dp4a variants (CuNNy/dp4a/*-Q) are typically 20%–50% faster than standard variants of the same size, but require Vulkan and compatible hardware.

Anime4K GPU recommendations

• Suggested GPUs (e.g. GTX 1080, RTX 2070, RTX 3060, RX 590, Vega 56, 5700XT, 6600XT) or higher:

  • Anime4K: Mode A (HQ)
  • Anime4K: Mode B (HQ)
  • Anime4K: Mode C (HQ)
  • Anime4K: Mode A+A (HQ)
  • Anime4K: Mode B+B (HQ)
  • Anime4K: Mode C+A (HQ)

• For GPUs below examples such as GTX 980, GTX 1060, or RX 570:

  • Anime4K: Mode A (Fast)
  • Anime4K: Mode B (Fast)
  • Anime4K: Mode C (Fast)
  • Anime4K: Mode A+A (Fast)
  • Anime4K: Mode B+B (Fast)
  • Anime4K: Mode C+A (Fast)

Frame interpolation

Press n. Interpolation is disabled by default. Repeated presses cycle through:

Interpolation Off ➜ Cinema Mode ➜ Anime Mode ➜ Ultra Smooth

• Cinema Mode: de-judder with no blur.
• Anime Mode: uses sphinx with a blur level of 0.65, aiming for smooth motion while tightening anime linework without making it blurry.
• Ultra Smooth: uses bicubic interpolation, which can look slightly soft by default; the setting is adjusted to -0.40 for a little sharpening and to reduce excessive smearing.
• This configuration uses mpv's built-in temporal interpolation (interpolation=yes). Based on AI summaries of online discussions, it is considered a good overall choice for many hardware setups. If you need a different interpolation method, you can ask an AI assistant; the setup is relatively simple.

---

4. General recommendations

1. Prioritize chroma upscaling: Place CfL_Prediction or KrigBilateral at the beginning of the chain to improve color edges.
2. Use only one luma upscaler: Do not combine ArtCNN and RAVU luma upscaling simultaneously; this can duplicate upscaling and waste performance.
3. Apply post-processing after upscaling: SSimSuperRes_RT and adaptive_sharpen_RT can correct artifacts after upscaling.
4. Use downscaling only for high → low resolution: SSimDownscaler_RT is suitable for playing 4K sources on a 1080p display, not for low → 4K upscaling.
5. Anime4K's Clamp_Highlights: Under vo=gpu-next, it may darken the image. Setting linear-downscaling=no may avoid this.
6. dp4a requires Vulkan: For the CuNNy dp4a version, set gpu-api=vulkan and gpu-context=winvk in mpv.conf.
7. Switch in real time: Use shortcuts in input.conf (such as Alt+F1–Alt+F12) to compare combinations. Press Shift+I, then 2, to view frame timing and confirm whether a shader is active.

---

Shader usage cautions

1. Do not stack every shader together.

   Some combinations in this project are already complete processing chains, for example:

   ArtCNN + RAVU
   FSRCNNX + SSimDownscaler + KrigBilateral
   Anime4K Mode A/B/C

   Treat these as complete presets. Do not stack their components on top of the same components again.

2. Higher output resolution is not always better.

   Native 4K video usually does not need further upscaling. High-bitrate, high-quality sources may not need strong AI upscaling either.

3. A blurrier image does not automatically need a stronger shader.

   For low-bitrate, heavily compressed, or very noisy video, strong upscaling may also amplify blocking, noise, and artifacts. Restoration, denoising, or structural reconstruction may be more suitable.

4. Actual results depend heavily on source quality.

   The same shader can behave very differently on anime, live-action, and game footage. Start by choosing according to content type and GPU performance, then compare results with A/B tests.