Skip to content

Built in ONNX Filters

animeojisan edited this page Sep 4, 2026 · 5 revisions

Built-in ONNX Filters

English | 日本語

This page covers the ONNX models bundled with cHiDeScaler-Neo.

Neo uses ONNX primarily for AI upscaling/restoration, temporal stabilization, and frame interpolation. ONNX processing is generally heavier than GLSL, and performance depends strongly on input resolution, GPU, backend, and model architecture.


AI upscaling and restoration

Model Scale Best use / characteristics
AnimeJaNai HD V3.1 Performance x2 For HD anime. AnimeJaNai V3 models aim to correct scaling-related blur and degradation while doubling resolution. V3 improves fidelity over older generations, including oversharpening, ringing, aliasing, line color/thickness, soft-shadow edges, and preservation of intentional depth-of-field blur. Neo bundles the fast V3.1 Performance model.
AnimeJaNai HD V3.1 Sharp1 Performance x2 Sharper V3.1 Performance variant. Choose it when you want stronger line definition; compare with Standard because some sources may look too aggressive.
AnimeJaNai SD V1 beta34 Compact x2 For SD anime. The AnimeJaNai project describes the SD family as attempting to make SD anime appear as if mastered in HD. With enough GPU performance it can be followed by an HD model for further upscaling.
AnimeJaNai V2L1 x2 x2 Older lightweight V2-family model. V2 development emphasized source fidelity and reducing V1-style oversharpening, line darkening, and detail loss. Neo uses it in some SD-anime chains.
cHiDeNoise Compact V1 x1 A model specialized for removing MPEG-2 compression noise from Japanese terrestrial and BS digital broadcasts. Compared with the UltraCompact version, it prioritizes stronger denoising and restoration quality at the cost of higher processing load.
cHiDeNoise UltraCompact V1.1 x1 Lighter cHiDeNoise variant prioritizing real-time speed. Denoise/detail recovery is more conservative than Compact.
ToHeart 35mm Compact V1.1 x2 Restoration/upscaling model aimed at retaining a 35mm-film-like character in 1990s cel anime. The author's documentation specifically mentions light cross-color noise, halo, ringing and deinterlacing errors while attempting to preserve source information. It can also work well with other film-produced 1990s anime and good DVD sources.
LosOdy X360 Compact V2 x2 Neo-bundled restoration/upscaling model used for 720p-era Xbox 360 / PS3-type sources. Compact is the quality-oriented variant.
LosOdy X360 UltraCompact V1 x2 Lighter LosOdy360 variant for the same 720p/X360/PS3-oriented use case.
umzi mahou RTMoSR x2 Very lightweight RTMoSR-family 2x model used by Neo's HD Anime preset. Useful when a low-cost HD-anime enhancement/upscale is preferred over heavier Compact-class models.

AnimeJaNai V3.1 Standard vs Sharp

Use AnimeJaNai HD V3.1 Performance as the neutral starting point, then try Sharp1 if you prefer stronger lines and crispness.

AnimeJaNai's V3 documentation emphasizes faithful preservation of line color/thickness, soft shadows, and intentional blur rather than simply maximizing sharpness.


Temporal stabilization

Model Scale Description
TemporalFix S1 v1.1 1x Temporal stabilization for small frame-to-frame shimmer, flicker, and unstable line details after AI processing. Resolution is unchanged.
TemporalFix S2 v1 1x Alternate TemporalFix tuning. Neo's bundled ToHeart 35mm + TemporalFix preset uses S2. Compare S1 and S2 by source.

The upstream vs_temporalfix project describes TemporalFix as a post-filter that averages/stabilizes results over multiple frames to reduce temporal inconsistencies such as fizzle, wiggly lines, shimmer, and temporal noise.


Frame interpolation

Model Use / characteristics
RIFE v4.22 Lite50 Lightest bundled Neo RIFE variant. Prioritizes speed. Like other optical-flow interpolation methods, subtitles, UI, scene cuts, occlusion and very fast motion can produce artifacts.
RIFE v4.22 Lite75 Middle option between Lite50 and the standard Lite model, trading additional compute for improved interpolation quality.
RIFE v4.22 Lite Highest-compute of Neo's three bundled v4.22 Lite variants. RIFE generates intermediate images between neighboring frames for smoother motion.
distilDRBA v2 Lite Anime-oriented interpolation. DRBA (Distance Ratio Based Adjuster) adjusts in-between frames using motion-distance ratios rather than simply treating every interval as uniform. It is useful when the goal is to preserve more of anime's held-frame / 2s-and-3s cadence instead of making all movement continuously smooth.

Neo's RIFE variants

Lite50 → Lite75 → Lite
lighter                heavier
faster                 higher-accuracy oriented

Raising interpolation from x2 toward x5 increases the number of generated frames and therefore increases GPU load.

RIFE vs DRBA

  • RIFE: best when continuously smooth motion is the goal
  • DRBA: best when anime-style held frames/cadence should remain more apparent

They target different interpolation aesthetics rather than being a simple quality ranking.


Backends

The normal Neo package can run ONNX through DirectML.

Supported NVIDIA RTX GPUs can optionally use the separately distributed TensorRT backend pack. Performance differences vary by model.

See DirectML and TensorRT.


Upstream references

For bundled Neo models that do not have a dedicated official documentation page, the descriptions are based on the actual model names and their intended use in Neo's built-in presets.

Clone this wiki locally