Frame interpolation models
This page lists the VFI models included in the current release and public-beta versions. Community items need no DLC; Professional items need the Professional DLC. Beta items are available only on the Steam public-beta branch.
Genre performance differs; auto pairing is in Presets. Model tags match the in-app labels:
Anime IRL Anti-Subtitle Decompress Turbo-OnlyRecommendation is product positioning (★★★★★ = current default pick).
Algorithm overview
| Algorithm | Edition | Tags | Rec. | AMD |
|---|---|---|---|---|
| Tariff | Pro | Anime IRL | ★★★★★ | × |
| RIFE | Community | IRL | ★★★★☆ | × |
| ncnn-RIFE | Community | IRL | ★★★☆☆ | √ |
| GMFSS | Community / Pro | Anime | ★★★★☆ | × |
| UMSS | Community / Pro | Anime | ★★★☆☆ | × |
| DRBA | Community / Pro | Anime | ★★★★★ | × |
| GIMM-VFI | Pro | IRL Turbo-Only | ★★★★☆ | × |
| Waft | Pro | Anime IRL | ★★★★★ | × |
| VFSS | Community | Anime IRL | ★★★☆☆ | × |
| TensorRT RIFE | Pro | IRL | ★★★☆☆ | × |
| MNN Tariff | Pro | Anime IRL | ★★★☆☆ | √ |
| Prometheus | Pro | Anime IRL Turbo-Only | ★★★★★ | × |
| IFUNet | Pro | Anime IRL | ★★☆☆☆ | × |
Tips
Weights named ncnn / mnn use ncnn or MNN and run on NVIDIA, AMD, and iGPU. Other CUDA models do not run on AMD or iGPU.
Warning
GMFSS / UMSS / some DRBA paths use a lot of VRAM; avoid 4K+ in one shot. If you also super-resolve, enable VFI then SR under output resolution, or split into two jobs. If artifacts get worse, go back to SR then VFI.
The UI recommends Tariff over GMFSS: similar quality, much faster.
Tariff
Current next-generation family. Quality can match GMFSS pg104 at higher speed. This family is available only in the Professional version.
| Model | Edition | Tags | Rec. | Strengths | Weaknesses |
|---|---|---|---|---|---|
| Tariff_neu2_upg | Pro | Anime IRL | ★★★★☆ | Near pg104 quality at about 2× the speed | — |
| Tariff_neu2_pwr / pwr_fix | Pro | Anime IRL | ★★★★☆ | About 2× faster than UPG; pwr_fix targets pan jitter | Stock PWR may shimmer on pans |
| Tariff_neu2_pge / pge_v2 / pge_v3_gram / pge_v3_real | Pro | Anime IRL | ★★★★★ | Quality-oriented PGE line; v3 splits gram / real | Slower than PWR |
| Tariff_neu2_stable / stable_v2 | Pro | Anime IRL | ★★★★☆ | Safer default look | Peak detail a bit behind UPG |
| Tariff_neu2_nb202 | Pro | Anime IRL | ★★★☆☆ | Lighter structure for iGPU / low VRAM | Below UPG / PGE quality |
| Tariff_neu2_nb202_mnn / pwr_mnn | Pro | Anime IRL | ★★★☆☆ | MNN path; Tariff can run on iGPU / AMD | Below CUDA Tariff quality |
RIFE
Fast, widely used. official_* and most rpr_* are included in the Community version.
| Model | Edition | Tags | Rec. | Strengths | Weaknesses |
|---|---|---|---|---|---|
| official_4.26 / 4.26_heavy / 4.22 | Community | IRL | ★★★★☆ | Latest IRL line for most live-action | heavy is slower and hungrier |
| official_4.18 | Community | IRL | ★★★★☆ | Bidirectional + dynamic flow can beat newer gens on some IRL | Extra options must be turned on |
| official_4.15 / 4.17 | Community / Beta | IRL | ★★★☆☆ | Bridge between 4.9 and 4.18 | 4.17 is available only in the public-beta Community version |
| official_4.9 | Community | IRL | ★★★☆☆ | Tuned for both anime and IRL | Speed similar to the previous gen |
| official_4.8 | Community | IRL | ★★★☆☆ | Anime-biased training | Weak on live-action |
| official_4.6 | Community | IRL | ★★★☆☆ | 2nd-gen, stable, non-integer ratios | Less fine than newer gens |
| official_2.3 | Community | IRL | ★★☆☆☆ | Classic 1st-gen | No non-integer ratios |
| rpr_v7_2.3 | Community | IRL | ★★★☆☆ | 3rd-gen combo, complex motion | Average speed |
| rpr_v7_2.3_ultra#2 | Pro | IRL | ★★★★☆ | Stronger combo on complex shots | Available only in the Professional version |
| rpr_v7_1.0 | Community Beta | IRL | ★★☆☆☆ | Early rpr | Available only in the public-beta Community version |
ncnn-RIFE
Community. Slightly below same-gen CUDA RIFE; much broader GPU support. Many versions ship ensembleTrue / ensembleFalse.
The Community version includes: rife-v2.3, rife-v4, rife-v4.6, rife-v4.9, rife-v4.10–4.11, 4.15, 4.16_lite, 4.18, 4.22, 4.22_lite.
| Strengths | Weaknesses |
|---|---|
| NVIDIA / AMD / iGPU; fast | Softer than CUDA RIFE; lite variants blur more |
GMFSS
Anime quality reference, but slow. The Community version includes pg_104 / pg_117 / union_v; the Professional version includes pg_104_lite / pg_104_pro. pg_119 is available only in the public-beta Community version.
| Model | Edition | Tags | Rec. | Strengths | Weaknesses |
|---|---|---|---|---|---|
| GmfSs_pg_104 | Community | Anime | ★★★★☆ | Among the strongest anime VFI | Slow; Pro users should prefer Tariff |
| GmfSs_pg_104_lite / pro | Pro | Anime | ★★★★☆ | lite is faster; pro leans quality | Still much slower than Tariff |
| GmfSs_pg_117 / pg_119 | Community / Beta | Anime | ★★★☆☆ | Post-104 experiments | 119 is available only on the public-beta branch |
| GmfSs_union_v | Community | Anime | ★★★☆☆ | Stable, smooth | Can blur more than basic |
UMSS
Same anime niche as GMFSS.
| Model | Edition | Tags | Rec. | Strengths | Weaknesses |
|---|---|---|---|---|---|
| Umss_v1 | Community | Anime | ★★★☆☆ | Some shots cleaner/smoother than pg104 | Slightly slower than pg104 |
| Umss_v4 | Pro | Anime | ★★★★☆ | Newer UMSS | Still slow |
DRBA
Keeps anime cadence: linear motion (backgrounds) is interpolated; non-linear motion (characters) keeps the original beat.
The left side is the input and the right side is the output. Backgrounds and other linear motion remain linear after interpolation, while characters and other non-linear motion keep their original anime cadence.
| Model | Edition | Tags | Rec. | Strengths | Weaknesses |
|---|---|---|---|---|---|
| DRBA_RIFE_v4.26 / 4.26_heavy | Community | Anime | ★★★★☆ | Preserves the original cadence; fast; Turbo can approach realtime on some GPUs | heavy is slower |
| DRBA_RIFE_v4.22_lite | Community | Anime | ★★★☆☆ | Lighter DRBA-RIFE | Included only in the release Community version; removed from beta |
| DRBA_GmfSs_pg | Community | Anime | ★★★★☆ | High-quality export that preserves the original cadence | Slow |
| DRBA_RIFE_v4.17 | Pro | Anime | ★★★☆☆ | DRBA-RIFE in the Professional version | Older than 4.26 |
| DRBA_Distill_v1 / v2_lite | Pro | Anime | ★★★★☆ | Accelerated distilled cadence path | lite trades some quality |
| DRBA_Tariff_neu2_pge | Pro | Anime | ★★★★★ | Tariff quality with DRBA cadence | Professional version required |
GIMM-VFI
Solid IRL VFI, GIMM-VFI. Professional; tagged Turbo-Only.
| Model | Edition | Tags | Rec. | Strengths | Weaknesses |
|---|---|---|---|---|---|
| GimmVfi_r_arb | Pro | IRL Turbo-Only | ★★★★☆ | Anytime IRL VFI | Blurry output |
| GimmVfi_r_arb_lpips | Pro | IRL Turbo-Only | ★★★★☆ | LPIPS-oriented look | Slower |
Waft
Professional models. waft_v1 is the anime-leaning version; waft_v2 / v2_lpips are newer versions.
| Model | Edition | Tags | Rec. | Strengths | Weaknesses |
|---|---|---|---|---|---|
| waft_v1 | Pro | Anime IRL | ★★★☆☆ | Anime-oriented Waft | Older than v2 |
| waft_v2 / waft_v2_lpips | Pro | Anime IRL | ★★★★☆ | Newer line; lpips has a more perceptual look | Needs helper model files such as depth |
VFSS
Community, multi-frame input. Vfss_v1_anime / Vfss_v1_real. Does not support resample-style dedup.
| Model | Edition | Tags | Rec. | Strengths | Weaknesses |
|---|---|---|---|---|---|
| Vfss_v1_anime | Community | Anime | ★★★☆☆ | Multi-frame context, anime | Dedup options limited |
| Vfss_v1_real | Community | IRL | ★★★☆☆ | Multi-frame context, IRL | Dedup options limited |
TensorRT RIFE
| Model | Edition | Tags | Rec. | Strengths | Weaknesses |
|---|---|---|---|---|---|
| rife_v4.6 / 4.8 / 4.18 / 4.22 / 4.26_heavy.onnx | Pro | IRL | ★★★★☆ | TensorRT same-gen RIFE | Long first compile; non-NVIDIA cards require VSPipe decoding and DirectML inference |
Prometheus
Recommended general-purpose VFI for live-action footage.
| Model | Edition | Tags | Rec. | Strengths | Weaknesses |
|---|---|---|---|---|---|
| prometheus_svfi / prometheus_v2_svfi | Pro | Anime IRL Turbo-Only | ★★★★★ | Stable on subtitles and UI footage | Uses substantial VRAM and may need tiling |