FILM
FILM (Reda et al., Google, ECCV 2022): makes the middle frame between two video frames, also for large motion, from a flow pyramid in both directions and features shared across scales. Apache-2.0 code and weights.
experimental Fork of dajes/frame-interpolation-pytorch@33bf477 · backends: portable, kernels
- Paper FILM: Frame Interpolation for Large Motion ECCV 2022 · first released 10 Feb 2022
Checkpoint film
Paper 2022 Added in 0.2.0 Evaluated with 0.2.0
Use it
pip install fastforks Lean install with only what this model needs: pip install fastforks-core
from fastforks import interpolation
model = interpolation.load("film") # kernels on a supported GPU, else portable PyTorch
result = model(frame1, frame2)
result.frame
result.backend # what ran: mode, device, dtype, kernels fastforks bench film Times the model on this machine. load(…, mode="portable") or mode="kernels" picks a backend; usage guide.
Speed and memory
Every backend on one GPU (RTX 5060 Ti 16 GB) and the same input, the authors’ code included. Kernels runs fp16. Portable runs fp16 on CUDA.
| Backend | dtype | Time per call (ms) | vs original fp16 | Peak memory allocated / reserved | Load (s) | First call (s) |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | not measured | – | – | – | – |
| Original fp16the authors' mixed-precision option (float16); the TorchScript release as inference.py, dt 0.5, zero pad to 64 (centred), --fp16: model.half() | fp16 | 464 | – | 2,953 MiB/3,556 MiB | 0.2 | 0.47 |
| Portable fp32 | fp32 | not measured | – | – | – | – |
| Portable | fp16 | 417 | 1.1× | 2,856 MiB/3,080 MiB | 0.4 | 0.42 |
| Kernels | fp16 | 332 | 1.4× | 2,856 MiB/3,080 MiB | 0.7 | 0.33 |
| Portable + torch.compile | 332 | 1.4× | 2,666 MiB/3,192 MiB | 41 (build) | – | |
| TensorRT fp16 | fp16 | 189 | 2.4× | 2,586 MiBenginenot comparable | 203 (build) | – |
Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-14 to 2026-11-16. First call, 3 warm-up, peak reset, median of 3 rounds x median of 5. Ratios divide two
times of this table (same GPU, same input). Original fp16 is the authors’ own precision option, else torch.autocast. torch.compile and TensorRT: build time instead of load time. TensorRT memory (engine, not comparable) is the GPU memory the engine took, read from the driver because TensorRT allocates outside PyTorch, so it can't be compared with the PyTorch peaks of the other rows.
Other devices
| Device | Backend | dtype | ms / call | Peak memory | Output vs reference | Input |
|---|---|---|---|---|---|---|
| Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-14 | Original | fp32 | 10448 | 12.5 GiB process | – | 1080x1920 |
| Portable | fp16 | out of memory | – | 1080x1920 | ||
| Portable | fp32 | out of memory | – | 1080x1920 | ||
| Kernels | fp16 | not implemented | – | 1080x1920 | ||
CUDA memory is torch’s peak allocated / reserved; Apple has no such counter, so the Mac rows show the process footprint (weights, Python and torch included). “Output vs reference” compares the device’s output with the stored golden output under the card’s tolerance; – = not checked on that device. Kernels “not implemented” = no kernels for this model on that backend yet: it runs portable there.
Licence and weights
Commercial use allowed
Our reading of the licences, not legal advice.
Weights
film-net-style-fp32.pt 138 MB· hosted on GitHubfastforks downloads and checks these files on first use (fastforks download film).
Accuracy
The test set through the authors’ code and through ours: same checkpoint, same protocol. The distance between the original’s fp32 and fp16 runs is what half precision alone costs; our runs should stay within it.
DAVIS 2017 480p
2 frame triplets per video (180), PSNR / SSIM
| Run | dtype | psnr ↑ | ssim ↑ | Δ psnr vs fp32 original | triplets | ms / triplet |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 27.63 | 0.8105 | reference | 180 | – |
| Original fp16 | fp16 | 27.46 | 0.8092 | −0.17 | 180 | – |
| Portable fp32 | not measured | |||||
| Portable | fp16 | 27.63 | 0.8105 | ±0.00 | 180 | – |
| Kernels | fp16 | 27.63 | 0.8104 | ±0.00 | 180 | – |
DAVIS 2017 trainval 480p (JPEG), frames 0 / 2 and 20 / 22 of each of the 90 videos give the frame 1 / 21 at time 0.5, compared with the real one: PSNR and SSIM (Gaussian 11 x 11, sigma 1.5) of the output rounded to uint8, on RGB in [0, 1], averaged over the triplets for the 180-triplet benchmark. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-14; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks d199e94c, 2026-11-14.
DAVIS 2017
the first triplet per video resized to 1080p (90), PSNR
| Run | dtype | psnr ↑ | ssim ↑ | triplets | ms / triplet |
|---|---|---|---|---|---|
| Original fp32 | not measured | ||||
| Original fp16 | not measured | ||||
| Portable fp32 | not measured | ||||
| Portable | fp16 | 28.71 | – | 90 | – |
| Kernels | not measured | ||||
The original code was not run on this dataset, so the table cannot show the distance to it.
DAVIS 2017 trainval 480p (JPEG), frames 0 / 2 and 20 / 22 of each of the 90 videos give the frame 1 / 21 at time 0.5, compared with the real one: PSNR and SSIM (Gaussian 11 x 11, sigma 1.5) of the output rounded to uint8, on RGB in [0, 1], averaged over the triplets for the 180-triplet benchmark. The fixed 1080p validation resizes the first three frames of each DAVIS video with bilinear interpolation and measures RGB PSNR before rounding the predicted middle frame. Measured on RTX 5060 Ti, torch 2.14.1+cu130, fastforks e3abc519, 2026-11-17.
Closeness to the original
Beyond the test-set scores: the raw output of ours against the original’s fp32 output on the same input, pixel by pixel. CPU, float32: max |diff| <= 5e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.001 and p99 <= 0.005 (absolute error of the output frame).
| Check | Tolerance | Measured |
|---|---|---|
| CPU fp32, max |diff| | ≤ 0.00005 | not recorded |
| GPU default dtypes, mean |diff| | ≤ 0.001 | 0.00034 |
| GPU default dtypes, p99 |diff| | ≤ 0.005 | 0.00154 |
Recorded for film on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-14, with these kernels: cudnn_hacc
Details
FILM (Reda et al., ECCV 2022): frame interpolation for large motion, the authors' film_net with the Style weights (the published default) as dajes/frame-interpolation-pytorch ports it. A 7-level image pyramid of both frames, cascaded features (one convolutional sub-tree extractor shared by all levels, so the coarse levels learn from fine training samples), a residual flow pyramid in both directions with one predictor shared by the coarse levels, both frames' images and features warped to the time on the 5 finest levels, and a U-Net-like decoder. Trained for the middle frame; other times run as upstream allows.
- Input / output
- Input: two RGB images (consecutive frames). Output: the frame between them at
timestep(0.5 by default; 0 = the first, 1 = the second), RGB in [0, 1]. - Input size
- any H x W >= 64, padded to a multiple of 64; tested up to 1988x2880 (13.0 GiB allocated)
- Options
cuda_graphs=true
- Outputs
frame- CUDA kernels
cudnn_haccconvolutions with fp16 accumulation (cuDNN)
- Apple (MPS)
- fp32, fp16
Full reference, generated from the repository
FILM (Reda et al., Google, ECCV 2022): makes the middle frame between two video frames, also for large motion, from a flow pyramid in both directions and features shared across scales. Apache-2.0 code and weights.
Commercial use: yes.
| variant | input | backend | time (ms) | peak memory (MiB, allocated / reserved) |
|---|---|---|---|---|
film | synthetic (1080x1920) | original portable kernels portable + torch.compile TensorRT fp16 | 464 417 332 332 189 | 2953 / 3556 2856 / 3080 2856 / 3080 2666 / 3192 2658 / 2658 |
Details
- Original code: https://github.com/dajes/frame-interpolation-pytorch at
33bf477ea2c6 - Code licence: Apache-2.0 (permissive); files:
LICENSE - Weights licence: Apache-2.0
- Licence evidence: https://github.com/dajes/frame-interpolation-pytorch/blob/33bf477ea2c65a2ed1b5f2280ee1f02be06910e8/LICENSE ; https://github.com/dajes/frame-interpolation-pytorch/releases/tag/v1.0.2 ; https://github.com/google-research/frame-interpolation/blob/69f8708f08e62c2edf46a27616a4bfcf083e2076/README.md
- Training data: Vimeo-90K (septuplets / triplets for frame interpolation) (google-research/frame-interpolation README "Datasets": Vimeo-90K triplets are the training set (film_net-Style.gin).)
- Options:
cuda_graphs= True (speed only) - Kernel libraries:
cudnn_hacc - Variants:
film - Outputs: frame
- Input size: any H x W >= 64, padded to a multiple of 64
- Tested up to (portable, fp16, RTX 5060 Ti 16 GB):
film1988x2880 (13.0 GiB peak allocated, 15.3 reserved) film1080x1920, first call: portable + torch.compile 41 s (cold Inductor cache), TensorRT fp16 203 s (ONNX + engine).filmoriginal 1080x1920: the authors' mixed-precision option (float16); the TorchScript release as inference.py, dt 0.5, zero pad to 64 (centred), --fp16: model.half()filmload / first call: original 0.29 s / 0.946 s, portable 0.43 s / 0.418 s, kernels 0.65 s / 0.334 s- Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5
Measured
DAVIS 2017 trainval 480p (JPEG), frames 0 / 2 and 20 / 22 of each of the 90 videos give the frame 1 / 21 at time 0.5, compared with the real one: PSNR and SSIM (Gaussian 11 x 11, sigma 1.5) of the output rounded to uint8, on RGB in [0, 1], averaged over the triplets for the 180-triplet benchmark. The fixed 1080p validation resizes the first three frames of each DAVIS video with bilinear interpolation and measures RGB PSNR before rounding the predicted middle frame. Original = the original code named under Details, same checkpoint and protocol.
| variant | dataset | original fp32 | original fp16 | portable | kernels | TensorRT | note |
|---|---|---|---|---|---|---|---|
film | DAVIS 2017 480p, 2 frame triplets per video (180), PSNR / SSIM | 27.63 / 0.8105 | 27.46 / 0.8092 | 27.63 / 0.8105 | 27.63 / 0.8104 | ||
film | DAVIS 2017, the first triplet per video resized to 1080p (90), PSNR | 28.71 / – | 28.71 / – |
Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks d199e94c / e3abc519, 2026-11-14 to 2026-11-17; kernels: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks d199e94c, 2026-11-14; tensorrt: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks e3abc519, 2026-11-17.
Parity
CPU, float32: max |diff| <= 5e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.001 and p99 <= 0.005 (absolute error of the output frame).
| variant | kernels error vs original, mean / p99 | GPU, torch | kernels used |
|---|---|---|---|
film | 0.00034 / 0.00154 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400 |
Example output
No example rendered for this checkpoint yet.