Feature request: CUDA feature extractor for CAMBI

[teb]

Summary
The CUDA-accelerated feature extractors (VIF, ADM, Motion) added via the NVIDIA/Netflix collaboration (VMAF-CUDA, libvmaf 3.0 / FFmpeg 6.1) do not include CAMBI. When a model requiring CAMBI (e.g. a VMAF v1 model) is loaded via libvmaf_cuda, initialization fails with:
libvmaf ERROR could not initialize feature extractor "Cambi_feature_cambi_score"
[Parsed_libvmaf_cuda_N] Failed to configure output pad
This reproduces consistently regardless of pixel format (yuv420p, nv12) or resolution, on a fresh libvmaf build (-Denable_cuda=true, CUDA 13.2, RTX 5070/sm_120). CAMBI works correctly on the identical build/model when run through the CPU (libvmaf) filter instead of libvmaf_cuda, confirming this is a missing CUDA implementation rather than a build or model-loading issue.
Ask
Is a CUDA implementation of CAMBI planned? If not currently planned, is there a technical reason CAMBI is harder to port than VIF/ADM/Motion (e.g. its multiscale/derivative-based approach), or is it simply not yet prioritized?
This matters increasingly since VMAF v1 models require CAMBI, so any pipeline wanting both VMAF v1 accuracy and CUDA-accelerated feature extraction currently has no path forward — you get one or the other.
Environment

libvmaf built from master (July 2026), -Denable_cuda=true
FFmpeg built from git with --enable-libvmaf --enable-ffnvcodec --enable-cuda-nvcc
CUDA Toolkit 13.2, driver 610.74, RTX 5070 (Blackwell, sm_120)
Ubuntu 24.04 in WSL2

[lusoris]

This is not specific to CAMBI. To load a VMAF v1 model, libvmaf_cuda needs a CUDA extractor for each of the model's four features; CAMBI is reported first only because it comes first in the model's feature list.

When a CUDA state has been imported and gpumask is 0, vmaf_use_features_from_model() (libvmaf.c:422-437 at 86da14d) looks up every model feature with the VMAF_FEATURE_EXTRACTOR_CUDA flag. vmaf_get_feature_extractor_by_feature_name() (feature_extractor.c:111) then skips any extractor that does not have it. The model path has no per-feature CPU fallback, so one missing CUDA extractor fails the whole model.

All eight vmaf_v1.0.16 and vmaf_v1.0.16_hfr models use the same four features:

Model feature	CUDA at 86da14d
Cambi_feature_cambi_score	No extractor. #1571 adds one.
Speed_chroma_feature_speed_chroma_uv_score	No extractor. #1574 adds one, but it rejects speed_prescale != 1.0 at init, and 3d0h_2160 (0.5), 5d0h (0.6) and their HFR variants set that option.
VMAF_integer_feature_adm3_score	adm_cuda does not provide it, and has no adm_csf_mode, adm_dlm_weight, adm_noise_weight or adm_min_val option.
VMAF_integer_feature_motion3_score	motion_cuda provides only motion and motion2, and has no motion_max_val option. The HFR models also set motion_five_frame_window and motion_moving_average.

So #1571 and #1574 together are not enough to load a v1 model on libvmaf_cuda, and I found no open PR that adds ADM3 or motion3 on CUDA. Until one does, v1 models need the CPU libvmaf filter.