Or can you project the stills onto some crude geometry? If you run the shots strung out as a single clip into a camera tracker, then you can generate a point cloud that you build meshes from and project each still from each frames camera onto that mesh then clean it up. This is essentially photogrammetry. You could then turn projectors and meshes on and off depending on where your virtual camera is located when you render the scene.
Without seeing the footage this may not be possible (Will a 3D track work on what was shot?) or a Gaussian splat approach may be simpler. It would not be a quick solution but could potentially work.
Got a solution working which involved Runway, Kling, Seedance, Neat Video, Topaz, Flame, After Effects and Baselight.
for those curious I tried: flame ML Timewarp, Silhouette ML timewarp, Kronos, Twixtor 7&8, flame motion timewarp, flame morph same as ML timewarp, @talosh timewarp, Gaussian Splat, and had a back up plan of rebuilding in cg with multi plane stuff.