Seamless Loop

Seamless Loop, under DaSiWa / video, turns a decoded RGB IMAGE batch into a looping IMAGE batch. Choose an interpolation backend and connect the output to Enhanced Video Combine or another saver. Keep source FPS for loop processing; align audio downstream if the frame count changes.

Interpolation backend

interpolation_backend defaults to RIFE / FILM, preserving existing workflows. RTX Frame Generation replaces only the transition interpolation with NVIDIA VFX; it does not increase the whole video's frame rate.

RIFE / FILM: model_name selects the native checkpoint described below. RTX controls are ignored.
RTX Frame Generation: requires a supported NVIDIA CUDA GPU and nvidia-vfx exposing VideoFrameGeneration.run_at_timestep. No RIFE/FILM checkpoint is needed; model_name is ignored, including the (no native models) choice on installations without checkpoints. rtx_mode offers LOW/MEDIUM/HIGH (HIGH default); rtx_device_id selects the CUDA index (0 default). Uses SDK RGB8 processing and preserves the IMAGE dtype, not HDR/10-bit processing.

Both backends run forward/reverse pair interpolation with complementary timesteps. RTX loads one effect per execution, copies SDK-owned results before reuse, and closes it on completion or failure. Automatic/forced cut detection is disabled for deliberate tail/head morphing. It uses the frame-generation node's synchronized default-stream helpers, not the upscaler's single-input runner. Small real RTX 5090 tests cover overlap/short-clip bridges, CPU/CUDA inputs and default/nondefault caller streams; this is not a guarantee of seamless visual quality on arbitrary footage.

Native model selector

Place a ComfyUI-compatible RIFE or FILM .safetensors checkpoint in models/frame_interpolation/ (or a configured equivalent path), then select it in model_name. The dropdown can be converted to an input for an external checkpoint filename.

Requires ComfyUI's native Frame Interpolation Model Loader; no additional dependencies. RIFE 4.26 inference was verified locally. FILM uses the same native loader but has not been inference-tested with a local checkpoint.

Automatic processing

Seam selection: compares reduced-resolution previews using exposure-compensated MSE/PSNR, local SSIM, edges, temporal differences and scene discontinuities. Tests trims of up to 10% at each end and several overlap lengths.
Transition: morphs overlapping tail/head sequences with bidirectional interpolation, eased timing and bounded exposure correction. Middle frames stay unchanged; the playback wrap lies in the original sequence. Frame count and starting frame can change.
Short/static clips: very short batches use an interpolated endpoint bridge. Static batches pass through unchanged; already-closed batches skip interpolation and drop the duplicated endpoint when exact_endpoint is off.
Memory: interpolates one frame pair at a time. RIFE/FILM uses native ComfyUI model management; RTX retains one SDK effect for the transition. Resolution and dtype are preserved; processed output uses the configured intermediate device. Full batches still require memory proportional to duration; the loop does not inherit the separate frame-generation node's mmap/output-budget controls.

Exact endpoints versus smooth playback

exact_endpoint is off by default. Leave it off for continuous cyclic playback. Enable it to append a copy of the first output frame, making the returned first/last frames pixel-identical but adding a duplicated frame on repeat. Pixel equality is not guaranteed after lossy encoding.

Arbitrary footage may still show visible morphing, especially with cuts, occlusion, camera travel or incompatible head/tail motion. Logs report trim, overlap, preview PSNR and temporal spikes; strong discontinuities trigger a warning. Preview PSNR is not a video-wide quality score.

References

RIFE: intermediate-frame flow estimation.
FILM: large-motion frame interpolation.
WhiteRabbit: inspiration for seam preparation and loop assembly.

4.2 KiB Raw Blame History