Smart Video Chunker

This is the node this pack is built around. Smart Video Chunker takes a long video with audio and slices it into segments for LTX 2.3 or Wan 2.2, putting the cuts on silence so a speaker never gets chopped mid-word, resampling to the target model's FPS, and snapping every chunk to the model's valid frame-count grid so the sampler doesn't reject it. In one node: chunk on speech boundaries, resample, and satisfy LTX's exacting 8n+1 (or Wan's 4n+1) frame rules - the three things that otherwise take a wall of math and switch nodes.

Why chunk at all? The README is blunt: chunking is required to get around VRAM issues. You generate long content as a series of smaller passes. And the frame-snapping matters because LTX 2.3's constraints are hard - off-grid frame counts fail rather than round, per the model card. If you've ever hit "invalid frame count" errors on LTX, this is the node that makes them stop.

How it works

Internally it calls the pack's TKSmartAudioChunker to find silence-based timing, then does the video work in steps:

1. Slice at native FPS. Using source_fps (the true fps of your incoming video tensor), it computes the frame window for the chunk and pads it slightly so resampling never comes up short.
2. Resample to target_fps (25 for LTX, 16 for Wan - the tooltip's defaults) via frame-index sampling.
3. Snap to the model's grid - model_type of LTX rounds to 8n+1, WAN to 4n+1, always rounding up so the model never gets fewer frames than the true duration.
4. Slice the audio to match, sample-accurately - padding with silence if the audio runs out.
5. Carry actual_end_time forward via the start_time_override input, so chunk N+1 starts exactly where chunk N ended, even after snapping trimmed or padded a boundary.

The index input (zero-based, from your loop) picks which chunk this pass produces; leave start_time_override at -1 for index 0.

The inputs that matter

• video + audio - the sources, an IMAGE tensor and AUDIO.
• index - loop driver.
• chunk_secs (target segment size) and variation (how far to search for silence). Lower chunk_secs for low VRAM.
• source_fps vs target_fps - the single most error-prone pair. source_fps must be your video's real rate (from a metadata node), target_fps what your model wants.
• model_type - LTX or WAN; it decides the frame grid.

Outputs

Eight sockets, but you mainly care about four: chunkOImages (the sliced IMAGE), chunkOfAudio (the matched AUDIO), numberFrames (frames after snapping - what the sampler actually generates), and num_chunks (total, for loop bounds). The timing outputs (actual_end_time, actual_start_time, chunk_duration) exist to be fed back into start_time_override for the next iteration so boundaries stay contiguous.

Common issues

Wrong source_fps is the classic: tell it 30 when the video is really 24 and every chunk's timing and frame count drift off. Pull the fps from your video loader, don't type from memory. And remember chunking is a compromise - Wan degrades past ~5s of content and LTX holds longer, so let your target model guide chunk_secs, not a round number. If you see generation_frames vs numberFrames differences in the log, that's the snap-vs-trim design working as intended, not an error.

Installing it

Part of trashkollector/TKNodes ("ComfyUI Handy Nodes"). Install via ComfyUI Manager (search "Handy Nodes") or:

cd ComfyUI/custom_nodes
git clone https://github.com/trashkollector/TKNodes

Restart. Needs the pack's stack - pydub + FFmpeg on PATH for audio, torch/torchaudio - but no model downloads beyond your usual video model.