Video Tile Disk Merge

The in-memory merge's problem is that it wants every processed tile in RAM at once. Video Tile Disk Merge is the disk-backed answer: it loads the saved .pt tiles from the job folder one at a time and merges them into the final IMAGE, so the only big tensor that has to exist is the output itself. It's the connected-graph version of the disk merge - it takes a tile_job handle rather than a folder path, which makes it the natural choice when you're in the same workflow that created the job.

The inputs that matter

• tile_job - the manifest from Disk Job or Disk Open Job.
• feather - same meaning as the in-memory Video Tile Merge: fraction of local tile width/height used for seam ramps.
• require_all_tiles (default true) - this is the good default. When true, the node stops and waits (reports the saved count and exact missing indices) until every expected tile file exists. Turn it off only if you want a partial merge.

Optional: feather_curve (linear, ease_in, ease_out, ease_in_out), blend_mode (alpha_over / weighted_average), merge_device (cpu default, auto, or cuda), and audio_override (an AUDIO that takes priority over stored audio.pt - mainly for older jobs created before audio storage existed).

What comes out

• IMAGE - the merged full frame batch, ready for your encoder (VideoHelperSuite-compatible).
• audio - stored or override audio, so the final export can carry sound without the original audio loader.

Why you'd pick this over Folder Merge

The pack has two disk merges now, and the split is simple: Disk Merge takes a tile_job connection and stays in the connected graph; Disk Folder Merge takes a folder path and can live in a completely separate export workflow. If you're already in the job's graph, Disk Merge is the least-friction path. The README keeps it "available for connected graphs" even as it pushes the folder-based version for standalone export.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/maDcaDDie2000/comfyui-video-tiler

Restart; it's under Video Tiler/Disk. No extra dependencies, no model downloads. Apache 2.0, vibe-coded, not actively maintained.

The memory math is worth internalizing: disk merge fixes the tile-cache memory, not the output memory. The final merged IMAGE is still one tensor, so merge_device=cpu is the lowest-VRAM path and cuda the fastest when VRAM allows. On a very long clip, plan for the output tensor's RAM regardless of which merge you use.