Memory-conscious video/image tiling for ComfyUI with overlap tiles, gaps, and feather blending. The normal slicers use tensor views where possible, and the merge nodes write into a single output buffer.

This pack was built and tested primarily for LTX 2.3 and MiniMax H3 tiled video upscale workflows. Other models or node stacks may work, but they are not the main target.

Hardware disclaimer: Tiled video upscaling is still a demanding workflow intended for systems with high hardware specifications. Tiling reduces per-step memory, but long clips and high resolutions can still require a lot of VRAM, system RAM, processing time, and disk space. If the in-memory path does not fit, use the disk-backed workflow to process and save one tile at a time instead of keeping every processed tile in memory.

Maintenance disclaimer: This node pack was vibe-coded for personal use and may not be actively maintained. Issues, compatibility updates, or pull requests are not guaranteed to be addressed. Use it at your own discretion.

What Changed Recently

Added optional merge_device to merge nodes: auto, cpu, or cuda.
- auto keeps the first tile's device.
- cpu merges in system RAM and returns a CPU IMAGE.
- cuda uses VRAM when CUDA is available.
Added disk-backed tile nodes for workflows where processed tiles are saved one by one and merged later; Disk Merge can wait until all expected tile files exist.
Fixed direct TILE_CONFIG handling for helper/merge nodes.
Fixed small-frame Reference Color Match crashes by falling back to replicate padding when reflect padding is too large.
Removed tracked Python bytecode and added .gitignore entries for cache/output folders.

Features

Video Tile Slicer (var. size): grid layout with explicit tile counts, gaps, seam tiles, and overlap extension.
Video Tile Slicer (fixed size): constant tile size, fractional overlap, and traversal patterns.
Video Tile Merge: reconstructs either layout with adjustable feather, feather curve, blend mode, and merge device.
Disk-backed workflow: process one tile per run, save numbered .pt tile files, and stream-merge saved tiles later.
Reference tile alignment: cuts matching reference tiles from the same tile_config.
Reference color match: post-merge low-frequency color pull toward a reference clip.
Audio present check: detects real audio streams/waveform energy from AUDIO bundles.
Sampler timing: measures sampler wall-clock time and exposes averages as workflow FLOATs.

Tiling reduces per-step VRAM, but total workflow memory still depends heavily on clip length, resolution, channels, dtype, and how ComfyUI caches nodes. Disk-backed tile saving is the lowest-VRAM path for the expensive upscale branch because processed tiles do not all need to remain in graph memory.

Installation

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

Restart ComfyUI. The regular nodes, including the sampler timing nodes, appear under Video Tiler. Disk-backed nodes appear under Video Tiler/Disk.

Node List

| Internal class | Display name | |---|---| | VideoTileSlice | Video Tile Slicer (var. size) | | VideoTileSliceFixed | Video Tile Slicer (fixed size) | | VideoTileMerge | Video Tile Merge | | VideoTileMergeOverlapSoft | Video Tile Merge (overlap soft) | | GetTile | Get Tile | | ReferenceTileSlice | Reference Tile Slice | | VideoTileReferenceColorMatch | Video Tile Reference Color Match | | VideoTileAudioFFprobeLTX | Video Tile Audio Present | | VideoTileSamplerTimerStart | Sampler Timer Start | | VideoTileSamplerTimerResult | Sampler Timing Result | | VideoTileDiskJob | Video Tile Disk Job | | VideoTileDiskOpenJob | Video Tile Disk Open Job | | VideoTileDiskIndexes | Video Tile Disk Indexes | | VideoTileDiskGetTile | Video Tile Disk Get Tile | | VideoTileDiskSaveTile | Video Tile Disk Save Tile | | VideoTileDiskMerge | Video Tile Disk Merge | | VideoTileDiskFolderMerge | Video Tile Disk Folder Merge | | VideoTileDiskPreview | Video Tile Disk Preview |

Example Workflows

Both supported targets include a normal in-memory example and a disk-buffered example:

| Target | Normal workflow | Disk-buffered workflow | |---|---|---| | LTX 2.3 | Video Tiler LTX 2.3 upscaling workflow | Video Tiler LTX 2.3 disk buffered workflow | | MiniMax H3 | Video Tiler MiniMax H3 upscaler workflow | Video Tiler MiniMax H3 disk buffered workflow |

The normal examples keep the processed tiles in memory. The disk-buffered variants save one indexed tile per queued run, then merge the saved tiles and export the video in a second pass. Each disk-buffered workflow contains notes describing when to increment the tile index and when to enable the merge pass.

Sampler Timing

Use the two timing nodes around a standard KSampler, KSampler Advanced, or custom sampler pipeline:

Connect the checkpoint/model loader's MODEL to Sampler Timer Start.
Connect its model output to the sampler's model input. For a custom sampling pipeline, connect it where the MODEL enters the guider/pipeline.
Connect the sampler's LATENT output to Sampler Timing Result samples.
Connect sampler_timer between the two timing nodes.

Sampler Timing Result passes the LATENT through unchanged and provides:

| Output | Meaning | |---|---| | average_seconds | Average wall-clock seconds per completed sampler call. This is the main output for tiled/list sampling. | | total_seconds | Total sampler time in the current queued workflow execution. | | seconds_per_step | Total time divided by the number of sampled sigma intervals. | | sampler_calls | Number of sampler calls included in the average. | | total_steps | Number of sigma intervals included in seconds_per_step. |

The FLOAT outputs can connect to any node that accepts a FLOAT. GPU work is synchronized at the measurement boundaries for useful wall-clock results. Model loading, VAE decode, and other nodes are not included. Timing data is reset for every queued workflow execution.

Standard In-Memory Workflow

Use this when the full processed tile list fits comfortably in memory.

Video Tile Slicer (var. size) or Video Tile Slicer (fixed size)
Process tiles with your upscale branch.
Connect processed tiles plus the original tile_config into Video Tile Merge.
Tune feather, feather_curve, blend_mode, and merge_device.

The slicer tiles output is a list. ComfyUI can run downstream nodes in list context, which is usually much lighter than processing the full frame at once. The final merge still needs the output image/video buffer, and in the in-memory workflow ComfyUI may also keep processed tile results cached.

Disk-Backed Low-VRAM Workflow

Use this when the expensive upscale branch cannot keep all processed tiles cached, or when you want to run very long clips on smaller PCs.

Pass 1: Save Processed Tiles

Run either slicer to get tile_config.
Connect tile_config to Video Tile Disk Job.
- job_name: stable folder name for this run.
- output_folder: defaults visibly to the folder used for disk tile jobs; change it if needed.
- audio (optional): connect the source AUDIO once to store it with the job for independent export.
- Output tile_job is the manifest path used by the other disk nodes.
Choose a tile index.
- Manual/reliable path: set tile_index yourself and queue one run per tile.
- Helper path: use Video Tile Disk Indexes to emit 0..tile_count-1 for list execution or queue automation.
Use Video Tile Disk Get Tile with original images, tile_job, and tile_index.
Send that single tile through your expensive upscale branch.
Save the processed result with Video Tile Disk Save Tile.

Saved files are exact PyTorch tensor files:

manifest.json
audio.pt          # present when source AUDIO was connected to Disk Job
tile_00000.pt
tile_00001.pt
tile_00002.pt
...

The numbering matches the original tile order in tile_config. The manifest records frame size, tile geometry, tile count, and saved tile metadata.

Pass 2: Independent Folder Merge and Video Export

Pass 2 can live in a completely separate ComfyUI workflow and does not need an active slicer, Disk Job node, source video, or tile-processing branch:

Add Video Tile Disk Folder Merge.
Paste the completed job folder or its manifest.json path into job_folder.
Set feather, feather_curve, blend_mode, and merge_device to the values you want for the final assembly.
Connect its IMAGE and audio outputs to your video encoder (for example, VideoHelperSuite Video Combine), then configure frame rate and format there.
Queue this output branch independently. The node rescans the folder on every run. It stops with the saved count and exact missing tile indices until every expected file is present, then streams the tiles into the final frame batch.

For jobs created before audio storage was added, connect the original audio loader to Folder Merge's optional audio_override input. The resulting audio output can still be wired to the encoder without regenerating any tiles. An override takes priority over stored audio.pt.

If an existing graph still uses a TILE_JOB connection, Video Tile Disk Open Job can open the folder and provide that connection without rerunning the slicer. The original Video Tile Disk Merge remains available for connected graphs.

This avoids loading every processed tile at once. The final merged IMAGE still exists as one tensor, so very long videos can still require a lot of system RAM if merge_device=cpu or VRAM if merge_device=cuda.

Review Tiles While Pass 1 Is Running

Use Video Tile Disk Preview as a separate review branch or in a small review-only workflow:

Point job_folder at the same disk job folder.
Choose tile_index and frame_index. Set frame_index=-1 to output every frame stored in that tile.
Connect preview to ComfyUI's Preview Image node.
Leave missing_tile=nearest_available to keep reviewing while the requested tile is pending, or choose error when you only want the exact tile.

The preview node reports the actual tile/frame shown and the current saved_count / tile_count. It rescans the folder each time it runs, so it does not need to be connected to or wait for the active tile branch.

Disk Node Details

Video Tile Disk Job

Creates or updates a disk job manifest from a slicer tile_config.

| Input | Description | |---|---| | tile_config | From either slicer. | | job_name | Folder-safe name for the tile job. | | output_folder | Folder where disk tile job folders are written. The widget shows the default path. | | audio | Optional source AUDIO saved as audio.pt for the independent merge/export workflow. |

Outputs: tile_job, manifest_path, tile_count, status.

Video Tile Disk Open Job

Opens a saved job independently from the slicer/processing workflow.

| Input | Description | |---|---| | job_folder | Job folder, manifest.json, or a parent folder containing exactly one job. |

Outputs: tile_job, manifest_path, saved_count, tile_count, status, and stored audio when available.

Video Tile Disk Indexes

Outputs an INT list of tile indices for a job. This can help queue all tile runs or drive ComfyUI list execution where your graph supports it.

| Input | Description | |---|---| | tile_job | Manifest from Disk Job. | | start_index | First tile index. | | end_index | Last tile index; -1 means final tile. |

Video Tile Disk Get Tile

Loads the manifest and extracts exactly one tile from the original IMAGE batch.

| Input | Description | |---|---| | images | Original full IMAGE batch [B,H,W,C]. | | tile_job | Manifest from Disk Job. | | tile_index | Tile index to extract. |

Outputs: tile, resolved tile_index, and tile_job passthrough.

Video Tile Disk Save Tile

Saves one processed tile as tile_XXXXX.pt and updates the manifest.

| Input | Description | |---|---| | tile | Processed tile IMAGE. | | tile_job | Manifest from Disk Job. | | tile_index | Tile index for numbering. | | overwrite | Replace an existing tile file for the same index. |

Video Tile Disk Merge

Loads saved tiles one by one from disk and merges them.

| Input | Description | |---|---| | tile_job | Manifest from Disk Job. | | feather | Same meaning as Video Tile Merge. | | feather_curve | Optional: linear, ease_in, ease_out, ease_in_out. | | blend_mode | Optional: alpha_over or weighted_average. | | merge_device | Optional: cpu default, auto, or cuda. | | require_all_tiles | Optional: default True; stop before merging until every expected tile file exists. | | audio_override | Optional AUDIO for older jobs; takes priority over stored audio.pt. |

Outputs: merged IMAGE and stored/override audio.

Video Tile Disk Folder Merge

Standalone final assembly node. It takes a folder/path widget instead of a TILE_JOB connection, always requires the complete expected tile set, and can therefore be queued in a separate export workflow. Connect its reconstructed IMAGE frame batch and stored/override audio outputs to your preferred video encoder.

| Input | Description | |---|---| | job_folder | Job folder, manifest.json, or a parent folder containing exactly one job. | | feather | Same meaning as Video Tile Merge. | | feather_curve | Optional seam-alpha curve. | | blend_mode | Optional: alpha_over or weighted_average. | | merge_device | Optional: cpu default, auto, or cuda. | | audio_override | Optional AUDIO for an older job without audio.pt; takes priority when connected. |

Outputs: merged IMAGE, status, tile count, and stored/override audio.

Video Tile Disk Preview

Loads one saved tile for interactive review, including while the job is incomplete.

| Input | Description | |---|---| | job_folder | Existing job folder or manifest path. | | tile_index | Tile to inspect. | | frame_index | Frame to inspect; -1 returns the complete saved frame batch. | | missing_tile | Use the nearest available tile or require the exact requested tile. |

Outputs: preview IMAGE, actual tile/frame indices, saved/total counts, tile path, and status.

Slicer Nodes

Video Tile Slicer (var. size)

Splits a video/image batch into a grid with normal tiles, gaps, and overlap/seam tiles.

| Input | Description | |---|---| | images | IMAGE [B,H,W,C]. | | tiles_x | 1-5 horizontal cells. | | tiles_y | 1-5 vertical cells. | | multiple | Tile size multiple, such as 8, 16, or 32. | | overlap_extension_x / overlap_extension_y | Pixels seams extend into adjacent tiles, snapped to multiple. |

Outputs: tiles, tile_config, visualization, tile_count, layout_label.

tile_config is geometry-only v4. Changing merge feather does not require re-slicing.

Video Tile Slicer (fixed size)

Uses fixed tile size and fractional overlap.

| Input | Description | |---|---| | images | IMAGE [B,H,W,C]. | | tile_width / tile_height | Tile size. | | multiple | Position/size snap multiple. | | overlap | 1/8, 1/4, 3/8, or 1/2 minimum neighbor overlap. | | pattern | row, column, spiral, or double_spiral. |

Outputs match the variable-size slicer. New slices emit tile_config v5. Legacy v3 configs still merge.

Merge Nodes

Video Tile Merge

Reconstructs full IMAGE output from processed tiles.

| Input | Description | |---|---| | tile_config | From either slicer. | | tiles | Processed tile list in slicer order. | | feather | 0.0 to 0.5, fraction of local tile width/height used for alpha ramps. | | feather_curve | Optional alpha remap. | | blend_mode | alpha_over or weighted_average. | | merge_device | auto, cpu, or cuda. |

alpha_over uses painter order with a coverage gate. weighted_average computes normalized geometry-weighted color sums.

Video Tile Merge (overlap soft)

Compatibility node for normalized weighted overlap merging. The main merge node can now do the same style with blend_mode=weighted_average.

Helper Nodes

Get Tile

Returns one tile by index from a full IMAGE or from a slicer tile list.

Reference Tile Slice

Cuts matching reference tiles from a reference image using the same tile_config.

Video Tile Reference Color Match

Post-merge color alignment. It resizes a reference clip, pulls low-frequency RGB toward it, and preserves merged high-frequency detail. Long batches are chunked to avoid PyTorch indexing limits.

Video Tile Audio Present

Returns true when an AUDIO bundle appears to contain real audio. If a path is available and ffprobe works, it requires at least one audio stream; otherwise it falls back to waveform energy.

Practical Notes

Disk tiles are .pt tensor files, not PNG/video files. This is intentional: it is lossless and preserves exact batch tensors.
Disk merge greatly reduces tile-cache memory, but the final merged IMAGE tensor still has to exist.
For the lowest VRAM during merge, use Video Tile Disk Merge with merge_device=cpu.
For fastest merge when VRAM is available, use merge_device=cuda.
Existing workflows that use the original slicer and merge nodes should continue to load because the disk workflow is added as separate nodes.

Compatibility

Primary tested targets: LTX 2.3 and MiniMax H3 workflows.
IMAGE tensors use ComfyUI layout [B,H,W,C].
Compatible with VideoHelperSuite-style IMAGE/AUDIO use.

License

Licensed under the Apache License 2.0.