ComfyUI VFX Nodes
by FugitiveExpert01
A growing collection of custom ComfyUI nodes built for VFX production pipelines — designed around the real demands of working with high-resolution imagery, video sequences, and AI-assisted visual effects workflows.
📌 Overview
This repo fills the gaps between off-the-shelf ComfyUI nodes and the specific needs of VFX work: large format images, temporal consistency across video frames, precise spatial control, and clean integration with diffusion-based upscaling and enhancement models.
Nodes are built with video batch support as a first-class concern — not an afterthought.
🧩 Nodes
🔲 TileSplit
Splits an image or video batch into an overlapping grid of tiles, ready to be passed individually to a model.
| Parameter | Type | Description |
|---|---|---|
image |
IMAGE | Input image or video batch |
tiles_x |
INT | Number of columns |
tiles_y |
INT | Number of rows |
overlap_percent |
FLOAT | Overlap between adjacent tiles as a fraction of tile stride |
alignment |
ENUM | Free · 8 (SD) · 16 (WAN / VACE) — snaps tile dimensions to the selected multiple |
Outputs: tiles (LIST) · debug_image · tile_calc
Features:
- Overlap calculated automatically and stored in
tile_calcfor seamless reconstruction - Outputs tiles as
(F, H, W, C)tensors — compatible with video model K-samplers - Debug image showing tile boundaries and overlap regions
- Alignment dropdown snaps tile dimensions to multiples of 8 (SD 1.5 / SDXL) or 16 (WAN 2.1 / VACE), preventing token count mismatches inside attention blocks
- Node footer reports tile count, per-tile tensor dimensions, and total memory footprint
🔳 TileMerge
Reconstructs a full image or video sequence from processed tiles using linear weighted blending for seamless, artifact-free joins.
| Parameter | Type | Description |
|---|---|---|
tiles |
IMAGE (LIST) | Processed tile sequences |
tile_calc |
TILE_CALC | Layout data from TileSplit |
feather_scale |
FLOAT | Scales the blend zone independently of overlap_percent. 1.0 = fade across the full overlap. 0.5 = tighter edge. 2.0 = wider, softer blend |
Outputs: image
Features:
- Linear feather masks at overlap regions — width controlled independently via
feather_scale - Weighted accumulation handles overlapping regions correctly
- Robust handling of tensor shapes returned by video models, including automatic resize if a model returns a slightly different spatial size
- Node footer reports output tensor dimensions and memory size
📌 ChromaPin
Pins a processed video's colours to a reference image by measuring colour drift at a single anchor frame and propagating the correction across the entire sequence.
The core workflow: supply the original reference image and tell ChromaPin which frame in the processed video corresponds to it. ChromaPin fits a colour-correction transform from the processed anchor frame to the reference, then applies that same transform to every frame — removing the model's colour drift uniformly without disturbing the natural colour variation between frames.
| Parameter | Type | Description |
|---|---|---|
video |
IMAGE | Processed video batch (F, H, W, C) |
reference_image |
IMAGE | Original reference image to correct towards |
reference_frame_index |
INT | 0-based frame index that corresponds to reference_image |
method |
ENUM | Colour transfer algorithm (see table below) |
strength |
FLOAT | Blend between original (0.0) and fully corrected (1.0) |
propagation |
ENUM | uniform or falloff — how the correction spreads across frames |
falloff_radius |
INT | (falloff only) Frames from the anchor at which strength reaches zero |
falloff_gamma |
FLOAT | (falloff only) Curve shape: 1.0 = linear, >1 = fast drop, <1 = slow drop |
Outputs: corrected_video · debug_comparison (three-panel: Reference / Before / After)
Methods:
| Method | Deps | Description |
|---|---|---|
mkl |
— | Monge-Kantorovich Linearization. Full 3×3 cross-channel Lab transform. Recommended default |
reinhard_lab |
— | Per-channel mean/std in CIE Lab. Good general purpose |
linear_rgb |
— | Per-channel gain/offset in sRGB. Fastest |
histogram |
— | Per-channel CDF matching. Best for non-linear shifts |
reinhard_lab_gpu |
kornia |
GPU-accelerated Reinhard; falls back to CPU if kornia is absent |
hm-mkl-hm |
color-matcher |
HM → MKL → HM compound. Best overall quality |
hm-mvgd-hm |
color-matcher |
HM → MVGD → HM compound |
hm |
color-matcher |
Histogram matching |
mvgd |
color-matcher |
Multi-Variate Gaussian Distribution transfer |
⚡ LoRA Load
Multi-LoRA loader with a custom folder-tree browser UI. Add any number of LoRAs from a searchable tree, toggle each on/off, and set per-LoRA model and CLIP strengths — all from inside the node.
| Parameter | Type | Description |
|---|---|---|
loras_json |
STRING (hidden) | Serialised LoRA list from the JS widget |
Outputs: lora_stack (FE_LORA_STACK)
Features:
- Folder-tree browser with search, per-row on/off toggle, and strength sliders
- Optional separate CLIP strength per LoRA (right-click the node)
- Module-level weight cache — identical files shared across multiple nodes or tile streams are read from disk only once
- CivitAI lookup via SHA-256: automatically fetches model name, base model, trained trigger words, and preview images; results are cached to a
.fe-info.jsonsidecar file
⚡ Apply LoRA
Applies a
FE_LORA_STACKfrom LoRA Load to a MODEL (and optionally CLIP). Architecture-agnostic: SD1, SDXL, Flux, WAN 2.1/2.2, HunyuanVideo, and others.
| Parameter | Type | Description |
|---|---|---|
model |
MODEL | Model to patch |
lora_stack |
FE_LORA_STACK | Stack from LoRA Load |
application_mode |
ENUM | Stack or Merge (see below) |
clip |
CLIP | (optional) CLIP to patch alongside the model |
Outputs: model · clip
Application modes:
| Mode | Description |
|---|---|
Stack |
Each LoRA applied as a sequential patch via the model patcher. Safe with any combination of LoRAs. Default |
Merge |
All LoRA weight deltas are pre-scaled and summed into a single combined dict, then one patch is applied. Best when LoRAs share many of the same target layers |
🔍 LoRA Trigger Analysis
Analyses LoRA weight deltas against all text encoders present in the wired CLIP to surface candidate trigger words. Architecture-agnostic — encoders are discovered dynamically, covering CLIP-L/G, T5-XXL, LLaMA/Gemma, and any dual/triple encoder combination.
| Parameter | Type | Description |
|---|---|---|
lora_stack |
FE_LORA_STACK | Stack from LoRA Load |
clip |
CLIP | CLIP object to analyse against |
top_k |
INT | Number of candidate tokens to return per encoder |
Outputs: candidate_triggers (STRING)
How it works: For each text-encoder layer in the LoRA whose in_features matches a discovered encoder's embedding dimension, the full token embedding table is projected through the lora_down input subspace and L2 activation norms are accumulated. High-scoring tokens are those most aligned with what the LoRA was trained to respond to. Results are labelled per-encoder when multiple are present.
🔤 Text List → Batch / Text Batch → List
Bidirectional converters between ComfyUI LIST and batched STRING types, with optional delimiter joining.
⚙️ Installation
1. Clone into your ComfyUI custom nodes folder:
cd ComfyUI/custom_nodes
git clone https://github.com/FugitiveExpert01/ComfyUI-FEnodes.git
2. Restart ComfyUI.
Nodes will appear in the node menu under the FEnodes category.
Core functionality requires no additional dependencies beyond what ComfyUI already provides (PyTorch, NumPy, Pillow).
3. Optional dependencies (install to unlock additional ChromaPin methods):
pip install kornia # enables: reinhard_lab_gpu
pip install color-matcher # enables: hm, mvgd, hm-mkl-hm, hm-mvgd-hm
🎬 Typical Workflows
Tiled video diffusion:
Load Video → TileSplit → [K-Sampler per tile] → TileMerge → Save Video
Tiled diffusion with colour correction:
Load Video → TileSplit → [K-Sampler per tile] → TileMerge
→ ChromaPin (+ reference frame) → Save Video
LoRA workflow:
Load Checkpoint → LoRA Load → Apply LoRA → [K-Sampler] → Save
LoRA Trigger Analysis → (use triggers in prompt)
Combined:
Load Video → TileSplit → [K-Sampler per tile] → TileMerge
→ ChromaPin
🗺️ Roadmap
| Node | Description | Status |
|---|---|---|
| TileSplit | Grid tile splitting for video batches with alignment dropdown | ✅ Released (v0.0.4) |
| TileMerge | Linear weighted tile reconstruction with independent feather control | ✅ Released (v0.0.4) |
| Text List → Batch | LIST to batched STRING conversion | ✅ Released |
| Text Batch → List | Batched STRING to LIST conversion | ✅ Released |
| ChromaPin | Anchor-based colour correction across video sequences | ✅ Released (v0.0.1) |
| LoRA Load | Multi-LoRA browser UI with CivitAI lookup and weight cache | ✅ Released (v0.0.6) |
| Apply LoRA | Stack or merge LoRA application, architecture-agnostic | ✅ Released (v0.0.6) |
| LoRA Trigger Analysis | Encoder-agnostic trigger word surface analysis | ✅ Released (v0.0.6) |
Have a node idea or a production use case that isn't covered? Open an issue.
🤝 Contributing
Pull requests are welcome. If you're adding a node, please:
- Keep video batch support (
F, H, W, C) as the primary tensor convention - Add a docstring describing what the node does and its input/output types
- Test with both single images and multi-frame video batches
- Set
CATEGORY = "FEnodes"so nodes appear grouped in the menu
📄 License
Apache License 2.0 — free to use in personal and commercial VFX pipelines.