Ten ComfyUI nodes for GPU-accelerated, float32 color adjustments, matching, and image downscaling, plus image batch selection, concatenation, and transitions.

cd /path/to/ComfyUI/custom_nodes
git clone https://git.colorized.life/gpu-layerstyles.git gpu-layerstyles
# Restart ComfyUI.

Nodes

| Display name | Node ID | Inputs and controls, in order | Outputs and behavior | | --- | --- | --- | --- | | GPU LayerStyles ColorBalance | GPULayerStyles_ColorBalance | image, cyan_red, magenta_green, yellow_blue, execution controls | image: balance RGB while preserving luminosity. | | GPU LayerStyles Brightness Contrast V2 | GPULayerStyles_BrightnessContrastV2 | image, brightness, contrast, saturation, execution controls | image: adjust brightness, per-frame contrast, then saturation. | | GPU LayerStyles ColorTemperature | GPULayerStyles_ColorTemperature | image, temperature, execution controls | image: negative values warm; positive values cool. | | GPU LayerStyles ColorMatch | GPULayerStyles_ColorMatch | image, image_ref, method, strength, execution controls | image: match RGB to the reference using MKL or MVGD; preserve target alpha. | | GPU LayerStyles ImageScaleDown | GPULayerStyles_ImageScaleDown | image, width, height, method, execution controls | image: downscale to exact dimensions with bicubic or Lanczos-3 filtering. | | GPU LayerStyles SliceImageBatch | GPULayerStyles_SliceImageBatch | image, slice | image: select frames with an index or START:END[:STRIDE]; negative indices and reverse strides are supported. | | GPU LayerStyles BatchConcat | GPULayerStyles_BatchConcat | images: growing sockets image_1 through image_100 | image: concatenate connected images or batches in socket order; dimensions, dtype, and device must match. | | GPU LayerStyles CrossFade | GPULayerStyles_CrossFade | images_1, images_2, start_index, frames, execution controls | image: keep the first batch's prefix, fade into the second batch over frames, then keep the rest of the second batch. | | GPU LayerStyles TwoBatchLoop | GPULayerStyles_TwoBatchLoop | images_1, images_2, blend_target, append_first_frame, execution controls | images: crossfade both batch boundaries into a loop; optionally append its first frame. Each input needs at least 2 * blend_target + 1 frames. | | GPU LayerStyles TwoBatchBridge | GPULayerStyles_TwoBatchBridge | images, blend_target | bridge_first: last blend_target frames; bridge_last: first blend_target frames; first/last: first frame of bridge_first, then last frame of bridge_last. |

Execution controls are output_device, then batch_size. All outputs are IMAGE batches; multiple outputs are listed in socket order.

Performance

42.11× faster with GPU output; 6.12× with CPU offloading in an informal, user-reported benchmark against ComfyUI_LayerStyle: a 720×1280 video with 525 frames, processed through ColorTemperature → ColorBalance → Brightness Contrast. Times cover the complete chain, with the execution cache cleared before each run. Hardware and repeat counts were not reported.

| Pipeline | Total time (ms) | Elapsed time reduction vs. upstream | Speedup vs. upstream | | --- | ---: | ---: | ---: | | GPU LayerStyles: CPU offloading | 11,157 | 83.7% | 6.12× | | GPU LayerStyles: GPU output | 1,621 | 97.6% | 42.11× | | ComfyUI_LayerStyle | 68,255 | 0.0% | 1.00× |

License

Christopher Hahne's color-matcher served as the original source for the MKL and analytical MVGD algorithms and supplies the numerical reference used by our tests. See COPYRIGHT for project and original-source attribution.

ImageScaleDown's Lanczos algorithm is adapted from Pillow. Its MIT-CMU notices and source attribution ship in source and wheel distributions alongside the project license.

The two vendored test photographs retain their upstream provenance and applicable licensing; see the fixture notices. The original upstream GPL text and attribution notices ship alongside the project license in source and wheel distributions.