Deep Exemplar Video Colorization (Original):
DeepExColorVideoNode colorizes a sequence of video frames using a color image as a visual reference. It applies Deep Exemplar’s reference-guided colorization method: the reference helps determine the colors, while each video frame supplies the scene detail and brightness. You can process frames independently or let the colorized result of one frame inform the next, which can help maintain color consistency through a sequence. The node also offers a faster, lower-resolution processing mode and optional model optimizations. It processes frames in their supplied order; the source does not show video-file encoding, audio handling, or a guarantee that every object will receive the intended color.
Deep Exemplar Video Colorization (Original) Input Parameters:
video_frames
An IMAGE input containing the frames you want to colorize. A standard batch of images is processed one frame at a time in batch order, so supply frames in the order you want the resulting sequence to follow. This input has no default or stated minimum or maximum frame count. Each frame is converted for processing, then collected into the output batch; an empty batch cannot produce a valid result.
reference_image
An IMAGE input containing the color image that guides the colorization. Choose an image whose palette and subject matter suit the look you want: the node extracts reference features once and uses them throughout the frame sequence. The source converts this input as a single image, so provide one reference image rather than a batch. There is no default or stated size range.
frame_propagate
A BOOLEAN setting that controls whether the previous colorized frame helps guide the next frame. Options are True and False; the default is True. When enabled, the first frame uses the reference image, and subsequent frames also receive color information from the immediately preceding result. This can improve continuity, though an unwanted color in one frame may influence later frames. When disabled, each frame is processed without that previous-frame color result.
use_half_resolution
A BOOLEAN speed-versus-detail setting. Options are True and False; the default is True. When enabled, the node runs its colorization models on images scaled to half the adjusted target width and height, then enlarges the predicted color information before producing the frames. This reduces the size of the model’s working images, but fine color boundaries may differ from full-resolution processing. Set it to False when preserving color detail matters more than processing speed; actual speed and quality gains depend on your hardware and footage.
target_width
An INT setting for the requested output width, with a default of 768, a minimum selectable value of 16, and a maximum of 4096. Before processing, the node adjusts the requested width to at least 64 and to a nearby multiple of 32, so the actual output width can differ from the value you enter. For example, 768 remains 768, while 16 becomes 64. The node pads or center-crops frames to the adjusted dimensions; it does not simply resize every source frame to fit.
target_height
An INT setting for the requested output height, with a default of 432, a minimum selectable value of 16, and a maximum of 4096. The same adjustment applies as for width: the processing height is at least 64 and a nearby multiple of 32. For example, 432 remains 432, while 16 becomes 64. Because frames are padded or center-cropped to the adjusted dimensions, check that your chosen height leaves important content visible.
use_torch_compile
A BOOLEAN option for attempting PyTorch model compilation. Options are True and False; the default is False. Enabling it may improve later processing speed, but the first run can take longer while compilation occurs. The source logs a compilation failure and continues rather than treating that failure as fatal, so a True value does not guarantee the optimization was applied. The reported setting records what you requested, not whether compilation succeeded.
use_sage_attention
A BOOLEAN option that enables the SageAttention path in the nonlocal model. Options are True and False; the default is False. This path is intended to speed up attention computation, but it depends on the installed sageattention package and a compatible environment. The node sets the option without checking installation or measuring a speed gain. Leave it disabled unless your setup supports it.
Deep Exemplar Video Colorization (Original) Output Parameters:
colorized_frames
An IMAGE batch containing the processed frames in the same order as the input. Each frame has the adjusted target width and height, with three RGB color channels; the node assembles the frames into a single batch for downstream image or video workflow nodes. These are colorized images, not an encoded video file, and this output does not contain audio or a frame-rate setting.
performance_report
A STRING containing a readable processing summary. It records the run date and time, number of frames, adjusted resolution, total elapsed time, average frames processed per second, time per frame, and the requested states of frame propagation, half-resolution processing, Torch compilation, and SageAttention. Use it to compare runs with different settings, while remembering that total time includes setup performed during the run and that the optimization flags do not prove an optimization succeeded.
Deep Exemplar Video Colorization (Original) Usage Tips:
- Start with
frame_propagate=Truefor a continuous shot, then inspect several frames throughout the sequence. If an incorrect color persists into later frames, tryFalseor process the affected shot separately. - Keep
use_half_resolution=Truefor an initial preview. For a final pass where small objects or color edges matter, compare a short sample withuse_half_resolution=Falsebefore committing to the full sequence. - Choose
target_widthandtarget_heightclose to your source framing and in multiples of32when practical. The node can center-crop or pad images to the adjusted size, so preview a frame to check for cropped subjects or added borders. - Use one reference image with colors relevant to the whole sequence. When a video changes scene or lighting substantially, process shots separately with references suited to each shot.
- Try
use_torch_compile=Trueprimarily for longer runs where later-frame gains might offset compilation time. Check the console for a compilation-failure message, and compare theperformance_reportacross runs on the same hardware. - Enable
use_sage_attentiononly after confirming the optional package works in your environment. If you encounter an attention-related error, turn it off and rerun. - Provide at least one frame and keep your GPU’s available memory in mind when increasing target dimensions or disabling half-resolution processing. The models are loaded onto CUDA, and larger working images require more memory.
Deep Exemplar Video Colorization (Original) Common Errors and Solutions:
Deep Exemplar checkpoints not found and download failed: ...
- Explanation: The node could not obtain the required colorization model files. The message includes the expected checkpoint paths.
- Solution: Check your internet connection and whether the node can write to its
checkpoints/video_moredata_l1directory. If your installation provides a manual checkpoint setup procedure, place the required files at the paths shown in the error.
VGG19 model not found and download failed: ...
- Explanation: The reference-feature model file could not be found or downloaded.
- Solution: Check download access and write permission for the node’s
datadirectory. Follow the installation instructions for obtainingvgg19_conv.pthif automatic download remains unavailable.
Torch not compiled with CUDA enabled or No NVIDIA driver on your system
- Explanation: The source moves its models and working images to CUDA, but the installed PyTorch build or machine cannot provide a working CUDA device.
- Solution: Run the node in a compatible NVIDIA CUDA environment with a suitable PyTorch installation and driver. Turning off
use_torch_compiledoes not remove the node’s CUDA requirement.
CUDA out of memory
- Explanation: The models or working images exceed available GPU memory. Larger target dimensions and full-resolution processing can increase memory use.
- Solution: Lower
target_widthandtarget_height, enableuse_half_resolution, and close other GPU-heavy tasks before retrying. Process a shorter batch if memory remains limited.
No module named 'sageattention' or an attention-related import error
- Explanation: The optional SageAttention path may require a package or compatible build that is absent from your environment; the exact message depends on the imported model implementation.
- Solution: Set
use_sage_attention=Falseto use the standard path. Enable it again only after installing and checking the package required by your node installation.
[ Deep Exemplar Video Colorization (Original)] torch.compile failed: ...
- Explanation: The requested compilation optimization failed. The source catches this error, prints its details, and continues processing, so this message alone does not mean the colorization failed.
- Solution: Leave
use_torch_compile=Falseif the run otherwise works. If you need compilation, use the printed details to check PyTorch and CUDA compatibility; do not interpretTorch Compile: Truein the report as confirmation of success.
RuntimeError: stack expects a non-empty TensorList
- Explanation: No processed frames were available when the node attempted to assemble its output. An empty
video_framesbatch is a likely cause. - Solution: Confirm the upstream node supplies at least one frame, then rerun.
Unsupported image shape or an error while converting an input image
- Explanation: The source converts each input to a PIL image and expects an image shape it can interpret, such as a standard ComfyUI image frame. A malformed frame or a reference batch containing more than one image may fail during conversion.
- Solution: Supply a valid
IMAGEbatch forvideo_framesand a singleIMAGEforreference_image. Check the upstream node’s image shape if the error continues.
An unexpected-size, crop, or padding result
- Explanation: This may be an output surprise rather than a software error. The node adjusts requested dimensions to a minimum of
64and a multiple of32, then center-pads or center-crops inputs to that size. - Solution: Check the adjusted resolution in the console or
performance_reportand preview an output frame. Revise the target dimensions or prepare the source framing upstream when important content is being cropped.
