OmniNodes Introduction
OmniNodes is a collection of custom nodes for ComfyUI. It gives you more ways to build image, audio, and video workflows: preparing prompts, comparing variations, refining details, inspecting models, processing finished images, and keeping track of how an artwork was made.
In ComfyUI, a node is a small step in a visual workflow. You connect nodes to pass an image, prompt, model, or other result from one step to the next. OmniNodes adds steps that can save repetitive wiring. For example, you can generate several seed variations with one sampling node, label the results in a comparison grid, and save a record of the settings used.
The nodes appear in ComfyUI under TensorVizion submenus. Some model and sampling tools appear in more than one submenu, so searching for a node by its name is often easier than browsing categories.
How OmniNodes Works
Think of a ComfyUI workflow as a production line for an artwork. One node supplies a prompt, another loads a model, a sampler generates an image, and later nodes polish or save it. OmniNodes fits into that line at different points:
- Prepare: Clean or vary a prompt, load a model, or create a mask.
- Generate: Sample with different seeds, combine conditioning, or refine part of an image.
- Inspect: Check a latent, read image metadata, or compare model and LoRA information.
- Finish: Adjust color, apply an upscale model, assemble video frames, or save the result.
- Record or automate: Write a workflow manifest, process a prompt list, or interact with a web service.
A latent is an intermediate representation used during image generation. It is not yet a finished picture. OmniNodes includes tools for blending, inspecting, and checking latents before they are decoded into images. Where noted below, sampling nodes use ComfyUI’s own sampling functions; their main benefit is bringing useful combinations of steps into a more convenient node.
OmniNodes Features
Audio and audio-reactive work
| Node | What you can use it for |
|---|---|
| Audio Beat Detect | Find likely beats, their timestamps, and an estimated tempo for timing visual changes. |
| Audio Loudness Match | Bring one clip closer to a reference clip’s perceived loudness. |
| Audio Mixer | Combine up to four stereo tracks, adjusting each track’s gain, pan, or mute setting. |
| Audio Normalize | Adjust peak or average loudness toward a chosen target level. |
| Audio Pitch Shift | Shift pitch by up to 24 semitones in either direction. |
| Audio Reverb | Add a sense of space using algorithmic or convolution-style reverb. |
| Audio Sidechain Duck | Lower one sound when another becomes louder, such as reducing music beneath speech. |
| Audio Spectrogram | Turn frequency information into an image; choose linear or logarithmic power scaling and a color map. |
| Audio Stem Splitter (Freq Band) | Separate bass, middle, and high frequencies. This is not an AI vocal-and-instrument separator. |
| Audio-to-Latent Modulator | Convert changes in audio level into values you can use for audio-reactive, frame-by-frame generation. |
| Audio Transient Shaper | Emphasize or soften the initial impact and continuing sound of a clip. |
| Audio Waveform | Make a waveform image for inspection or presentation. |
Image editing, comparison, and saving
| Node | What you can use it for |
|---|---|
| Background Remover | Isolate a foreground subject and produce a mask. Choose a neural cutout, OpenCV GrabCut, or a flat-color chroma key; the neural option falls back to GrabCut when rembg is unavailable. |
| Image Resize | Set an exact size or scale factor. Choose whether the image stretches, fits within, or fills the requested dimensions, along with an interpolation filter. |
| Apply Upscale Model | Run a loaded upscale model on an image; optionally resize the result to a target long-edge size. |
| Quick Save Image | Save a single image in ComfyUI’s managed output area, with optional seed, model, and notes stored in the PNG. |
| 3D LUT Apply | Apply a .cube color look and adjust its strength. It supports 3D LUTs, not 1D LUTs. |
| Contact Sheet Maker | Arrange a batch into an unlabeled grid for quick browsing. |
| Custom Folder Batch Saver | Save a batch to a chosen folder with numbered filenames. Unlike Quick Save Image, this does not save into ComfyUI’s managed output area. |
| Aspect Ratio Bucket | Snap an image to a nearby standard SD or SDXL training aspect ratio. |
| Image Blend | Combine two images using a blend mode, ratio, strength, and optional mask. |
| Image Color Grade | Adjust exposure, contrast, saturation, gamma, lift, gain, temperature, and tint. |
| Image Grid Compare | Make a labeled comparison grid, useful for judging seeds, samplers, or strengths side by side. |
| Face Detect & Crop | Locate faces and return crops and a detection mask for later refinement. |
| Image Mask Composite | Apply localized darkening, brightening, blur, or color effects through shape-based masks. |
| Image Noise Inject | Add grain-like noise, including a monochrome option. |
| Image Sharpen & Blur | Sharpen or soften an image in one node. |
| Image Vignette & Glow | Darken the edges and add glow around bright areas. |
| Load Image + Recovered Metadata | Load a PNG and expose available generation settings, such as seed, prompts, steps, CFG, sampler, and checkpoint, as separate outputs. |
| Mask Morphology | Grow, shrink, feather, or invert a mask. |
| Resize to Multiple | Pad, crop, or stretch an image to dimensions divisible by a chosen number, such as 8. This serves a different purpose from ordinary Image Resize. |
| Text Overlay | Place text on an image and control its font, size, color, position, outline, and background box. |
Working with recovered metadata: Connect the recovered seed to a seed tool or the positive prompt to a text-encoding node instead of retyping them. The node can read supported ComfyUI workflow information or A1111/Forge-style PNG text. If a picture has no supported embedded data, it still loads, but metadata_found is false and metadata outputs use empty or zero defaults. In workflows with several text-encoding nodes, its positive/negative prompt identification is a guess; check the recovered text before relying on it.
Latent inspection and control
| Node | What you can use it for |
|---|---|
| Latent Anomaly Mask | Mark unusual latent regions and return a corrected latent and mask. |
| Latent Blend | Mix two latents by a chosen weight. |
| Latent Channel Mixer | Adjust how latent channels contribute to the result. |
| Latent Histogram | View the distribution of values in each channel and an outlier statistic. |
| Latent Interpolate | Create transitions between two latents using linear or spherical interpolation. |
| Latent Mask | Make rectangle, ellipse, or gradient masks in latent space. |
| Latent Noise Inject | Add a controlled amount of noise to a latent. |
| Latent Palette Extractor | Extract a latent signature for comparison. |
| Latent QC Gate | Check for invalid values, suspiciously flat output, or excessive outliers before decoding; optionally substitute a fallback latent. |
| MiniMax H3 AV QC Gate | Check the video and audio portions of a MiniMax H3 audio/video latent separately. An ordinary latent check cannot handle this combined structure. |
| Latent Structure Probe | Display a heatmap of latent activity. |
| Latent Visualizer | Create a viewable preview of latent channels and their statistics. |
These checks are diagnostics, not aesthetic judgments. A passing latent can still produce an image you dislike, while an unusual-looking latent might be intentional.
Model, LoRA, and conditioning tools
| Node | What you can use it for |
|---|---|
| Universal Checkpoint Loader | Load a checkpoint’s model, text encoder, and VAE without choosing an SDXL-only loader; read its filename-based family guess as a clue, not proof. |
| Simple SDXL Loader | Load the model, text encoder, and VAE for an SDXL checkpoint together. |
| Batch Folder Loader | Select a checkpoint by name from a registered checkpoint subfolder. |
| VAE Loader | Load a VAE and see a summary. |
| Upscale Model Loader | Load an upscale model to use with Apply Upscale Model. |
| ControlNet Loader | Load a ControlNet model and see a summary. |
| ControlNet Preprocessor | Turn an image into canny edges, line art, or a lightweight depth estimate. The depth_lite estimate is a visual heuristic, not a trained depth model. |
| ControlNet Apply (Advanced) | Apply a ControlNet image to conditioning, controlling its strength and the portion of sampling during which it acts. |
| ControlNet Stack | Combine up to four ControlNets, each with its own image, strength, and timing window. |
| Image Prompt Loader | Load a vision encoder and style model pair for reference-image conditioning. |
| Image Prompt Apply | Apply a reference image through that style model, with adjustable strength. |
| CLIP Text Compare | Compare how two prompts are represented by text conditioning. |
| CLIP Text Weight | Increase or decrease the influence of CLIP conditioning. |
| Dual Model Merger / Model Merge Weighted | Mix two models with a chosen weighting. |
| Model Block Freeze | Freeze selected U-Net blocks for partial fine-tuning workflows. |
| Model Info Inspector | Read a loaded model’s key count, architecture guess, and precision. |
| LoRA Info Inspector | Inspect a LoRA’s rank, alpha, and target modules. |
| LoRA Metadata Diff | Compare two LoRA files’ metadata for structural compatibility. |
| LyCORIS Format Inspector | Check whether a file’s tensor names indicate LoRA, LoHa, LoKr, DoRA tagging, or a mixture. Run it before choosing a specialized loader. |
| LoHa Loader / LoKr Loader / DoRA Loader (Custom) | Apply those specialized adaptation formats using their respective weight-merging methods. Check each node’s summary for skipped layers. |
| LoRA Stack | Apply several LoRAs to a model/text-encoder pair. |
| Quick LoRA Stacker | A simpler option for a single LoRA. |
| Multi-LoRA Weight Sweep | Try one LoRA at a range of strengths. Downstream generation can run once per strength; connect its labels to Image Grid Compare to identify the results. |
| Trigger Word Extractor | Find a LoRA trigger word in a prompt and return the remaining text. |
| Krea 2 Guidance Helper | Check a guidance value against the node’s Turbo or Raw recommendation and see an approximate indication of its effect. |
| Krea 2 Variant Validator | Flag a likely mismatch between a Krea 2 Turbo or Raw checkpoint and its sampling settings. |
| VRAM / Model Size Estimator | Estimate memory needs from checkpoint and LoRA file headers. Its result is a lower-bound estimate, not a guarantee that a workflow will fit. |
| Smart Unloader | Release models and run cleanup to help free memory while passing its input onward. |
| Metadata Embed / Metadata Reader | Save custom metadata and read embedded metadata as text or parsed JSON. |
Specialized LoRA-family loaders have limits. The LoHa, LoKr, and DoRA loaders support 2D linear-style weights, but not 4D convolution layers. A file may therefore apply only partially. Their layer-name matching is also best-effort, particularly outside familiar SD1.5 and SDXL U-Net layouts. Review the reported skip counts. For more predictable DoRA results, apply its loader before other nodes that patch the same model layers. A related ComfyUI auto-detection issue is documented in ComfyUI issue #8683.
GGUF file tools
GGUF is a model-file format that can store quantized weights. OmniNodes can inspect, load, check, and create these files.
| Node | What you can use it for |
|---|---|
| GGUF File Info | Inspect a file’s model metadata and the actual quantization types used by its tensors. Start here with an unfamiliar file. |
| GGUF Diffusion Model Loader | Open a GGUF diffusion model and output a regular ComfyUI model. |
| GGUF CLIP Loader | Open one or two GGUF text encoders and output regular ComfyUI CLIP conditioning components. |
| GGUF VAE Loader | Open a GGUF VAE and output a regular ComfyUI VAE. |
| GGUF Checkpoint Converter | Convert a supported checkpoint to GGUF, choosing a listed Q4–Q8 type or F16/F32. |
| GGUF Quant Validator | Sample tensors to flag invalid, all-zero, or empty data before using a questionable download or conversion. |
Important memory distinction: OmniNodes’ GGUF loaders fully dequantize weights when loading them. The resulting model behaves like a normally loaded model, but it needs the corresponding full-precision memory rather than retaining the GGUF file’s reduced-memory advantage during inference. If saving VRAM is your reason for using GGUF, these loaders do not provide that benefit.
The converter retains tensors that cannot use its chosen block-quantization type as F16 and reports how many required that fallback. It does not offer K-quants.
Prompts and sampling
| Node | What you can use it for |
|---|---|
| CLIP Skip | Stop text encoding a chosen number of layers early, where a checkpoint or style calls for it. |
| Conditioning Composer | Combine, concatenate, or set an area for two conditionings. |
| CLIP Text Encode (Simple) | Encode prompt text with a concise summary output. |
| Embedding Helper | Format textual-inversion embedding tokens for a prompt. |
| Negative Prompt Presets | Choose a prepared negative-prompt block. |
| Prompt Cleaner | Remove duplicate tags, extra spaces, and malformed weighting syntax. |
| Prompt Combiner | Join prompt fragments with a selected separator. |
| Prompt Random Line | Select one line from a text block using a seed. |
| Prompt Token Counter | Check prompt length and warn about possible truncation. |
| Prompt Weight Scheduler | Produce a prompt fragment whose attention weight changes according to a schedule. |
| Wildcard List Inspector / Wildcard Loader / Wildcard Prompt Builder | Inspect wildcard text files, resolve __wildcard__ choices with a seed, and assemble prompts from multiple categories. |
| Empty Latent Image | Set up a blank latent with a chosen resolution and batch size. |
| Simple KSampler | Run ComfyUI sampling with an added summary of the settings. |
| KSampler Base+Refiner | Run an SDXL base stage and refiner stage with a handoff point controlled by switch_fraction. |
| KSampler Conditioning Blend | Mix two positive conditionings with a blend_ratio, then sample. |
| KSampler Masked Inpaint | Supply an image and mask directly for an inpainting sampling pass. |
| KSampler Seed Variator | Generate a batch using consecutive seeds; set num_variations to choose how many. |
| Latent Upscale (Hires Fix) | Upscale a latent and optionally resample it at low denoise. Turn resampling off to use it only as an upscaler. |
| Detailer (Crop-Inpaint-Paste) | Crop a masked detail, refine it at higher resolution, and blend it back into the original image. |
| Seed Stepper | Progress through seeds by incrementing, choosing unused random values, or cycling through a fixed list across queue runs. |
| VAE Encode / VAE Decode | Convert between images and latents, with an added summary output. |
For a quick comparison, try KSampler Seed Variator to explore compositions, or Multi-LoRA Weight Sweep when you want to judge how strongly a particular LoRA should affect an image. Change one factor at a time to make the comparison easier to interpret.
Video
ComfyUI commonly handles video frames as a batch of images. These nodes help you prepare, alter, and assemble those frames.
| Node | What you can use it for |
|---|---|
| MiniMax H3 Duration Calculator | Convert a requested duration into a compatible MiniMax H3 frame count at 24 fps and check the requested resolution. |
| Video Load | Load frames and report the source frame rate and count. |
| Video Trim / Extract | Keep a selected range of frames. |
| Video Concat / Splice | Join image batches into a longer sequence. |
| Video Frame Interpolate | Add frames between existing frames for smoother apparent motion. |
| Video Speed Ramp | Vary playback speed across the sequence. |
| Video Loop Composer | Build looping or back-and-forth sequences. |
| Video Motion Trail | Add a ghosted movement trail. |
| Video Scene Detect | Find hard cuts and report scene boundaries. |
| Video Color Match | Bring frames closer to a reference image’s color appearance. |
| Video Mux Audio | Add an audio track to a saved video file. |
| Video Save | Export an image batch as MP4, WEBM, or GIF, subject to the format’s required packages. |
Web services and workflow organization
| Node | What you can use it for |
|---|---|
| HTTP Request (WebAPI) | Send a web request and inspect its response, status code, and headers. |
| OAuth2 Token Manager (WebAPI) | Obtain an access token from a compatible service. |
| RSS Feed Parser (WebAPI) | Read feed entries, optionally filtering by a keyword. Some unusual or malformed feeds may not parse correctly. |
| Webhook Listener (WebAPI) | Receive posted data through a local server and return the oldest unread payload on a workflow run. |
| Model/LoRA Downloader | Download a file into a registered ComfyUI model folder, skipping it when the target file already exists. |
| JSON Field Extractor | Read one value from a web response using a path such as data.items.0.title; return a fallback if it is missing. |
| JSON Builder | Assemble a request body from up to four key/value pairs and optional additional JSON. |
| Endpoint Poller | Check an endpoint repeatedly until it reports the expected result or the time limit is reached. |
| Response Saver | Save a response as text or JSON without overwriting a previous numbered file. |
| Discord Notify | Send a completion message and, optionally, the first image in a batch to a Discord webhook. Keep its webhook URL private. |
| Folder Watcher | Find the next file not yet marked as processed in a folder. |
| Any Switch / Conditional Gate | Route a value or choose between branches according to a condition. |
| Batch Counter | Count queue runs and derive a seed or index. |
| Timer Start / Timer Stop | Measure how long a named part of a workflow takes. |
| Prompt List Iterator | Step through prompts in a text file or folder; pair its index with Batch Counter. |
| Try/Catch (Value Guard) | Replace a returned error-like or invalid value with a fallback. It cannot catch a node that crashes before returning a value. |
| Workflow End | Accept up to four inputs and produce a run summary. |
| Workflow Manifest Writer | Save a JSON record of settings associated with an output, such as its prompt, checkpoint, seed, sampler, and LoRAs. |
Web API tools may depend on the outside service, its permissions, and its response format. The nodes that make HTTP requests require the requests package. All GGUF tools require gguf. Video Save’s MP4/WEBM output requires imageio and imageio-ffmpeg; GIF requires imageio. These are optional dependencies for their respective tools, not prerequisites for every OmniNodes workflow.
What’s New with OmniNodes
The supplied history names two dated groups of changes but does not provide a complete version-by-version release history.
- August 8, 2026 — Recover settings from finished images. Load Image + Recovered Metadata added separately usable outputs for embedded generation details. This makes it easier to revisit a previous render or use a shared PNG as the starting point for a new experiment.
- September 14, 2026 — More direct image and model workflows. The dated changes include Background Remover, Image Resize, ControlNet Stack, Model/LoRA Downloader, and Latent Upscale (Hires Fix). They reduce the number of separate steps needed for common cutout, resizing, multi-ControlNet, model-file, and high-resolution refinement workflows.
- September 14, 2026 — A working webhook listener. Webhook Listener changed from an earlier nonfunctional stub to a local listener that receives posted payloads and checks for new data on subsequent runs.
- Other additions described in the supplied readme, without individual release dates. These include the six GGUF tools; specialized sampling and inpainting nodes; image prompting and ControlNet application; Krea 2 checks; MiniMax H3 helpers; and additional inspection, comparison, and record-keeping tools. The latent-mask syntax error described in the source was also fixed, but no separate version number is supplied for it.
Troubleshooting OmniNodes
A node does not appear in ComfyUI
- Restart ComfyUI completely.
- Search for the node’s exact name. Model tools can appear under
TensorVizion/Model UtilitiesorTensorVizion/Model, and sampling tools may appear underModel UtilitiesorSampling. - Check ComfyUI’s terminal output for
[OmniNodes]messages.Loadedindicates registration;No NODE_CLASS_MAPPINGSindicates a skipped file;Error importingindicates an import failure with details beneath it. - If an error names a missing package, check whether the node belongs to a category with an optional dependency.
A Web API or GGUF node reports a missing package
- Read the package name in the error. HTTP-dependent Web API nodes require
requests; GGUF nodes requiregguf. - Have the missing package made available in the Python environment that runs ComfyUI.
- Restart ComfyUI and try the node again. Other OmniNodes categories do not require those two packages merely to operate.
Video Save cannot produce MP4 or WEBM
- Check whether
imageioandimageio-ffmpegare available to ComfyUI. - Confirm that the incoming image batch contains the frames you intend to save.
- Retry the export after the required video packages are available. GIF uses
imageiowithout the additional MP4/WEBM package.
A loaded GGUF file uses more memory than expected
- Confirm that you are using an OmniNodes GGUF loader.
- Remember that these nodes dequantize the model at load time, so the smaller file does not remain a reduced-memory model during inference.
- Use a quantized-inference solution that keeps weights quantized if reduced VRAM use is essential.
A specialized LoRA-family loader has little or no visible effect
- Inspect the file with LyCORIS Format Inspector to identify its format.
- Use the matching LoHa, LoKr, or DoRA loader.
- Read the loader’s summary for skipped layers. Convolution layers are not supported, and some model architectures may have layer names that cannot be matched.
- For DoRA, try placing its loader before other model-patching nodes that affect the same layers.
Recovered image settings look incomplete or incorrect
- Check
metadata_found. Images without supported embedded generation data still load, but do not provide recovered settings. - Try the appropriate
sourcechoice when you know whether the PNG came from ComfyUI or an A1111/Forge-style workflow. - Verify recovered positive and negative prompts manually when the original ComfyUI workflow contained several prompt-encoding nodes. The parser uses a heuristic for those cases.
A ControlNet depth guide does not match the scene’s actual depth
depth_lite makes a quick estimate from visual cues; it is not a trained depth estimator. For work that depends on accurate spatial depth, use an appropriate depth-model preprocessor instead.
The Value Guard did not stop a workflow error
Try/Catch (Value Guard) can inspect only a value that an earlier node successfully returns. It cannot intercept a crash in that earlier node. Check the failing node’s error directly, or use a node that returns an error value the guard can evaluate.
Frequently asked questions
Do you need every optional package to use OmniNodes?
No. The readme identifies requests for HTTP-dependent Web API tools, gguf for GGUF tools, and the relevant imageio packages for Video Save. Most audio, image, latent, prompt, sampling, model, and workflow functions use components available with ComfyUI.
Is Audio Stem Splitter the same as separating vocals from instruments?
No. It splits frequency bands into bass, middle, and high outputs. Sounds can overlap across those bands.
What is the difference between Contact Sheet Maker and Image Grid Compare?
A contact sheet is an unlabeled overview of a batch. Image Grid Compare adds per-image labels so you can identify settings in a controlled comparison.
Will a passing Latent QC Gate guarantee a good image?
No. It checks for certain technical problems in a latent, not composition, style, or artistic quality.
Can Quick Save Image and Custom Folder Batch Saver be used interchangeably?
They save to different places. Quick Save Image uses ComfyUI’s managed output area for a single image; Custom Folder Batch Saver writes batches to a directory you choose.
Learn More about OmniNodes
- OmniNodes repository: The extension’s source and available project documentation.
- ComfyUI repository: Background on the application and its node-based workflow system.
- ComfyUI issue #8683: Technical context for the LyCORIS format-routing concern discussed above.
For practical experimentation, begin with a small, familiar workflow and add one OmniNodes tool at a time. A labeled Image Grid Compare after a seed or LoRA-strength sweep makes changes particularly easy to see.
