ComfyUI-BCNodes Introduction
ComfyUI-BCNodes is a collection of utility nodes and interface tools for ComfyUI. It helps you handle common tasks around an image-generation workflow: preparing masks, resizing images, organizing prompts, loading LoRAs, creating captions, comparing results, and saving finished work.
You can use one tool or combine several. For example, you might remove an image’s background, soften the resulting mask, apply a film-inspired look, compare the edited image with the original, and save it with a descriptive filename. Other nodes help with more specialized work, including training-caption preparation, SeedVR2 video processing, and checking how a post might be treated on social platforms.
Type BCNodes in ComfyUI’s node library to find the nodes. They are also organized into categories such as BCNodes/mask, BCNodes/image, and BCNodes/workflow. Each node has its own BC_ registration key so the pack can coexist with other custom-node packages.
How ComfyUI-BCNodes Works
A ComfyUI workflow passes information from one node to another. BCNodes provides small tools for particular steps in that process:
- Image nodes work with pictures or batches of pictures. A batch is simply a group of images processed together.
- Mask nodes work with selections. Think of a mask as a stencil: it identifies where an edit should take effect.
- Logic and workflow nodes decide which connected input to use or whether part of a workflow should run.
- Language-model nodes turn text and, for supported models, images into captions, prompts, or observations.
- Output nodes prepare and save the results.
For instance, you could connect a mask to MaskGrow to expand a selection and soften its edge. You could then connect that mask and an image to Draw Mask On Image to paint a color only where the mask allows it. You control the steps by connecting nodes and adjusting their settings.
Some BCNodes tools are not ordinary nodes. Align adds buttons to the toolbox when you select multiple items, while Process Monitor shows memory use and information about workflow runs.
ComfyUI-BCNodes Features
Masks and Backgrounds
- Is Mask Empty checks whether a mask contains a selection. Its true/false result can help a workflow decide what to do next.
- Mask Fill Holes fills enclosed gaps inside a mask.
- MaskGrow expands or shrinks a mask and then blurs it. Expanding can cover more of a subject; shrinking can pull the selection away from an edge.
- Draw Mask On Image paints a chosen color through a mask onto an image batch. Opacity controls how strongly that color appears.
- Blockify Mask turns a mask into blocks within its bounding box, retaining blocks that contain part of the selection.
- Repeat Mask Batch repeats a mask batch the specified number of times.
- BiRefNet Remove Background removes an image background using BiRefNet.
Image Preparation and Saving
- Image Scale By Aspect Ratio scales an image or mask using a target aspect ratio and side length. Its mask output is always present rather than
None. - Image Resize resizes images and masks to a specified width and height. Depending on the option you choose, it can stretch the result, preserve proportions, add padding, or crop. Use padding when you want to retain the whole image; use cropping when filling the target frame matters more.
- Join Image Lists combines image lists and supports additional inputs.
- Depth Anything produces a depth map, with nearer areas shown in white. You can choose a ControlNet-preprocessor size or an exact width and height.
- Social Media Export creates a platform-ready version for each selected platform, using a minimal crop according to its specifications.
- Save Image builds folder and file names from prompt-widget values, saves in a Pillow-supported format, and embeds prompt and workflow information. Its previews appear in the gallery rather than beneath the node.
- Save Image With Caption saves PNG images and, when a caption is connected, a matching caption file beside each image. This is useful when preparing a training dataset.
- Skin Texture adds fine skin detail inside a SAM 3 mask by enhancing detail already in the image and adding a synthetic pore pattern.
- Frequency Merge combines the broad structure and color of one image with the fine texture of another image of the same size.
Looks and Quality Checks
The PostFx nodes help you build and apply film-emulation looks:
- PostFx Apply applies a look to an image batch. You can select a theme and condition, adjust strength, and optionally provide a different look or a mask.
- PostFx Theme starts a look from a named, built-in theme.
- PostFx Custom Look builds a look with common controls or adjusts an incoming look.
- PostFx LUT uses a 3D
.cubeLUT on its own or layers it over a theme. - PostFx Signature Sheet creates a labeled contact sheet showing every theme in a category on one image. It can help you compare looks before choosing one.
The analysis nodes provide checks rather than edits:
- Image Quality Gate evaluates blur, sharpness, noise, clipping, and entropy, then displays scores and a
PASS,SO-SO, orFAILverdict. - Caption Audit checks a folder of LoRA training captions with
caption-auditand displays a report card on the node.
Prompts, Logic, and Workflow Control
- Prompt List turns one prompt per line into a list.
- Aspect Prompt List groups prompts beneath
[WxH]headings and outputs corresponding prompt, width, and height lists. - Show Text displays incoming text on the node while passing it onward.
- Logic Boolean converts a value between 0 and 1 into
BOOLEAN,NUMBER,INT, orFLOAToutputs. - Math Expression calculates an arithmetic expression using inputs
a,b, andc. - Any Switch takes the first connected input that is not
None. It accepts different input types and additional inputs. - Select Switch chooses a named input. Only the selected branch runs, which can be useful when switching between alternative approaches.
- Seed provides a seed setting. Set it to
-1to draw a new random seed on every run. - Auto Bypass automatically bypasses target nodes when a watched source is empty.
- Anything Everywhere supplies unconnected inputs of a matching type when the workflow runs.
- Fast Groups Bypasser provides a bypass toggle for each group.
- Image Comparer shows two images on a node with a sliding divider, making visual differences easier to inspect.
- Align adds align and distribute buttons to the toolbox for a selection of items. It is an interface tool, not a node.
Language Models, Captions, and Prompts
- Qwen LM runs a supported Qwen chat model through ComfyUI core. You can provide system and user text, a prefill, images, and one LoRA; it outputs
textandthinking. It uses the models’ official chat templates and sampling defaults. - LM Config adjusts language-model sampling settings. Only settings you change or connect replace the model’s official defaults.
- LM Loader selects the model used by Captioner, Prompter, or PreFlight Observe. You can choose the family, model, precision, one LoRA and its strength, and whether to keep the model loaded. The loader selects settings for another node; it does not load a model by itself.
- Captioner creates one training caption per image frame. Choose the built-in
tagsorproseprompt, ornonewhen using your own image-to-caption LoRA. You can also provide an instruction or LM Config. It outputs frames and captions as lists for Save Image With Caption. - Prompter creates one text-to-image prompt from a short brief, a reference image, or both. Its built-in choices can describe a person’s appearance or leave identity to a character LoRA;
noneis available for your own prompter LoRA. It outputs text and, when CLIP is connected, positive conditioning encoded from that text.
Pre-Publication Checks
The PreFlight nodes form a sequence for assessing an image or video frame before publication:
- PreFlight Observe uses the selected chat model to describe visible details such as garments, exposure, framing, pose, and text in the image. It returns structured observations as JSON. Its greedy generation setting is intended to produce the same result for the same input; errors are returned as
{"error": ...}. - PreFlight Report combines those observations with a caption. Its rules engine returns verdict ranges for Instagram, TikTok, and X:
OK,RISKfor possible reach demotion, andBLOCKfor possible removal. It also provides a summary and logs its prediction. - PreFlight Outcome records what actually happened after a post was published: clean, demoted, or removed.
- PreFlight Calibrate displays feedback as per-platform and per-rule calibration tables. It runs again when the feedback store changes.
Treat these verdicts as predictions, not guarantees about a platform’s decision. Recording real outcomes gives you a way to compare predictions with experience.
LoRAs, Models, and Performance
- Power Lora Loader applies any number of LoRA rows to a
MODELinput and outputs aMODEL. It does not take CLIP. Each row uses the key handling provided by Lora Loader (Key Fix). - Lora Loader (Key Fix) applies one LoRA after renaming certain keys that ComfyUI’s own loader may otherwise leave out, including lightx2v
.diff_mand PEFT keys without thediffusion_model.prefix. The console reports how many keys were renamed and how many matched no module. - Auto Model Downloader lists models used by a workflow and fetches missing ones into
models/. - Patch Sage Attention changes attention for one connected model rather than enabling SageAttention globally.
autoselects the most accurate kernel;qk_quantcontrols Q·K quantization. If a call is unsupported, the kernel errors, or its output contains a NaN, the node uses the original attention. The console counts these cases per run. See the model documentation for its requirements and an A/B test.
SeedVR2 Video Processing
The SeedVR2 nodes cover several stages of a SeedVR2 workflow:
- SeedVR2 Framing Downscale selects a downscale factor from the tallest SAM 3 face box, distinguishing close-up, medium, and far framing. It also outputs the measured face fraction.
- SeedVR2 Resize prepares original frames for encoding, including resizing and padding, and outputs a color reference.
- SeedVR2 VAE Encode streams frames from RAM in slices so VRAM use does not grow with frame count.
- SeedVR2 Chunk Size estimates the longest compatible
4n+1chunk from latent size and the graphics card’s total memory. Connect its result to Split SeedVR2 Latent’s manualframes_per_chunksetting. - SeedVR2 VAE Decode streams decoded slices to RAM so VRAM use does not grow with frame count.
- SeedVR2 PostProcess processes output one frame at a time into a float16 result without full-size temporary copies.
- SeedVR2 Preprocess (Compact) combines resizing and VAE encoding, outputting a latent and the plan needed by its matching compact post-process node.
- SeedVR2 PostProcess (Compact) decodes only retained frames and corrects them using a reference rebuilt from the original frames for each chunk.
Process Monitor
Process Monitor is an interface tool rather than a node. It shows live RAM and VRAM use, an estimate before a run, and measurements for individual nodes. It can report why a run was killed and perform a full memory clear back to startup levels without restarting ComfyUI. Its setting to drop stale outputs is off by default.
ComfyUI-BCNodes Models
BCNodes is primarily a utility pack, not a collection of interchangeable image-generation checkpoints. These are the model choices identified in the provided documentation:
| Model or family | Used for | When you might choose it |
|---|---|---|
| Qwen3.5-4B and Qwen3.5-9B | Qwen LM and LM Loader workflows | Chat-model tasks such as generating text, captions, or prompts, according to the connected node. |
| Qwen3.8-27B | Qwen LM workflows | A listed Qwen chat-model option when your setup can accommodate it. |
| Qwen3-VL-8B | Qwen LM workflows | Tasks that use an image input alongside language-model instructions. |
| Abliterated Qwen3.5-9B and Qwen3-VL-8B builds | Qwen LM workflows | Alternative listed builds of those models. |
| Qwen3.5-9B INT8 ConvRot | PreFlight Observe | The node’s default observation model. |
| BiRefNet | BiRefNet Remove Background | Separating a subject from its background. |
| Depth Anything V2 Small | Depth Anything | Producing a depth map from an image. |
| Depth Anything 3 Small, Base, Mono-Large, and Metric-Large | Depth Anything | Alternative listed depth-map model choices. |
The Qwen LM model file can be found in models/Qwen-LM or models/text_encoders, or downloaded on first use. LM Loader narrows its available choices according to the selected model family and model. The LM documentation covers model files, LoRAs, and sampling in more detail.
Troubleshooting ComfyUI-BCNodes
A node is missing from the library
- Search the node library for
BCNodes, then check the relevant category. - If the missing node is a PostFx or Caption Audit node, check that its supporting package,
postfxorcaption-audit, is available. The rest of BCNodes can still load without those packages. - Remember that Align and Process Monitor are interface tools, not nodes. Align appears in the toolbox above a multi-selection.
Patch Sage Attention does not appear to change attention
- Check the Patch Sage Attention documentation for its SageAttention 2.2.x requirement; that package is not included in BCNodes’
requirements.txt. - Review the console counts. The node uses the model’s original attention when a call cannot use the selected kernel, a kernel fails, or the result contains a NaN.
- Use the documented A/B test to compare the patched and original behavior for your workflow.
A LoRA seems to apply only partly
- Try Lora Loader (Key Fix) for one LoRA, or Power Lora Loader for multiple LoRAs.
- Check the console for the number of renamed keys and keys that matched no module.
- See the loaders documentation for the specific key formats these nodes address.
PreFlight Observe returns an error
- Inspect the returned JSON for an
"error"entry. - Check the model selected in the connected LM Loader.
- Refer to the PreFlight documentation for the expected observation-to-report workflow.
Frequently Asked Questions
Do I need every node in a workflow?
No. The pack contains independent utilities as well as nodes designed to work together. Use only the steps your project needs.
Does PreFlight tell me exactly what a social platform will do?
No. Its OK, RISK, and BLOCK ranges are predictions. PreFlight Outcome lets you record what actually happened.
Can I keep the entire image when resizing to a new shape?
Image Resize offers a padding option for preserving the image’s proportions within the target dimensions. Cropping fills the frame by removing some content; stretching changes proportions.
Why is my saved image preview not beneath Save Image?
That node shows its preview in the gallery.
Where can I compare two results?
Use Image Comparer to inspect two images with a sliding divider.
Learn More about ComfyUI-BCNodes
Start with the guide for the task you are working on:
- Mask tools and image tools for selections, resizing, depth, exports, and saving.
- PostFx looks and analysis tools for styling and checking images.
- Text tools, logic tools, and workflow tools for organizing prompts and controlling a graph.
- LM tools and PreFlight tools for captions, generated prompts, observations, and publication-risk predictions.
- SeedVR2 tools for video-processing stages and memory-conscious options.
- Process Monitor for memory and run diagnostics.
- Measurements for recorded time, RAM, and output-size comparisons; unused outputs for the behavior of unconnected whole-batch outputs.
BCNodes is released under the MIT license. Bundled BiRefNet architecture code retains its MIT notice, and bundled Depth Anything V2 architecture code retains its Apache-2.0 license.
