ComfyUI-Impact-Pack: Detectors, Detailers and Iterative Upscalers for ComfyUI

What ComfyUI-Impact-Pack actually adds to a ComfyUI graph

ComfyUI ships as a node graph editor for diffusion pipelines, and everything outside its built-in node set arrives as a custom node pack. ComfyUI-Impact-Pack is one of those packs. Its README describes the goal plainly: it helps "conveniently enhance images through Detector, Detailer, Upscaler, Pipe, and more." The audience is people who already build ComfyUI workflows and want detection and refinement stages inside the same graph rather than in a separate editing tool.

The pack's own project metadata goes further. The pyproject.toml description says it offers "various detector nodes and detailer nodes that allow you to configure a workflow that automatically enhances facial details" and provides an "iterative upscaler." That is the concrete promise: find a region, refine it at higher fidelity, then upscale the result, all as connected nodes.

The nodes are grouped in the README under Detector nodes. SAMLoader (Impact) loads a SAM model. ONNXDetectorProvider loads an ONNX model to provide BBOX_DETECTOR. CLIPSegDetectorProvider wraps CLIPSeg for the same output type, and the README notes you need the ComfyUI-CLIPSeg extension installed for it. SEGM Detector (combined) and BBOX Detector (combined) return a mask from an input image. SAMDetector (combined) takes SEGS plus an image and outputs a unified mask. SAMDetector (Segmented) splits the detected segments instead, emitting combined_mask and batch_masks, with the README admitting the grouping policy is arbitrary sets of three and "expected to be improved in the future." That last note is worth reading as an honest limitation rather than a feature.

The repository layout backs this up. modules/ holds the implementation, js/ the front-end extensions, locales/ the translations, wildcards/ and custom_wildcards/ the wildcard text, example_workflows/ a set of ready graphs, and tests/ plus troubleshooting/ the checks and failure notes. node_list.json enumerates the nodes the pack registers. install.py sits at the root, which is the script the README tells you to run by hand.

How SEGS, detectors and detailers move data through the graph

The pack revolves around SEGS, a structured list of detected regions that carries bounding boxes and masks between nodes. A detector node produces SEGS, a detailer node consumes them, and the refined regions are recombined into the output image. Simple Detector (SEGS) is the convenience layer: it operates "primarily with BBOX_DETECTOR" and, when SAM_MODEL or SEGM_DETECTOR is also supplied, generates improved SEGS through mask operations on both the bounding box and the silhouette. The README calls it "a convenient tool to simplify a somewhat intricate workflow," which is a fair description of what it replaces.

Video paths exist too. Simple Detector for Video (SEGS) runs detection on each frame individually and builds a SEGS object holding a batch of masks rather than one mask. SAM2 Video Detector (SEGS) does something different: it uses SAM2 video tracking to produce that batched SEGS. The distinction matters. Per-frame detection can flicker between frames because each frame is decided independently; tracking carries identity forward. The README states support for facebookresearch/sam2 models arrived in V8.18, and sam2 is listed as a dependency in both pyproject.toml and requirements.txt, pulled from the Git repository rather than a package index.

Pipe nodes are the plumbing that keeps these graphs manageable. The README warns that when wildcard support was added to FaceDetailer, "the structure of DETAILER_PIPE-related nodes and Detailer nodes has changed," and that existing workflows may malfunction. That is the trade-off of a pack this size: the node graph is the API, and changing node structure breaks saved graphs.

Wildcards are a second mechanism worth understanding before you build around them. The README notes the selection weight syntax changed from a single colon to a double colon at V3.16, and points to a separate tutorial file in the ComfyUI-extension-tutorials repository. If you copy wildcard prompts from an older workflow, the syntax may no longer parse the way it did.

Installing ComfyUI-Impact-Pack manually

The README marks ComfyUI-Manager as the recommended install route. Manual installation is documented and is the path to follow if you do not use the manager. Note the NOTICE entry for V7.6: automatic installation is no longer supported, and the README says to install using ComfyUI-Manager, or manually install requirements.txt and run install.py.

Start by cloning into the custom_nodes directory. The README gives the clone target name as comfyui-impact-pack, and the second command changes into that directory.

cd ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Impact-Pack comfyui-impact-pack
cd comfyui-impact-pack

Dependencies differ by environment. On Windows Portable, the README runs pip through the bundled interpreter, using a relative path back up to python_embeded.

..\..\..\python_embeded\python.exe -m pip install -r requirements.txt

For venv or conda, activate the environment first and then install from the same requirements file.

pip install -r requirements.txt

The README also states that install.py must be run to complete installation. After restarting ComfyUI, the nodes should appear in the node menu. If you need UltralyticsDetectorProvider, the README is explicit that it is not part of this pack and requires ComfyUI-Impact-Subpack installed separately. Since V8.0 the subpack is no longer installed automatically.

Version coupling is the real operational cost

The NOTICE list is the most useful part of the README for anyone deciding whether to adopt this pack, because it documents how often the pack has required a matching ComfyUI version. V8.24 requires ComfyUI 0.3.63 or higher due to structural changes in DifferentialDiffusion. V5.0 dropped compatibility with ComfyUI versions before 2024.04.08. V4.77 required an October 8th build or later. V3.6 required a specific August 8 commit. Two entries describe outright failures: loading the pack on ComfyUI versions released before June 27, 2023 fails, and PreviewBridge may not function correctly on versions before July 1, 2023.

Some breaks are structural rather than version-gated. V8.19 removed legacy nodes including mmdet. V4.12 changed MASKS to MASK. V4.20.1 changed the parameter order in RegionalSampler, which the README says causes malfunctions in previously created RegionalSamplers. V4.85 is incompatible with an outdated ComfyUI IPAdapter Plus, requiring a build dated March 24th or later. V4.7.2 is incompatible with old versions of ControlNet Auxiliary Preprocessor when using MediaPipe FaceMesh to SEGS.

The practical consequence: upgrading ComfyUI and upgrading this pack are not independent decisions. If you pin ComfyUI for stability, you also pin the pack, and you inherit whatever detector models and node shapes that pair supports. The repository carries tests/, troubleshooting/ and docs/ directories, but the README does not document a rollback procedure for a workflow that broke after an upgrade. That silence is itself a risk to weigh.

Where ComfyUI-Impact-Pack is the wrong tool

This pack is not a standalone application. It is a set of custom nodes, so it only does anything inside a running ComfyUI installation. If your goal is to upscale a folder of images from a shell script, this is not the route; you would be installing a graph editor to reach an upscaler.

It is also not self-contained for detection. UltralyticsDetectorProvider lives in ComfyUI-Impact-Subpack, and since V8.0 that subpack is not installed automatically. The README states this twice, once at the top and once in the Companion Pack section. A user who installs only this repository and then looks for YOLO detection models will not find them.

There are hard environment floors. The pack is incompatible with ComfyUI versions before 2024.04.08 as of V5.0, and V8.24 raises the floor to 0.3.63. If you are running an older ComfyUI because another custom node depends on it, this pack may be the node that forces the upgrade, and the upgrade may break the other node. The README's own warning about ComfyUI IPAdapter Plus is exactly this scenario.

Finally, the segmentation output is not exact. SAMDetector (Segmented) groups detected segments "arbitrarily in sets of three," by the README's own wording. If your workflow needs clean per-object masks, the documentation does not promise that today. And the pack's dependency list is not small: segment-anything, scikit-image, piexif, transformers, opencv-python-headless, scipy, numpy, dill, matplotlib and sam2 all land in the same Python environment as ComfyUI, so a conflict with another custom node's pins is a real possibility.

How it compares with running detection outside ComfyUI

The alternative most people already have is a two-stage pipeline: run a detector such as a YOLO model or SAM in a separate Python script, save masks or cropped regions to disk, then feed the crops into ComfyUI as ordinary images and composite the results afterward. That approach is more work to wire up, but each stage is replaceable without touching the other, and a broken detector does not invalidate a saved workflow.

The difference in approach is where the state lives. In the external pipeline, the interface between detection and generation is files on disk, which is stable and inspectable. In ComfyUI-Impact-Pack, the interface is SEGS flowing between nodes in memory, which is faster to iterate on and keeps everything in one graph, but which the README shows is subject to structural change across releases. DETAILER_PIPE nodes changed shape when wildcard support reached FaceDetailer. MASKS became MASK at V4.12. Legacy mmdet nodes were removed at V8.19.

A second alternative is to stay inside ComfyUI but use only the built-in nodes and a lighter custom node set. You lose SEGS-based region handling, the combined detector nodes and the iterative upscaler, and you gain a much smaller compatibility surface. Whether that trade is worth it depends on how much of your workflow is region refinement versus whole-image generation.

Licence and maintenance status

ComfyUI-Impact-Pack is licensed under GPL-3.0, with LICENSE.txt at the repository root and pyproject.toml declaring the licence by file reference. GPL-3.0 is a copyleft licence. If you redistribute a modified version of this pack, or distribute a larger work that incorporates it, the licence terms apply to that distribution. Running it locally to produce images is a different question from shipping it inside a product, and the repository does not offer guidance on that boundary. This is not legal advice; read LICENSE.txt and consult someone qualified if you plan to redistribute.

The repository is not archived. Its last push was on 2026-04-19, so it is not accurate to describe the project as actively maintained on the basis of push activity alone. The version in pyproject.toml is 8.28.3, and the README's NOTICE list documents a long release history with frequent compatibility patches. The project also links a YouTube tutorial playlist and a separate ComfyUI-extension-tutorials repository for node documentation, so the README is a changelog and index rather than a full manual.

Upgrade cost follows from the same NOTICE list. There is no documented migration tool, no rollback procedure, and no compatibility shim for the node renames the README describes. A team running this in production should keep a copy of the working ComfyUI and pack versions together, because the pack's own history shows that a version pair is the unit that works.

Editorial conclusion

Adopt ComfyUI-Impact-Pack if you already run ComfyUI and want detection, face detailing and iterative upscaling inside the same graph, and if you accept pinning ComfyUI to a version the pack supports. Do not adopt it if you want a standalone image tool, if you need UltralyticsDetectorProvider without installing ComfyUI-Impact-Subpack, or if you cannot absorb the migration cost that past releases imposed on existing workflows. Before committing, check the NOTICE section against your ComfyUI build, install requirements.txt and run install.py as the README states, and open one of the graphs in example_workflows/ to confirm the nodes resolve on your machine.

Frequently asked questions

Official sources

1. Issues
2. License: GPL-3.0
3. ltdrdata/ComfyUI-Impact-Pack on GitHub
4. README