Castcut nodes for ComfyUI¶

An optional ComfyUI node pack that runs some of Castcut's checks inside the render job instead of as separate ComfyUI calls. Everything works without it.

With the pack	Without it
Best of two for hard poses (Day) renders both takes in one job: one model load, one encode, the pose check runs inside ComfyUI and the closer take is saved as the still, the other beside it.	Best of two queues the second take after the first and checks each one with a separate ComfyUI call.
Cut-outs (Isolate on white, plates) repair the matte and composite onto the fill in the same job.	The matte comes back to the app, is repaired there and composited in a second step.

The pack is one Python file, castcut_nodes.py, using only what ComfyUI already has (numpy, Pillow). It adds five nodes under the Castcut category: Pose Score, Pick Best, Face Distance, Mask Repair and Report.

Install from Castcut (easiest)¶

Open Settings → ComfyUI → Castcut nodes. The card says whether this ComfyUI has the pack and which version (Missing, Installed v1.4.0, Outdated), and offers one way to install it:

ComfyUI-Manager is installed → Install with ComfyUI-Manager. When it is done, Restart ComfyUI… restarts ComfyUI through the Manager. It asks first, because a restart stops whatever ComfyUI is rendering and empties its queue for everyone, and it stays disabled (Wait for 2 jobs) while ComfyUI has jobs running or waiting. The card then waits for ComfyUI to come back and checks again.
No Manager, or the Manager refuses → a copy-paste command for the ComfyUI machine. The app reads ComfyUI's /system_stats (its start command, OS, Python) to find the custom_nodes folder and serves the file itself at /api/castcut-nodes/file. There are versions for curl and wget, Windows PowerShell (incl. the portable build), git, comfy-cli and Docker. Then restart ComfyUI and click Check again.

The Best of two switch on Day and the Play checks list in Settings link to the card when the pack is missing.

When ComfyUI-Manager says no¶

The pack is on the Comfy Registry as castcut-nodes (publisher doodersrage), so a Manager that reads the Registry installs it by name. A Manager that installs it from its Git URL instead (https://github.com/doodersrage/castcut-nodes) — recent versions, 3.41 and later — only does so when both of these are true:

allow_git_url_install = true under [default] in the Manager's config.ini — on current ComfyUI that is ComfyUI/user/__manager/config.ini (older ComfyUI: ComfyUI/user/default/ComfyUI-Manager/config.ini). It is read at start-up, so restart ComfyUI after changing it.
ComfyUI listens on this machine only (--listen 127.0.0.1, or no --listen). With --listen 0.0.0.0 Git-URL installs are always refused.

Restarting from the app needs the Manager's security_level at normal (the default) or lower. If you would rather not change Manager settings, use the copy command: it needs nothing from the Manager.

Install by hand¶

Any one of these, then restart ComfyUI.

Copy the file (smallest). From the Castcut checkout:

# A ComfyUI system service (here /opt/comfyui):
sudo install -m 644 comfyui-nodes/castcut/castcut_nodes.py /opt/comfyui/custom_nodes/castcut_nodes.py
sudo systemctl restart comfyui

# A ComfyUI in your home folder:
cp comfyui-nodes/castcut/castcut_nodes.py ~/ComfyUI/custom_nodes/

Or fetch it from a running Castcut (curl -fsSL http://<castcut>/api/castcut-nodes/file -o …); when the app has sign-in or PROMPT_API_TOKEN on, add -H "Authorization: Bearer <API key>" (a personal key from Profile → API keys works).

Windows portable: put castcut_nodes.py in ComfyUI_windows_portable\ComfyUI\custom_nodes\. No python_embeded packages are needed.

ComfyUI-Manager → Install via Git URL: https://github.com/doodersrage/castcut-nodes (see the settings above). It clones the whole Castcut repository; its root __init__.py loads only the node pack.

Comfy Registry: ComfyUI-Manager → search Castcut nodes, or comfy node install castcut-nodes (published as version 1.1.0; 1.2.0 adds the routes below).

Docker: copy the file into the container's custom_nodes (or onto the volume mounted there) and restart the container.

Check it worked: http://<comfyui>:8188/object_info/CastcutPoseScore returns the node, and the card shows Installed v1.4.0. With 1.2.0, http://<comfyui>:8188/castcut/info answers too.

Routes (1.2.0 – 1.5.0)¶

Besides the nodes, the pack adds a few HTTP routes to ComfyUI, so some checks no longer go through the render queue:

Face checks answer right away. The face match (identity score, partner and face-lock checks) and finding the face on a plate used to be queued as small graphs, which waited for the render in progress — 45–95 s on Edit 2511 for about half a second of work. They now run on the CPU inside ComfyUI, with the InsightFace model kept loaded, and give the same numbers (checked against the queued graphs on 12 stills: distances within 7×10⁻⁸, identical boxes).
No round trip for a still that is checked. A finished still used to be downloaded and uploaded again as an input; ComfyUI now copies it itself, under the same name.
object_info only when something changed. The app asks for a fingerprint of the node list and model files and refetches the ~8 MB object_info only when it moved.
1.3.0: pose checks and the Face finish probe too. The pose read behind Auto-review and the pose checks runs DWPose on the CPU inside ComfyUI (~0.3 s a still, the same keypoints as the node), and Face finish's "which face is hers" probe runs with the face checks.
1.3.0: delete unused uploads. Settings → ComfyUI → Input folder offers Delete N files after a scan. It asks first, scans again, deletes only what that scan offers, and the pack keeps anything younger than a day or named by a queued job. Without 1.3.0 you get the command as before.
1.3.0: health. One call for the queue, free GPU memory and what the pack can check; the Castcut nodes card uses it instead of reading the whole queue.
1.4.0: two-person pose reads too. The read that keeps a couple in contact as two bodies (person masks, each person alone, DWPose) runs in-process with the Impact Pack's own nodes and YOLO on the CPU, instead of a queued graph.
1.4.0: graphs without the picture. Fix area, Face finish and the LoRA check read the graph a still was made with from its PNG; the pack sends only that text, not the ~1.5 MB file.
1.4.0: is it being used? The Castcut nodes card shows what the pack answered since ComfyUI started (per check, with the average time) and how often the app had to fall back to the queue — run a Day and look there.
1.5.0: Open in ComfyUI opens the graph. The still's graph is staged in the pack's example_workflows/ folder, which ComfyUI serves as templates, and the link opens it straight onto the canvas (?template=…&source=…) instead of leaving you to find it in the Workflows sidebar. Needs the pack installed as a folder (Comfy Registry, git or ComfyUI-Manager — not the single castcut_nodes.py file) and one ComfyUI restart after installing, so ComfyUI registers the folder. The copy under workflows/Castcut/ is still saved either way.

Without 1.2.0 (or when a route fails) everything works as before, through queued graphs. The queue position of a waiting still now comes from ComfyUI's jobs list rather than the whole queue, on any ComfyUI that has it (pack or not).

Updating¶

The card compares the installed version (each node's description ends with [castcut-nodes 1.4.0]) with the version this app ships. Outdated means a newer one came with the app: install again the same way — copying the file over the old one, git pull in a clone, or the Manager. A copy from before version 1.1.0 shows as Installed (an older version).

Use one install, not two: a castcut_nodes.py file and a castcut folder in custom_nodes both load and the later one wins.

For maintainers¶

The pack's source, tests and Comfy Registry metadata (pyproject.toml) live in comfyui-nodes/castcut. scripts/castcut-nodes-release.sh <dir> copies it into the standalone repository (doodersrage/castcut-nodes) that is published to the Registry (comfy node publish). When a node changes, bump CASTCUT_VERSION in castcut_nodes.py, version in pyproject.toml and CASTCUT_NODES_BUNDLED_VERSION in src/lib/castcut-nodes-setup.ts together — the Python and TypeScript tests check they agree.