ComfyUI workflow models

ComfyUI workflow models

Drop ComfyUI API-format workflow exports here. comfyui_adapter.py registers each *.json as a remote model, so it shows up in Pallaidium as [Remote] <name>.

Easiest way: in Pallaidium's Remote Backend preferences, with the ComfyUI adapter selected, click Import Workflow to copy a .json in here (the running backend reloads automatically), or Open Folder to manage these files. You only need the manual steps below when crafting the export.

1. Export a workflow from ComfyUI

1. In ComfyUI: Settings → enable "Dev mode" (adds API-export buttons).
2. Build/open the workflow you want (e.g. your SCAIL-2 character replacement, Bernini-R video edit, Stable Audio, or Boogu edit graph).
3. Save (API Format) → save the .json into this folder. The file name (without .json) becomes the model id and default label, e.g. bernini-r.json → [Remote] bernini-r.

The adapter also accepts a normal UI export that wraps the graph under a "prompt" key, but API format is preferred.

2. Mark the nodes the adapter should fill in

The adapter injects the request's parameters by node title (in ComfyUI: right-click a node → Title). Rename nodes to:

Title	Node it goes on	Filled with
prompt (or positive)	the positive CLIPTextEncode / text node	the prompt
negative	the negative text node	the negative prompt
width / height	EmptyLatent / size node	resolution
steps	sampler	inference steps
cfg	sampler	guidance scale

Also handled automatically (no title needed):

• Seed — any node with a seed or noise_seed input is set to the request seed.
• Reference image(s) — every LoadImage node receives an uploaded image, in title order (image, image2, …). One LoadImage → img2img / i2v; several → multi-reference (e.g. a character reference). Leave the LoadImage's own default in place for txt2-only runs (it's only overwritten when an image is sent).
• Source video — for video edit / character replacement, a LoadVideo node (VHS_LoadVideo / VHS_LoadVideoPath / VHS_LoadVideoUpload / VHS_LoadVideoFFmpeg / core LoadVideo) receives the source video you select in the VSE. The adapter advertises such a workflow with the control mode, so Pallaidium shows the video input picker; the chosen strip is uploaded and bound to the first video loader. Character replacement = LoadVideo (source) + LoadImage (the new character).

Anything you don't title is left exactly as saved, so the workflow runs as designed.

3. Media type is auto-detected

From the workflow's output node:

• SaveImage / PreviewImage → image
• VHS_VideoCombine / SaveAnimatedWEBP / SaveWEBM / SaveVideo → video (use VideoHelperSuite's VHS_VideoCombine with format video/h264-mp4 for mp4)
• SaveAudio / SaveAudioMP3 / SaveAudioOpus → audio

i2i/i2v is enabled when the graph has a LoadImage node.

4. Optional overrides — <id>.meta.json

Put a sidecar next to the workflow to override the label or detected values, e.g. bernini-r.meta.json:

{
  "display_name": "Bernini-R: Video Edit",
  "type": "video",
  "modes": ["i2v"],
  "max_ref_images": 1,
  "default_steps": 30
}

Use this to set the pretty names from your screenshots (the file name can't contain :), or to correct anything the auto-detector got wrong.

Explicit node bindings (for complex graphs)

When a workflow's parameters don't live on conventionally-titled nodes — e.g. the prompt sits on a PrimitiveStringMultiline, width/height are PrimitiveInt nodes using a value input, an LLM prompt-enhance chain feeds the encoder, or there are several identically-titled CLIPTextEncode nodes — titling can't disambiguate them. Add a bindings map to the sidecar that names the exact {node, input} each request field maps to (node id = the JSON key, e.g. "320:319"; input = the key inside that node's inputs):

{
  "display_name": "LTX-Video 2.3 i2v",
  "type": "video",
  "modes": ["i2v"],
  "max_ref_images": 1,
  "bindings": {
    "prompt":   {"node": "320:319", "input": "value"},
    "negative": {"node": "320:313", "input": "text"},
    "width":    {"node": "320:312", "input": "value"},
    "height":   {"node": "320:299", "input": "value"},
    "fps":      {"node": "320:300", "input": "value"}
  }
}

• Bindable fields: prompt, negative, width, height, steps, cfg, fps, num_frames, strength, seed. A field you omit (or that the request didn't supply) leaves that node at its saved value.
• A field may map to several nodes — use a list: "seed": [{"node": "a", "input": "noise_seed"}, {"node": "b", "input": "noise_seed"}].
• Bindings override title-matching, and the automatic handling still applies: the source image is uploaded and bound to LoadImage, the source video to a LoadVideo, and any seed/noise_seed input is set even without an explicit binding.
• Find node ids by opening the API-format JSON — each top-level key ("320:319", "269", …) is a node id; its class_type and _meta.title tell you which is which.

A complete worked example ships in this folder:

• ltx-2.3-i2v.json
• ltx-2.3-i2v.meta.json.

5. Use it

Restart the adapter (it scans this folder at startup), then in Pallaidium click Refresh Remote Models. Your workflows appear in the matching dropdown (Movie / Image / Audio).