Import a ComfyUI workflow as an app
tools/import_workflow.py turns a ComfyUI workflow into a draft Freefield app: apps/<id>/app.json and
apps/<id>/workflow.api.json. The draft passes npm run validate; you check what it guessed, then mark it ready.
python tools/import_workflow.py my_workflow.json --id my-app --name "My App"
python tools/import_workflow.py ui_workflow.json --id my-app --object-info object_info.json # a UI workflow
The file can be in API format (Export (API) in ComfyUI) or a UI workflow with nodes and links. A UI workflow is
converted with tools/ui2api.py; when it uses a node that tools/object_info.json does not list, pass
--object-info with GET /object_info from a ComfyUI that has that node. --section (image, video, audio,
utility), --out (another folder) and --force (overwrite) are optional.
What it reads from the graph
| In the workflow | In app.json |
|---|---|
LoadImage, LoadVideo, LoadAudio |
a required image, video or audio input |
the words of a text encoder that reach a positive input, or the PrimitiveString node that feeds it |
prompt |
the same for negative, when it has words |
negative |
seed / noise_seed on samplers and noise nodes |
one seed input that writes all of them |
steps on samplers and schedulers |
a slider from half to twice the value |
literal width, height, length on an Empty…Latent… node |
selects (768–1536) and a frames slider |
ResolutionSelector |
an aspect-ratio select with the node's options |
PrimitiveInt, PrimitiveFloat, PrimitiveBoolean with a title |
a slider or toggle labelled with the title |
Save… nodes (PreviewImage becomes SaveImage) |
outputs, saved under freefield/<id> |
every loader value that ends in .safetensors, .ckpt, .pt, .pth, .gguf, .bin |
models |
A model file that one of packs/*.json already lists gets its repo, path, size, folder, license and gate from
there, and its phase from the apps that use it. The pack is the existing one that holds every model; otherwise the
app gets a new pack named after its id. A file no pack lists gets "repo": "TODO/unknown" and a size of 0.001 GB,
and the report names it.
Example
The official FLUX 3 Klein edit template, in UI format:
ui-test: 7 inputs, 1 outputs, 4 models → apps/ui-test
inputs:
- image (image) → 11.image
- prompt (textarea) → 48.value
- prompt2 (textarea) → 49.prompt
- steps (slider) → 7.steps
- width (select) → 44.width
- height (select) → 44.height
- seed (seed) → 7.seed
outputs:
- image ← 9 (SaveImage)
models matched in packs/:
- flux3-image-klein-int8-convrot.safetensors black-forest-labs/flux-3-image-klein-comfy 5.83 GB (denoise)
- qwen3vl_4b_int8_convrot.safetensors black-forest-labs/flux-3-image-klein-comfy 4.86 GB (encode)
- flux2-vae.safetensors black-forest-labs/flux-3-image-klein-comfy 0.34 GB (decode)
models Freefield does not know:
- qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors (loaded by CLIPLoader 46.clip_name; folder guessed as text_encoders)
check by hand:
- converted from the UI format with object_info.json
pack: ui-test (new: no existing pack holds every model)
validator: 0 errors, 0 warnings
It exits 1 when the app does not validate, 2 on a bad id or an existing folder without --force.
What it cannot know
- Which values a user should change. Every titled primitive becomes an input, frame rates and switches included; delete the ones that are internal.
- Which seed matters. All seed fields get the same input; a two-stage graph often wants only one of them.
- Prompts that are not literal text: words assembled by
StringConcatenate, or a prompt-enhancer's own instructions, which show up as a second prompt (prompt2above). - Where an unknown model comes from. Find the file on Hugging Face, fill
repo,fileandgb, then runpython tools/catalog.py all; until then the new pack lists a TODO repo. - Presets, gallery selects, templates around
{input}values, credits and a showcase lab case: those are written by hand, as in the other apps.
Tests: python test/import_workflow_test.py imports z-image, remove-bg, edit and video and compares the
result with their hand-written app.json.