Profiles

CodeComfy covers six capability profiles. One command reaches all of them:

Ctrl+Shift+P → CodeComfy: Run… (all profiles)

The flow is always the same — profile → preset → inputs.

Inputs come from the preset, not the profile

The questions you get asked are derived from the chosen preset’s own workflow graph, not from a form hardcoded per profile.

That means two presets in the same profile can ask for different things. An image-to-video preset asks you for a source image because its graph needs one; a text-to-video preset sitting right next to it does not. Nothing is special-cased — the graph is the source of truth.

The inputs you may be asked for:

Input	When
Prompt	The graph has a text-conditioning node
Negative prompt	The graph has a reachable negative-conditioning path
Style tags	Audio presets — genre, instrumentation, mood
Edit instruction	Image-edit presets
Query	Inference detect / segment — what to look for
Source image	Image-to-video, image-to-mesh, edit, and every inference preset
Source audio	Stem separation

Files are uploaded to your ComfyUI server automatically. CodeComfy uses the filename the server reports back, not the one it sent — with overwrite off, ComfyUI renames on collision, so the returned name is the only reliable handle.

Nothing is submitted before it can succeed

Before any graph is sent, CodeComfy checks it against your server:

Node types — via GET /object_info/{class}. A missing node names the pack that provides it.
Model files — via GET /models/{folder}. A missing model names the file and the folder it belongs in.

If anything is missing, nothing is submitted and no GPU time is spent. If your ComfyUI is too old to list models, the model check is skipped rather than reported as a failure — an old server should never look like a missing file.

The profiles

Image

Text-to-image, image editing, and ControlNet.

Preset	Notes
Qwen txt2img (2512)	UNETLoader path, 16-channel latent
Qwen edit (2511)	Takes reference images into the edit encoder
ControlNet union — Qwen	Union ControlNet, optional core Canny preprocess
ControlNet union — SDXL	Same code path, SDXL stack

Video

Text- and image-to-video on real temporal models. These take length in frames, not a batch of stills — the frame count is snapped to the legal 4n + 1 grid before submission, because ComfyUI accepts an off-grid value without complaint and the model then mishandles it.

Preset	Notes
Hunyuan 1.5 i2v	Requires a start image
Hunyuan 1.5 720p	Higher resolution
Wan 14B
LTX standard
Mochi
Core output	Ends in `CreateVideo` → `SaveVideo` — no FFmpeg needed

Most video presets terminate in VHS_VideoCombine, which needs the Video Helper Suite node pack. The core-output preset avoids that dependency entirely and is encoded by the server.

Audio

ACE-Step 1.5 text-to-music, plus stem separation.

Preset	Notes
Music	Standard length
Music (long)
Music (MP3)	MP3 rather than FLAC
Jingle	Short-form
Draft	Fast, low-step
Stem separation	Needs the audio-separation node pack and a source audio file

Music presets take style tags rather than a prose prompt — genre, instrumentation, timbre, mood.

3D

Hunyuan3D-2 image-to-mesh, exported as GLB. All core nodes.

Preset	Notes
Draft	Fastest
Standard
Detail	Highest quality

Give it one source image; you get a .glb in .codecomfy/outputs/.

Inference

Florence-2 image understanding. Needs the comfyui-florence2 pack.

Preset	Output
Caption	Short description
Detailed caption
More detailed caption
Tag	Tag list
Detect	Bounding boxes — takes a query
Segment	Segmentation mask image — takes a query
OCR	Text found in the image

Most inference presets write their result as text; CodeComfy saves it as a .txt alongside your other outputs so it is a real artifact, not just a log line.

Metadata

Read the workflow ComfyUI embeds in its output PNGs. This profile never contacts the server — it is a local file read, so it works with ComfyUI closed.

CodeComfy: Read Workflow from PNG opens the extracted graph in a new editor tab. It prefers the prompt chunk (API format — the graph actually submitted, and the one that can be re-submitted) over the workflow chunk (editor format, which /prompt does not accept).

This is the fastest way to recover exactly what produced an image somebody sent you.

Where the workflows come from

CodeComfy does not author workflow graphs. The 27 reference workflows are vendored from comfy-headless, where every node type is verified against the live ComfyUI catalog.

This matters more than it sounds. A wrong workflow graph does not raise an error — it runs to completion and returns nothing. Keeping a second, hand-maintained copy of that knowledge would be a silent-failure generator, so maintainers sync it instead:

npm run kb:sync

npm run kb:check fails if the vendored copy has drifted from upstream.

Progress, cancelling, and the queue

Live progress. CodeComfy opens ComfyUI’s WebSocket before submitting, so the status bar shows real sampler steps (Step 12 / 20 (60%)) rather than a single static message. If your VS Code build has no WebSocket, or the socket drops mid-run, CodeComfy falls back to polling automatically — outputs are unaffected either way, because the socket is only ever an optimisation.

Cancel (CodeComfy: Cancel Generation) clears the pending queue and interrupts the running job. The order matters: /interrupt on its own aborts only what is running, and the server immediately promotes the next pending item — which is why cancelling used to look like it started the next job.

Clear queue (CodeComfy: Clear ComfyUI Queue) drops pending work and leaves the running job alone. This is the honest version of “pause”: mainline ComfyUI has no pause primitive and no resume-at-step-N, so stopping more work from starting is the part that can genuinely be done.