Protoface
Official Protoface API nodes for ComfyUI with automatic full-catalog coverage. Video generation with MiniMax H3, LTX 2.5 Fast, Seedance, Kling, Veo, Runway and Luma. Image generation and editing with FLUX, Nano Banana, GPT Image and Qwen. Music, speech and 3D generation.
Nodes (104)
Qwen Image 3.0 Edit Without the 20GB Download
Pay Per Picture, Skip the Checkpoint
First Frame In, Fifteen Seconds Out
30-Second Clips With the Audio Already In
Ten Images, and the Prompt Decides Which Is Which
Animate That Still You Just Spent an Hour On
480p Is Your Friend
Character Consistency Without Training a LoRA
Point At Media You've Already Uploaded
The Editing Model You Can't Run, Wired Into Your Graph
Text That Actually Renders
Four Reference Images and No Patience Required
Iterate Until It's Right, Then Switch Models
The Gated Model, Doing the Thing Locally
Adding Twenty Seconds to the End of a Shot
9:16 by Default, Audio Included
Two Stills, One Shot, No Morphing
The Cheap Seat on a Very Good Model
Change the Shot You Already Shot
Four More Seconds, Cheaply
The tier that only makes sense at 720p
The draft machine you'll queue fifty times
Restyle footage you already shot without reshooting it
Bolt another five seconds onto a clip without a seam
The cheapest way into a closed video model
The full-fat tier, 4K included, priced accordingly
Prompt-driven video editing that actually holds the subject
Chaining 15-second segments into a real shot
4K, references, and the best-looking invoice you'll ever dread
Thirty seconds, one shot, no aspect ratio override
Same shot, less waiting, same bill
The prompt gets a prefix and you lose two dropdowns
The 720p feedback loop for people who hate waiting
30-second continuations, and no end-frame steering
Thirty seconds from a prompt, at flagship prices
720p/1080p only, and worth it anyway
Layers and masks out of a flat render
Multi-reference editing for four and a half cents
Five cents a render, and it rewrites your prompt
Forty voices, one kilobyte of text per call
The voice everyone benchmarks against, wired into your graph
A prompt-to-song node for when ACE-Step won't cut it
Keep the performance, swap the pipes
Google's music model, with an image as the brief
Reference in, same character out
The closed model that made everyone stop arguing about prompts
The Veo you can actually afford to iterate with
Give a still a camera and a soundtrack
Turn an 8-second shot into a scene
The one with the audio baked in
Your subject, still your subject
The one that can actually spell
One input, two useful outputs
Motion that starts where your frame left off
Multi-shot storyboards from one node
Your character, someone else's dance
The fast pass on your key frame
1080p drafts that don't cost a session
The hosted version of the model you'd otherwise run at home
Start frame, end frame, five seconds
Fixing the clip Luma already generated
Three knobs, one bill
A talking head with no prompt field at all
One image in, one edit out
Four fields and a bill
Five seconds, 1080p, no knobs
The two-field video node
The pack's flagship, all fourteen inputs of it
Lyrics, a prompt, and a finished song
Multi-image in, no seed, argue with it in English
Quality tiers, no seed, no undo
The cheapest way to make a still move
Prompt, length, size, done
Chaining stills into a moving sequence
MiniMax H3 with the audio you can't switch off
Every input at once, watermark included
The cheap fast one you'll actually use for drafts
The node that lets you type a resolution
You specify megapixels, not a scale factor
The design model with two fields
Recraft's boring-named node is the one for clean, graphic-looking stills
Real SVG out of a ComfyUI node, and why it won't plug into anything
Stop re-uploading the same image to every node in your chain
Your browser crashed and the job already billed you. This gets it back.
Prompt-driven video editing with keyframe anchors — and a brutal per-second bill
When the turbo draft isn't good enough
Iterate here, finish somewhere else
One still into five seconds of motion, at 28 credits a second
The node that saves the things Save Image can't touch
Dub the clip you already have, no per-face training
A mesh in one call, and a face budget you should actually set
Hunyuan3D V3 from images
Tripo H3.1 text-to-3D
Quads, orientation, and where the texture lands
The Aleph 2 keyframe position helper
One helper per gap, and the arithmetic will bite you
The durations have to add up to the clip
The Kling 3.0 Pro shot helper with a real prompt box (image-to-video)
Yes, there are two identical Kling 3.0 Pro Shot nodes. This is the text-to-video one.
One image in, one edited image out, six credits
Grok Imagine Image 2.0 in ComfyUI, With No Weights and No VRAM Bill
Give Grok Imagine a Still and Get Motion Back — No Video Model on Your Disk
Text to Video for People Without a Video GPU
Grok Imagine's Reference Mode
Protoface for ComfyUI
Official Protoface API nodes for ComfyUI with automatic full-catalog coverage. Generate video with MiniMax H3, LTX 2.5 Fast, Seedance, Kling, Veo, Runway and Luma. Generate and edit images with FLUX, Nano Banana, GPT Image and Qwen, and create music, speech and 3D assets.
One extension provides searchable model/task nodes with native media connections and controls: currently 44 models and 95 tasks. Browse the catalog by model name or task, then connect it to your ComfyUI workflow.
Latest approved release: 0.1.2. Protoface on Comfy Registry, package protoface-nodes, publisher protoface. Representative image, video, audio and reference workflows have been tested against the production API. Mock coverage includes every bundled task; it does not establish every model's production quality or latency.
Install
Requires ComfyUI 0.41.0 or newer and Python 3.10 or newer. ComfyUI Desktop and local/self-managed hosted installations can load this custom-node package. Comfy Cloud does not permit arbitrary independent custom-node installation. This extension bills your Protoface account directly; it is independent of Comfy Partner Nodes and Comfy credits.
With ComfyUI-Manager, open Extensions (or Manager), search Protoface, install protoface-nodes, then restart ComfyUI and reload the browser. If Manager does not show Protoface, refresh its catalog. You can also use the tagged Git or ZIP installation below. After installation, continue with API-key setup in step 4.
For a manual installation of 0.1.2:
- Clone the release tag from the ComfyUI directory:
Alternatively unpack the 0.1.2 Registry source ZIP intogit clone --branch v0.1.2 --depth 1 https://github.com/protoface-ai/protoface-comfyui.git custom_nodes/protofaceComfyUI/custom_nodes/protoface. - With ComfyUI's Python interpreter, install
requirements.txt:
Desktop users should use its managed environment; Windows portable users should usepython -m pip install -r custom_nodes/protoface/requirements.txtpython_embeded/python.exe -m piprather than a system Python. - Restart ComfyUI and reload the browser.
- Click Protoface in the lower-right corner and enter an API key from your Protoface account. Interactive keys stay in server memory and are excluded from exported workflows.
- Search Protoface, MiniMax H3, a model name, or a task such as Text to Video in ComfyUI's Add Node search. Choose a model/task node, fill its controls, then queue it.
For a ready-made graph, open Templates → the Protoface extension and choose MiniMax H3 Text to Video, LTX 2.5 Fast Image to Video, Qwen Image 3.0 to MiniMax H3 Asset Reuse, or Resume Run.
Headless/persistent setup uses operator environment variables:
export PROTOFACE_API_KEY='your-api-key'
# Alternatively set PROTOFACE_API_KEY_FILE to a private file containing the key.
Never put a key in workflow JSON. Anyone allowed to execute workflows on a shared ComfyUI server can use its configured operator key; use separate server instances or an authenticated deployment with appropriate access controls for separate accounts.
Work with media and controls
Each visible model/task has its own native controls. There is one shared implementation underneath, rather than separately maintained code per model.
- Prompts, enums, numbers and booleans come from public schemas. Default omits an optional parameter; Set supplies it; Null explicitly sends null where supported. Blank text is distinct from omission.
- Media inputs offer native Image, Video or Audio modes. Connect Load Image, native video, or native audio nodes. List inputs grow as you connect assets; input batches count toward the model's reference limit.
- References mode accepts Protoface Asset Reference (existing asset ID/public URL) or Protoface Reuse Result Asset. Reuse result assets to chain remote jobs without reuploading their media.
- Kling multi-shot and Pika transition controls use generated typed item nodes. Search the model name plus Shot or Transition, connect one helper per item, and select Set on the list control.
- Images, masks, videos, audio, SVG and 3D files use native typed outputs. Multiple results are output lists, including heterogeneous image sizes. ComfyUI maps connected downstream nodes over these lists. A mask uses native inverted-alpha semantics.
- files contains every original downloaded asset, including transparent PNGs and non-media artifacts. Protoface Save File writes originals to ComfyUI's output directory. A 3D file's actual format remains authoritative; a GLB-only downstream node may need a conversion if the provider returns another format.
- run id is available for recovery and support. Protoface Resume Run retrieves an existing job without submitting another generation.
Paid generation, caching and interruption
Keep generation_id fixed to reuse or resume the same job. Increment it, or choose ComfyUI's increment-after-generation option, to request a fresh paid generation. This is separate from a model's random seed. Changed prompts, assets or model parameters constitute a new request.
The extension journals a request identity before submission and reuses its API idempotency key on ambiguous failures. Requeueing unchanged inputs after a restart resumes the recorded run. Interrupting execution requests cancellation; API cancellation is best effort and does not imply a refund. A polling timeout leaves a recoverable job rather than submitting another one. An unresolved submission near the API's 24-hour idempotency limit stops and asks you to find the run in your account.
Recorded runs resume without uploading their references again. For a new request, cached uploads are checked against the API; a deleted asset is uploaded again. If an ambiguous submission's reconstructed body changes, the adapter stops for recovery rather than retrying a different paid request under the same key.
Every model node contains a serialized, non-secret request_scope assigned by the frontend. Headless workflows must supply their own stable unique workflow/node scope and preserve it when resuming. Deleting the state directory, assigning a new scope, or changing the generation ID loses that deduplication intent and can lead to a new paid job.
Catalog updates
The bundled catalog loads offline without an API key or startup network request. Protoface → Refresh catalog checks the live public schemas and caches a validated snapshot. Restart ComfyUI and reload the browser to load added/changed nodes. Ordinary new models using supported schema/media types need a catalog refresh rather than an extension-code update.
Nodes keep an immutable schema identity. Old definitions remain loadable from bundled/local history and are marked archived after a schema change or retirement. Archived definitions refuse new submissions; select the current model node or use Resume Run for previous jobs. Existing results are never automatically regenerated because a package version changed.
New schema constructs, media modalities, or provider-specific behavior can require an extension update. Compatible tasks still load when other tasks need an unsupported control or media adapter; Protoface settings lists the skipped tasks and their reasons. A malformed catalog, or one with no supported tasks, keeps the working snapshot. Publishing a model-list endpoint alone does not prove a future model is usable in ComfyUI.
See installation, updates and troubleshooting, release and Registry setup, development checks, architecture, and example workflows.
Operator limits
PROTOFACE_STATE_DIR selects the private catalog/job/media cache (default ~/.cache/protoface-comfyui). Preserve this directory for job recovery; do not share it between accounts or delete it to troubleshoot an unresolved paid submission.
Optional operator settings are PROTOFACE_CONCURRENCY (default 2), PROTOFACE_JOB_TIMEOUT_SECONDS (7200), PROTOFACE_POLL_INTERVAL_SECONDS (2), and PROTOFACE_MAX_DOWNLOAD_BYTES (2,000,000,000). PROTOFACE_BASE_URL is trusted server configuration for a proxy/test deployment; it is never a workflow input. Changing credentials changes the cache namespace; use Resume Run for an existing job after API-key rotation.