Protoface FLUX.2 Pro · Text to Image

If what you need is a product shot with legible copy on the packaging, this is the node. FLUX.2 Pro's selling point in one line is production-grade stills with sharp text rendering and natural detail - and text rendering is exactly where open models tend to fall over and where Black Forest Labs' line has been strongest since the first Flux.

The context you need

Flux 2 Dev never became a local default. The KB's account: the 32B release was rejected within a week for being too heavy, ~53GB for model plus Mistral encoder against Qwen-Edit's ~29GB, and the LoRA ecosystem never formed because training it needs data-centre hardware. Klein became the local editor of 2026 instead. Pro is the tier above both, API-only, and the honest reason to use it is that the local ceiling and this ceiling are not the same height.

Also worth knowing before you spend: BFL's image models carry provider-side content filtering, and FLUX 3 - the multimodal successor - shipped as gated early access with the open Dev backbone merely promised. So "wait for the open version" is not a strategy.

How this node is put together

Protoface is a catalog and an adapter, not a model runtime. On startup it reads bundled schemas for every model/task it supports and generates native ComfyUI V3 nodes from them - this one is black-forest-labs/flux-2-pro × image.generate. All nodes share one HTTP adapter: validate against the schema, submit, poll with a progress bar, download, decode into ComfyUI types. Since there's no input media on a text-to-image job, the upload step is a no-op; what's left is the journal, which records the request identity locally before submission so that an unchanged requeue resumes rather than re-bills.

Four controls

prompt is the real interface. Flux-2 responds to a fairly plain description; you don't need the comma-spam incantations that SDXL-era prompts trained us into. Name the subject, the medium, the light, and the text you want rendered - in quotes, so it's unambiguous.

size is a plain string field defaulting to 1024*1024, and the format is width*height with an asterisk. This trips people constantly. It's not 1024x1024, it's not two number widgets, it's one string. Non-square is fine: 1344*768, 768*1344.

seed is an optional integer; -1 means random, anything else is reproducible. If you're producing a set of variations that share a look, hold the seed and change only the text.

generation_id is the paid thing, not the creative thing. Leave it fixed and the pack reuses or resumes the same job; increment it and you've bought a fresh generation. It's deliberately separate from seed because "same image again" and "same prompt, new take" are different requests.

There is no negative prompt, no steps, no guidance, no sampler. Gradio-era habits don't apply.

Outputs

image - an output list, so downstream nodes map over any multi-result response automatically. mask - derived from the result's alpha, in ComfyUI's inverted convention, useful if the model hands back a cutout. files - the original downloaded asset (PROTOFACE_FILE) for Protoface Save File. run id - a string, and your recovery handle.

Install

Manager is one route: Protoface → install protoface-nodes → restart → reload. Manual is the other:

cd ComfyUI/custom_nodes
git clone --branch v0.1.0 --depth 1 https://github.com/protoface-ai/protoface-comfyui.git protoface
python -m pip install -r protoface/requirements.txt

That's four small pure-Python dependencies - httpx, jsonschema, Pillow, av - and zero gigabytes of weights. Needs ComfyUI 0.41.0+ and Python 3.10+. Desktop users should pip into the Desktop-managed environment; Windows portable users should use python_embeded\python.exe -m pip, not a system Python.

Then the key: click Protoface in the lower-right corner and paste it. It lives in server memory only, so it does not survive a restart and can never leak through an exported workflow - which is the right tradeoff, if mildly annoying.

Troubleshooting

"No Protoface nodes" after a manual install. The folder must be custom_nodes/protoface with __init__.py directly inside it. An extra nested folder from unzipping a release archive is the classic cause, and having two copies installed is the second.

A model that isn't listed. Protoface → Refresh catalog, restart ComfyUI, reload the browser. The catalog refresh pulls current public schemas; ordinary new models appear without a pack update, but a genuinely new control type needs one.

Errors that read like silence. Auth and credit problems surface as a job that doesn't produce, not a stack trace. Check the settings dialog shows https://api.protoface.com and the account has credit.

A scripted run fails on scope. Headless workflows must supply request_scope themselves - the frontend fills it in for you, the API can't guess it.

A 1024x1024 job came back rejected. You typed an x. Asterisk.