Replicate tencentarc/photomaker-style

If you read the article for tencentarc/photomaker, this node is its sibling - same family, different temperament. Both take one to four photos of a person and generate that person somewhere new, but where the plain PhotoMaker defaults to photorealistic output, this style variant is the one tuned for pushing a face into an art style. The model's own description calls it the "Stylization version," and its default example prompt is telling: "A girl img riding dragon over a whimsical castle, 3d CGI, art by Pixar." That's not a photo; that's a character design brief.

So the honest use case: you have a face you want to keep recognizable while the medium changes - Pixar render, comic book, low-poly game art. That's this node's whole job, and it's the exact kind of thing that's painful to do locally, where you'd normally be chaining IP-Adapter with a stack of style LoRAs and hoping nothing fights.

How it works

Mechanically it's identical to the standard PhotoMaker node - the same ID-embedding stack that fuses your reference photos, the same img trigger token, the same Replicate proxy through Comflowy's cloud. The difference is tuning: this variant defaults to "(No style)" for style_name (the plain node defaults to "Photographic"), and it's biased toward stylized interpretation rather than faithful photorealism. Same inputs, same output (IMAGE), same credit burn per run.

The inputs that matter

• input_image (required) + input_image2/3/4 - your reference photos. Keep them all the same person; that rule doesn't change.
• prompt - keep the img token, then spend your words on the style and scene. The style templates below reinforce it, but the prompt does the heavy lifting.
• style_name - the same 11 templates ("(No style)", "Cinematic", "Disney Charactor", "Digital Art", "Fantasy art", "Neonpunk", "Enhance", "Comic book", "Lowpoly", "Line art"). With the default "(No style)", the prompt alone drives the look - which is often what you want for fine-grained control.
• style_strength_ratio (default 20, 15–50) - cranking this makes the chosen template dominate; keeping it low preserves the reference face better.
• num_steps, guidance_scale, num_outputs, seed, disable_safety_checker - same as the plain node.

Which one do you actually want?

That's the real question, and the answer is short: photorealistic portraits in new settings → the plain tencentarc/photomaker. Same face in an illustration, comic, or 3D-render style → this one. They're near-identical to set up, so if you're not sure, run the same reference photo and prompt through both and keep whichever matches your look. Nobody's judging you for having both nodes on the canvas.

Install

Part of Comflowy's Custom Nodes, installed like everything else:

cd ComfyUI/custom_nodes
git clone https://github.com/6174/comflowy-nodes

Or ComfyUI Manager → search "Comflowy's Custom Nodes". Restart, set your key once with the Comflowy Set API Key node, and you're set. No model downloads; it all runs on Comflowy's side.

Where people get burned

The img token still matters - forget it and the identity signal never attaches. Stylization is also where likeness goes first: the more aggressively you push style_strength_ratio and a heavy template, the more the face drifts toward "inspired by" rather than "that's them." Keep the strength low if recognition matters. And same as every Comflowy node, a black placeholder output means the API call failed - network or key - not that the model rejected your reference.