Replicate okaris/omni-zero

Omni-Zero is the niche one in this pack, and it's genuinely cool if you're in the niche. It's a zero-shot stylized portrait pipeline: you hand it a face you want to keep recognizable plus a style you want it dressed in, and it produces a portrait that has both - no fine-tuning, no training a LoRA, no waiting. The model itself is open source (it's on GitHub under okaris/omni-zero), but running it means wrangling insightface, a diffusion pipeline, and several control signals. This node hands the whole mess to Replicate via the Comflowy API and returns a finished image.

If that sounds like "style transfer but for identity," that's exactly the frame. It's the kind of thing you reach for when you want a consistent character across a stylistic look - think stylized avatars, themed portraits of the same person, or keeping a subject recognizable through a heavy artistic treatment.

How it works

Like every Replicate node in the pack, this one is generated from the model's schema and proxied through Comflowy's backend: your inputs go to app.comflowy.com/api/open/v0/flowy with your Comflowy key, Comflowy runs the model on Replicate, and the output image comes back as a standard IMAGE tensor. Images you plug into IMAGE-typed inputs get base64-encoded automatically; the STRING inputs take URLs.

The pipeline fuses several control signals at once - identity, style, optionally composition and depth - each with its own strength knob. That's the whole trick: it's not one prompt, it's a weighted bundle of "who," "how," and "where."

Inputs worth knowing

• identity_image and style_image (both required, both STRING) - these take URLs. The identity image is the face to preserve; the style image is the look to apply.
• image and depth_image (optional IMAGE) - drop actual tensors in here, from any ComfyUI node. image acts as the base/composition; image_strength (default 0.15) controls how hard that base is held.
• model - omni-zero or omni-zero-realism; the latter leans toward photographic realism over the stylized default.
• prompt / negative_prompt - yes, it still has those. Default prompt is just "A person".
• guidance_scale (3), number_of_steps (10), number_of_images (1-4), seed (42).
• The strength family: identity_strength, style_strength, composition_strength, depth_strength - all 0–1, all defaulting to 1 except depth at 0.5. Lower a strength to relax that signal.
• force_rerun - same caching trick as the other Replicate nodes; toggle it to force a fresh prediction.

Output is a single IMAGE.

Installing and the gotchas

Same story as every node in this pack: install Comflowy's Custom Nodes via ComfyUI Manager (or git clone https://github.com/6174/comflowy-nodes into custom_nodes, then restart), set your key with the Comflowy Set API Key node, and you're off. The only dependency is requests.

Where people get burned here:

• URLs only for the required inputs. identity_image and style_image are plain strings - paste a URL, not a file path. If you have the images locally, load them and wire them into the optional IMAGE inputs instead.
• The identity won't hold if the strength is wrong. This is a zero-shot model, not a fine-tune. Crank identity up and keep the style from overwhelming it, or you'll get a great style on a stranger's face.
• It costs credits. All Comflowy API nodes bill in Comflowy credits, and Replicate models are priced by compute time - Omni-Zero isn't the cheapest one in the pack. Test with number_of_images at 1 and low step counts.

If you've been fighting face-consistent styling with IPAdapter and it's not quite landing, this is worth a spin. Just budget for the fact that you're paying per attempt, and each attempt is a few seconds of server-side diffusion.