Replicate okaris/omni-zero-couples

Couples portraits are the classic case where identity tools get complicated. Most identity-preservation methods handle one face well and start fighting themselves on the second - you're juggling IP-Adapter weights, controlnets, and a prayer that both faces survive. Omni-Zero Couples is built for exactly this job: give it a photo of each person, and it generates a stylized portrait of the two of them together, no training, no LoRA, no per-person tuning.

The "zero" in the name is the pitch. It's a zero-shot pipeline - the model's paper and the okaris implementation run without fine-tuning for your specific people. You're renting that capability through Comflowy's cloud rather than running the pipeline locally, which spares you the multi-GPU dance this kind of method usually demands. The node's default prompt is already tuned for it - "Cinematic still photo of a couple..." - so even a lazy first run produces something usable.

How it works

It's another auto-generated Replicate node, so the flow is the pack standard: images get base64-encoded, the whole request goes to app.comflowy.com with the pinned okaris/omni-zero-couples version, and the output image comes back as an IMAGE tensor. What makes this node interesting is how much control it gives you over the composition, which is more than most one-click identity nodes.

The inputs that matter

• identity_image_1 and identity_image_2 - the two people. These are the whole point; without them there's no identity to preserve. Each has a matching identity_image_strength_ input (default 1).
• prompt / negative_prompt - the defaults are solid (the negative prompt bans anime, blur, text, and deformation). Keep them, tweak the scene words in the prompt.
• style_image + style_image_strength - a reference for the artistic style you want the couple rendered in. This is the "make it look like a painting / this illustrator's style" input.
• base_image + base_image_strength (default 0.2) - a starting composition, useful if you want a specific pose or framing and you let the model refine it.
• depth_image - if you have a depth map (from a ControlNet-style preprocessor or a depth-estimation model), you can hand it in to steer structure.
• guidance_scale (default 3), number_of_steps (default 10), number_of_images (up to 4), seed (default -1 for random).

For a first pass, honestly: two identity images, keep the default prompt, hit go. Add a style image when you want the look bent, and only start fiddling with depth and base images when the couple composition comes back wrong.

Install

It lives in Comflowy's Custom Nodes, so:

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

(Or ComfyUI Manager → search "Comflowy's Custom Nodes".) Restart, set your API key with the Comflowy Set API Key node, and the node appears under Comflowy Replicate. No models to download; the only dependency is requests.

Where people get burned

Cost and likeness are the two complaints you'll actually hit. This runs a full diffusion pipeline in the cloud, and it generates up to number_of_images of them - a run is a real credit spend, so set that to 1 while you're dialing in. On likeness, temper expectations: like every zero-shot identity method, it gets close and sometimes stops. It preserves identity best in stylized output; if you need forensic accuracy on a face, you'd be better off with a training-based approach or a proper face-swap rather than a stylized portrait generator. And as always with this pack, a black placeholder output means the API call failed - check your network and key before you blame the model.