Replicate fofr/face-to-many

Feed it a photo of a face and it spits that person back out as a 3D character, an emoji, a video-game render, pixel art, a clay figure, or a toy - while keeping them recognizable. It's fofr's popular face-to-many model, and it went semi-viral for exactly that "make me a Pixar character" trick. This node runs it on Replicate's cloud, so there's nothing to install or download beyond the wrapper, no VRAM needed, and it works on any machine - but you need a Replicate account and a REPLICATE_API_TOKEN, and each run costs Replicate credits.

How it works

The likeness-keeping is the interesting part, and it's built on InstantID: a zero-shot face adapter that reads your reference face's geometry (via facial landmarks) and injects it into the generation, so the identity survives a total style change. On top of that it uses a depth ControlNet to hold the head's structure and a style LoRA per look. So "3D" or "Clay" isn't a filter - it's a fresh generation that's been pinned to your face. That's why a clean, front-facing reference photo matters so much: InstantID gets its likeness from one image, and a bad one gives it little to work with.

The inputs and outputs that matter

• image - the face. Clear, well-lit, roughly front-on. This is the whole game.
• style - 3D, Emoji, Video game, Pixels, Clay, or Toy. The look.
• prompt (default "a person") - describe the character if you want to push it ("a person as a wizard").
• instant_id_strength (1) - how hard it locks onto your face. High keeps the likeness; too high can fight the style.
• denoising_strength (0.65) - how far it departs from the source. Higher = more stylized but looser likeness.

Output is a single IMAGE. Other knobs: prompt_strength (4.5) for prompt adherence, control_depth_strength (0.8) for how tightly it holds head structure, negative_prompt, custom_lora_url + lora_scale to layer your own LoRA, seed, and force_rerun to bypass Replicate's cache.

How to install it

One install brings every Replicate node.

• ComfyUI Manager: search ComfyUI-Replicate, install, restart.
• Manual:

  cd ComfyUI/custom_nodes
  git clone https://github.com/replicate/comfyui-replicate
  cd comfyui-replicate
  pip install -r requirements.txt
  

  then restart.

Nothing downloads locally - just the replicate pip package. Set your token first: export REPLICATE_API_TOKEN="r8_..." (replicate.com/account/api-tokens).

Common issues & troubleshooting

The likeness is off. Almost always the reference photo. InstantID pulls identity from a single image, so give it a sharp, front-facing, well-lit face with no heavy occlusion (sunglasses, hands, extreme angle). Then raise instant_id_strength toward 1 and ease denoising_strength down a touch so it doesn't stylize the face away.

It's too stylized and stopped looking like the person - or the opposite, too photo-real and barely styled. These two knobs pull against each other: denoising_strength and instant_id_strength. More style, more denoise; more likeness, more InstantID strength. Find the balance for the look you want; the defaults are a decent starting point.

No face detected / it errored on the input. InstantID needs to actually find a face. A profile shot, a tiny face in a wide frame, or a heavily obscured one can fail detection. Crop closer to the face and try again.

Same result on a re-run. Replicate cached it for identical inputs. Change the seed or set force_rerun for a new take.