Nodes/FunUI/FunUI Identity Transfer
ComfyUI Node

FunUI Identity Transfer

Carrying a Face Into an LTX Clip Without Training a LoRA

By digital-garbage·Created about a year ago·Updated about 10 hours ago· 21
FunUI Identity Transfer
  • model
  • vae
  • positive
  • negative
  • latent
  • image
  • model
  • positive
  • negative
  • status
◄identity_projectorNone►
◄source_id2►
◄phase_scale1.0►
◄id_strength1.0►
◄arcface_modeauto_adjust►

Identity in video is normally either a LoRA you trained or an adapter bolted onto a UNet. FunUI Identity Transfer does neither: it encodes your reference face with the picture VAE and appends it to the attention sequence as extra tokens the model can look at. The author calls it Best-FaceID style, and it's a port of LTX's own identity-transfer phase tag.

The honest framing first: with no image wired, everything passes through untouched. That's not a fallback mode, it's the default - this node is a no-op until you give it a face.

What it actually does

The reference image is resized to the clip's latent frame size, encoded by the VAE you hand it, and installed on a clone of the model as a reference latent segment. That segment gets tagged with a source phase - source_id × phase_scale - which is how the model knows "this is reference material, not the current shot". The default source_id of 2 is the value the trainer used; setting it to 0 turns the rotation off while leaving the tokens in place, which is a switch you'll actually use when a reference leaks visibly into the frame.

Optionally there's a second channel: an ArcFace embedding passed through a small projector network, appended to the positive conditioning as text-style context tokens (and a zeroed version to the negative). That's the "FaceID" half. Per the tooltip it's a weak channel - 5–20 is a normal id_strength to test, against a default of 1 - and the overlap tokens carry most of the identity either way.

Failure is treated as a downgrade, not a crash. No face detected? "ArcFace found no face: overlap tokens only" in the status. insightface not installed? Same, with the pip line to fix it. The clip still generates.

Inputs and outputs

The required set is the whole graph around it: model, vae (the picture VAE used to encode the reference), positive, negative, latent (the clip's starting latent - the reference is sized to its frame, so this has to be the real one, not a placeholder).

Then the interesting ones:

  • image - the reference face. Leave it unwired and nothing happens.
  • identity_projector - an ArcFace projector .safetensors from models/loras, or None. It's not a LoRA in any functional sense; it just lives in that folder.
  • source_id - the phase id, 0–8, default 2.
  • phase_scale - multiplies that phase.
  • id_strength - multiplies the ArcFace tokens only, not the overlap tokens. 0–50.
  • arcface_mode - auto_adjust (retries detection zoomed out and up), as_is (detect on the image only), disable (skip ArcFace entirely).

Outputs: model, positive, negative, status. Wire all three first outputs onward - the model carries the reference tokens, and the conditioning carries the ArcFace ones. status is your only witness; it says which of the two channels fired.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/digital-garbage/FunUI
cd FunUI && pip install -r requirements.txt

ComfyUI Manager: search FunUI. Restart. (The pack page reads olivv-cs/ComfyUI-FunPack; the repo is digital-garbage/FunUI now - same project, renamed.)

For the ArcFace half you also need InsightFace, which is deliberately not in requirements.txt:

pip install insightface

No InsightFace means overlap tokens only, which is a workable mode.

Where people get burned

LTX 2.x is untested in this pack. The README says it in bold: it has never been run on a real LTX model and nobody plans to test it soon. This node names LTX in its own description, so you are squarely in "you try it, you file the issue" territory. Set expectations accordingly and don't debug your prompt until you've confirmed a plain single-shot clip works.

A 4D latent. It needs a video latent with frames, height and width; a still-image latent raises. Build the latent with FunUI Empty Latent and the matching model.

Expecting IP-Adapter strength. Token injection is a lighter touch than a FaceID adapter at weight - it nudges, it doesn't lock. If the face keeps drifting, don't crank id_strength to 40; fix the reference (straight-on, well lit, face filling a decent part of the frame) and try the ArcFace channel at all before deciding the node is weak.

InsightFace licensing. The Python library is MIT, the pretrained models aren't. If this is commercial work, that's a thing to check before it becomes a dependency.

Forgetting the negative. Both conditioning outputs carry changes when ArcFace fires. Wire only the positive and the negative branch of your CFG is now inconsistent with it.

CategoryFunUI/Conditioning

Inputs (11)

NameTypeDefaultDescription
modelMODEL—
vaeVAEThe picture VAE, to encode the reference.
positiveCONDITIONING—
negativeCONDITIONING—
latentLATENTThe clip's starting latent: the reference is sized to its frame.
imageoptIMAGEThe reference face. None = no identity transfer.
identity_projectoroptCOMBONoneArcFace projector .safetensors (models/loras). None = the overlap tokens only (they carry most of the identity).
source_idoptFLOAT20–8Source-phase id for the reference tokens (the trainer used 2). 0 turns the rotation off, leaving the tokens in.
phase_scaleoptFLOAT1.00–4—
id_strengthoptFLOAT1.00–50Multiplies the ArcFace tokens. A weak channel: 5-20 to test.
arcface_modeoptCOMBOauto_adjustauto_adjust retries detection zoomed out/up; as_is detects on the image only; disable skips ArcFace.

Outputs (4)

NameTypeDescription
modelMODEL—
positiveCONDITIONING—
negativeCONDITIONING—
statusSTRING—