LTX Identity Transfer

Most "identity" nodes in ComfyUI are IP-Adapter or InstantID descendants: they pull a face embedding out of your reference photo and steer a diffusion model toward it. LTX Identity Transfer is a different animal entirely. It doesn't touch CLIP or ArcFace at all - it patches the LTXV model at runtime so your reference image gets encoded as its own set of tokens, sharing the video latent's coordinate space, and the model treats it as "the thing to copy" because that's literally what it was trained to do. And critically, "identity" here isn't limited to faces - the author's own description is blunt about it: any subject works, a person, an object, an animal, whatever your checkpoint was trained on.

The catch is that this only works if you're running a checkpoint trained for it. This node is conditioning glue for a specific class of LTXV LoRA (the "reference/source_phase" style, related to the Edit Anything family this pack also ships nodes for) - load that LoRA onto your model with a standard ComfyUI LoRA loader first, then route the model through this node before your sampler.

What you actually touch

The required inputs are the usual LTXV sampling chain - model, positive, negative, vae, latent - plus reference_image, the thing you want copied in. It'll take a batch of images too, not just one: if your checkpoint was trained on multiple stacked references (layout='strata'), each image in the batch becomes its own reference block.

Three settings decide whether the transfer actually lands:

• layout - overlap (default), st_drc, or strata. This isn't a quality slider, it's a coordinate convention, and it has to match what the checkpoint was literally trained with. Get it wrong and identity transfer doesn't degrade gracefully - it just doesn't work, because the model never learned that convention.
• ref_resize_mode - match_target crops-then-resizes your reference to the output's pixel size (fine for a single centered face, but it silently discards the edges of a mismatched-aspect image like a wide character sheet). match_target_letterbox fits the whole reference in with padding instead of cropping. native_resolution encodes the reference at its own size - required if your checkpoint trained on a fixed reference resolution (a composite sheet, say) independent of the output video's size.
• source_id - the training default is 2; leave it unless you know your checkpoint used something else.

There's also reference_guidance_scale, an optional reference-flavored CFG pass (ST-DRC style): at 1.0 it's off, above that it runs a third forward pass per step with the reference dropped and amplifies the reference's own contribution the same way CFG amplifies your text prompt. Start around 2-4 if the model is ignoring your reference.

Outputs are the patched model, updated positive/negative, the latent to hand to your sampler, a debug string, and two preview images worth actually looking at: ref_preview shows exactly what got encoded, and crop_overlay draws a green box over what survived a match_target crop versus what got thrown away. Don't skip checking these - it's the fastest way to catch a resize mode that's quietly eating half your reference.

Installing it

Through ComfyUI Manager, search "ComfyUI-BFSNodes" and install. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/alisson-anjos/ComfyUI-BFSNodes.git
cd ComfyUI-BFSNodes
pip install -r requirements.txt

Restart ComfyUI. You'll also need a compatible reference-transfer LoRA - drop it anywhere under ComfyUI/models/loras/ (subfolders are fine) and load it onto your LTXV model with the standard LoRA loader before this node touches it. The pack's README is explicit that a plain LoRA loader by itself isn't the whole story for this conditioning style - the LoRA carries the weights, this node carries the runtime patch that makes them mean anything.

Where people get stuck

The single biggest failure mode is a layout that doesn't match the checkpoint - since it's a coordinate convention rather than a strength knob, the symptom is "nothing happens," not "it's a bit off," which throws people because it doesn't look like a normal LoRA-strength problem. Second most common: match_target's default center-crop quietly cutting off exactly the part of your reference sheet you needed (a face closeup off to one side, say) - check crop_overlay before assuming the model is broken, and switch to crop_anchor or match_target_letterbox once you see what's actually getting cropped. Because this node works by monkey-patching LTXV internals at runtime, it also wants to be the thing touching model right after your LoRA loader - don't sandwich other model-patching nodes between the two without checking they play nicely together.