LTX Identity Transfer

If you've already read about LTXIdentityTransfer in this pack, you can skip most of this - LTXIdentityOverlapConditioning ships with the exact same display name ("LTX Identity Transfer"), the exact same description, and an identical input/output schema. It looks like the class the pack settled on internally, with LTXIdentityTransfer as either the newer name or a compatibility alias for older workflows. Either way, functionally this is the node: a runtime model patch that injects your reference image as its own set of tokens sharing the video latent's coordinate space, so an LTXV checkpoint trained for it can copy a subject - any subject, not just faces - into the generation without going through I2V at all.

The mechanics, briefly

You wire it into the normal LTXV chain (model, positive, negative, vae, latent), point reference_image at whatever you want copied in, and it hands back a patched model plus updated conditioning and latent for your sampler. It also accepts a batch of reference images rather than just one - useful for checkpoints trained on multiple stacked references, where layout='strata' means each image in the batch becomes its own reference block at its own coordinate slot.

Three fields decide whether it works at all:

• layout (overlap / st_drc / strata) has to match the coordinate convention your checkpoint was actually trained with - it's not a quality dial, it's a protocol, and using the wrong one means the model never learned that layout, so identity transfer just doesn't happen.
• ref_resize_mode controls how your reference gets fit to the output's frame size: match_target crops-then-resizes (fast, but discards anything outside center on a mismatched aspect ratio), match_target_letterbox pads instead of cropping, and native_resolution keeps the reference at its own size for checkpoints trained on a fixed reference resolution independent of video size.
• source_id defaults to 2, matching what the reference checkpoints in this family were trained on - change it only if you know better.

Push reference_guidance_scale above its default of 1.0 and you get an extra reference-flavored CFG pass that amplifies how much the reference actually shows up, the same way normal CFG amplifies your prompt - a reasonable first move if the output is basically ignoring the reference.

The outputs worth actually opening are ref_preview and crop_overlay: the first shows exactly what got encoded from your reference, the second draws a green box over what a match_target crop kept versus threw away. If your identity transfer looks half right, this is where you find out why.

Getting it installed

Search "ComfyUI-BFSNodes" in ComfyUI Manager, or do it by hand:

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

Restart ComfyUI afterward. This node is only conditioning glue - it needs an LTXV reference-transfer LoRA loaded onto model with a regular LoRA loader before it runs; a bare LoRA load alone isn't sufficient for this style of conditioning, per the pack's own README. Drop the checkpoint under ComfyUI/models/loras/ (organized subfolders are fine).

What actually goes wrong

The recurring one is layout mismatch - because the failure mode is total silence rather than a quality dip, people burn time tweaking strength values on a node that has none, when the real fix is just picking the layout the checkpoint expects. Second: trusting the default match_target crop on an aspect-mismatched reference sheet without checking crop_overlay first, then wondering why a face or panel that was clearly in the source image never made it into the output. And because this works by patching LTXV internals live, keep it adjacent to your LoRA loader in the graph rather than routing other model-patching nodes between the two.