I added a second masking mode while keeping the original behavior available. With the original mode, the mask limits which reference tokens are used for feature transfer, but Klein can still see the full reference image for context.

The new mode isolates the masked region more strictly. When connected unmasked reference tokens are blocked as attention sources, so the model only receives context from the selected area of that reference.

This should be especially useful with multiple references. For example, one image can provide full identity context while another contributes only an outfit,, face, or other specific region. It should also help with outfit swaps, close-ups, and references containing distracting backgrounds or unrelated details. The full documentation explains how the node works and how the two masking modes differ. I recommend reading the masking section before testing it.

I still recommend masking only what you need from the photos whether one or multiple as it give cleaner results :)

Names of masking modes :

Old behavior is focus_only

New behavior is zero_unmasked_tokens (recommended)

The node's documentation

The mask behavior documentation

I updated Identity Feature Transfer to remove the need for stacked/chained nodes.

clearer screenshot of the wf since reddit compresses the photos

Now the workflow is simpler:

• Use Multi ReferenceLatent for multiple reference images.

• Use Identity Feature Transfer Final for the identity pull.

• If you use masks, connect each mask directly to the matching mask input on the node.

  • subject_mask_1 = mask for reference 1

  • subject_mask_2 = mask for reference 2

  • etc.

The node handles the multi-reference setup internally, so you no longer need multiple stacked identity nodes for each reference.

Presets are still available, similar to the previous version.

For custom tuning, the two main knobs are:

• Temperature

• Similarity

Temperature is the main identity-strength control. Lower temperature gives a stronger, more direct 1:1 identity pull.

Similarity works more like a refiner/filter. It controls how selective the match needs to be before the node pulls from the reference.

So in practice:

• Lower temperature = stronger identity / more faithful match

• Higher temperature = softer, looser identity influence

• Lower similarity = allows more reference matches

• Higher similarity = stricter matching, more selective pull

example workflow (update to version 3.4.1 as there was a conflict with a node from a different repo causing the multireference latent node to be replaced if you had the other custom node installed and now that has been fixed)

Also just a little side note, this Final version uses a bit diff technique in term of pulls so 1:1 is achievable but needs to be careful enough to get it.

Previous posts for context:
multi ref latent

Iden transfer v3