Fix Face ID reference prefix with LTX keyframe conditioning

[jjdejong]

Summary

Fixes a shape mismatch when LTX Reference Enable / Best Face ID reference-token injection is combined with LTX guide/keyframe conditioning, including looping sampler workflows.

Reproduction

The sampler fails in LTXVModel._process_output() with an error such as:

shape mismatch: value tensor of shape [9248, 128] cannot be broadcast to indexing result of shape [1, 8959, 128]

For a 31-frame, 17x17 latent tile, LTX keyframe filtering leaves 8,959 target tokens. The Face ID patch then prepends one 17x17 reference frame (289 tokens), producing 9,248 output tokens. The base LTX output path still applies the original 8,959-position grid_mask.

Root cause

LTXVModel._process_output() reconstructs guide/keyframe outputs before the existing instance-level unpatchify wrapper runs. The wrapper therefore removes the Face ID prefix too late.

Fix

Patch LTXVModel._process_output() so it removes the pending Face ID reference prefix from both the model output and video timestep tensor before LTX performs keyframe reconstruction. The pending length is cleared so the later unpatchify wrapper does not strip the prefix twice.

No node API, workflow format, or default behavior is changed for graphs that do not inject a reference prefix.

Validation

• Python syntax compilation passed.
• git diff --check passed.
• The repository has no pytest suite; pytest is not installed in the test environment.
• Full GPU generation validation is still needed on an LTX 2.3 ComfyUI setup.