Most text-to-video models struggle to reliably transfer a specific person's identity into a generated clip. This LoRA tackles that gap by injecting a reference latent into the target sequence with a distinct RoPE "source tag" and training with an ArcFace identity loss — a pragmatic recipe that turns LTX-2 into a dependable reference-to-video identity transfer module.
What Sets It Apart
-
Overlap reference + TASS‑RoPE source-phase: the reference latent shares frame-0 spatial grid (overlap) but is multiplied by a distinct RoPE phase (source_id), so the model cleanly separates "who is the reference" from "who to generate". So what: reduces identity confusion and enables multiple reference sources for multi-subject scenes.
-
ArcFace identity supervision: decoded predictions are encoded with an ArcFace projector during training and pulled toward the reference embedding. So what: enforces perceptual face similarity beyond token conditioning and improves face-level fidelity.
-
Character-sheet continuation checkpoint: trained on 4-panel inputs (face + front/side/back) with a native-resolution requirement. So what: lets the model reproduce clothing and body build consistently, not just facial identity.
-
ComfyUI integration (BFS Nodes) and prompt tooling: includes an LTX Identity Transfer node and a Prompt Enhancer that extracts identity attributes from the reference image. So what: ready-to-run for users of the ComfyUI LTX workflow without ad-hoc graph hacks.
Who it's for and trade-offs
Great fit if you need to generate short, identity-preserving video clips from one or a few reference images and you are running LTX-2 in ComfyUI. The model is most reliable with clean, frontal close-up references (or a 1536×1024 character-sheet for clothing/body). Expect better identity when prompts use the recommended ref_t2v: prefix and describe the person's attributes and actions.
Look elsewhere or adjust expectations if you need large-angle/full-body identity transfer without a character-sheet, want zero "first-frame copy" artifacts, or require guaranteed metric-robust identity scores for tiny/occluded faces. Practical constraints: requires the LTX-2 base model, BFS Nodes dependencies (insightface, Gemma encoder), and careful handling of ref_resize_mode for the character-sheet checkpoint.
Technical note: the optional ArcFace projector is available but had marginal gains in the author's experiments; identity is carried mainly by the overlap reference latent and prompt conditioning.
