LTX-Best-Face-ID: Identity-Preserving Reference-to-Video LoRA for LTX-2 With ArcFace Loss and TASS-RoPE

Alisson Pereira (@Alissonerdx) has released LTX-Best-Face-ID, an identity-preserving reference-to-video LoRA for LTX-2 (22B). Give it a reference photo and a text prompt, and it generates a video that keeps that person's identity — all from a single image, without requiring a white-background reference.

One photo + a text prompt → consistent identity video. No special backgrounds needed.

Sample Outputs

What It Does

LTX-Best-Face-ID performs reference-to-video (ref_t2v) generation: one reference image of a person → a video of that same person performing the described action. The identity is injected through three techniques working together:

1. Overlap reference conditioning — the reference latent shares the frame-0 grid position with the first generated frame
2. TASS-RoPE source-phase tagging — a distinct rotary position encoding phase ("source tag") lets the model distinguish the reference from generated content
3. ArcFace identity loss — during training, the decoded face is compared against the reference embedding in recognition space

The result: identity consistency without the usual tricks like requiring clean white backgrounds or multi-view reference sheets.

How TASS-RoPE Works

The technical innovation is TASS-RoPE (Temporal-Adjacent Spatial-Shifted RoPE), from Chen et al.'s ST-DRC paper (arXiv:2606.02441). Instead of simple concatenation, the reference latent gets a multiplicative RoPE phase tag:

phase[d] = source_id · phase_scale · θ^(−d/L)
target tokens: source_id = 0 (exact no-op)
reference:     source_id = 2 (distinct rotary "tag")

This "source tag" lets the attention mechanism separate who is who in the sequence, strongly improving identity transfer. Because the tag is positional, the same mechanism gracefully handles multiple references (source_id 2, 3, 4...) for multi-subject conditioning.

Usage in ComfyUI

Requires the companion BFS Nodes: ComfyUI-BFSNodes

1. Install ComfyUI-BFSNodes via ComfyUI Manager (dependencies install automatically)
2. Load LTX-2 checkpoint + Gemma-3 text encoder / CLIP as usual
3. Add the LTX Identity Transfer (overlap + source-phase) node; feed it the reference image
4. Load LTX-Best-Face-ID LoRA on the MODEL path
5. Write a prompt prefixed with ref_t2v: and queue

Prompt Tips

• Always prefix with ref_t2v: — the model was trained with this format
• Describe the action, setting, framing, camera in detail
• Describing identity attributes (skin tone, hair, eyes, facial hair, glasses) noticeably improves results
• The included workflow has a Prompt Enhancer that uses Gemma-3 to auto-enrich your prompt with identity attributes

ref_t2v: A light-skinned man with long dark-brown hair and narrow rectangular
metal-frame glasses is folding clothes in a laundry room, medium-wide shot.

Best Reference Image

This model was trained on face-focused references:

• Close-up / bust crop — roughly chest-up, face large and clearly visible
• Frontal or near-frontal (slight 3/4 angles fine)
• Single subject, centered
• No white background required but clean, well-lit crop helps

Training Details

Links

• LTX-Best-Face-ID on HuggingFace
• ComfyUI-BFSNodes (Required)
• Alissonerdx on HuggingFace
• TASS-RoPE / ST-DRC Paper