An identity-preserving reference-to-video LoRA for LTX-2.3 (22B). Give it a reference
photo of a person + a text prompt, and it generates a video that keeps that person's identity.
Built with overlap reference conditioning + TASS-RoPE (source-phase RoPE) and a differentiable
ArcFace identity loss. Runs in ComfyUI via the companion BFS Nodes.
Status: this is the first (overlap/TASS-RoPE) model. Other experimental variants
(native Gemma-vision conditioning, timestep-split texture injection) may be released later if
they prove out.
🎬 Examples
With 2 Pass (Upscaled)
What it does
Reference-to-video (ref_t2v): one reference image → video of that identity performing the
prompt's action.
Identity is injected by placing the reference latent in the target's frame-0 RoPE grid
(overlap) and tagging it with a distinct source phase so the model knows it is a reference,
not the first frame to generate.
An auxiliary ArcFace face-similarity loss on the decoded prediction sharpens the identity.
How it works (technique)
Overlap reference + TASS-RoPE (source-phase)
The reference latent is concatenated to the video sequence sharing the frame-0 grid (classic
IC-LoRA "overlap"). To stop the reference from leaking into / being confused with the generated
first frame, each source gets a distinct multiplicative RoPE phase:
This "source tag" lets the model separate who is who in the sequence and strongly improves
identity transfer. Because the tag is positional, the same mechanism generalizes to multiple
references (source_id = 2, 3, 4, …) for multi-subject conditioning.
TASS-RoPE (Temporal-Adjacent Spatial-Shifted RoPE) is from ST-DRC, Chen et al.,
Spatial-Temporal Decoupled Reference Conditioning for Identity-Preserving Text-to-Video
Generation — arXiv:2606.02441. This LoRA uses an
overlap + per-source RoPE-phase variant of that idea.
ArcFace identity loss
During training the flow-matching prediction is decoded to pixels, the face is aligned and passed
through ArcFace (buffalo_l w600k_r50), and a cosine identity loss pulls the generated face toward
the reference embedding — the same recognition space used to evaluate identity.
This model was trained on face-focused reference images. Looking at the training data, every reference follows the same pattern:
Close-up / bust crop — roughly chest-up, with the face large and clearly visible.
A single subject, centered in the frame.
Frontal or near-frontal (slight 3/4 angles are fine, but the face is always well exposed).
Tightly cropped on the face/upper body — the training crops are about 460×406 (nearly square).
Recommendation: for the best identity transfer, give the model a reference that looks like this — crop your image to the face / upper body. Full-body shots, wide shots, or images where the face is small and far away work noticeably worse, because the model never saw references like that during training.
In short: a clean, frontal, well-lit close-up of the face is the ideal reference.
Captions (IMPORTANT)
This model was trained on captions in the ref_t2v: format that describe the scene and the
person's action (not just "a person"). Identity in LTX is strongly prompt-driven — the more the
prompt describes, the better the identity holds. For best results:
Prefix prompts with ref_t2v:.
Describe the action, setting, framing, camera in present-progressive.
Describing identity attributes of the reference person (skin tone, hair, eyes, facial hair,
glasses, face shape) noticeably improves the result.
The shared ComfyUI workflow includes a Prompt Enhancer that looks at the reference image (via
the model's own multimodal Gemma-3 text encoder) and automatically enriches your prompt with the
person's identity attributes — so you don't have to describe them by hand.
Example (enhanced):
ref_t2v: A light-skinned man with long dark-brown hair past his shoulders, narrow rectangular
metal-frame glasses, light blue-gray eyes and light stubble is folding clothes in a laundry room,
medium-wide shot. He places the folded clothes into a basket, then carries it across the room.
Neutral indoor lighting, simple domestic environment.
Install ComfyUI-BFSNodes (ComfyUI Manager → Install via Git URL, or clone into custom_nodes/).
Dependencies (insightface, transformers, …) install automatically.
Load LTX-2 (checkpoint + the Gemma-3 text encoder / CLIP) as usual.
Add the LTX Identity Transfer (overlap + source-phase) node; feed it the reference image.
Load this LoRA on the MODEL path.
Use the shared workflow (it wires the Prompt Enhancer → identity node → sampler).
A ready-made workflow JSON is included alongside the model.
About the ArcFace projector (optional — you don't need it)
The identity node has an optional ArcFace projector dropdown (it projects a face embedding into
extra conditioning tokens). You can leave it on None. In our experiments the projector's
effect was marginal — identity is carried almost entirely by the overlap reference latent (and
the prompt), so the projector adds very little on top and is not required for this model. It is kept
only as an optional extra; skipping it keeps the graph simpler with no meaningful quality loss.
Tips for best results
Use a clean, frontal, well-lit close-up as the reference — identity quality tracks reference
quality closely.
Let the Prompt Enhancer add the identity attributes (or write them yourself).
Keep the reference framing simple; busy/occluded reference faces degrade identity.
Limitations
First-frame / "mask" tendency: because identity comes from the raw reference latent, the
reference appearance (and framing / accessories like glasses) can partially copy into the video —
it can look like the reference is "pasted"/masked rather than freshly generated. Using a clean
frontal reference and the trim-first-frame node mitigates this.
Prompt dependence: identity is much stronger when the prompt describes the person. A bare,
generic prompt gives weaker identity — use the Prompt Enhancer.
Close-up bias: the training data skews to close-up / frontal faces, so full-body or
large-angle shots hold identity less well.
Metric caveat: ArcFace similarity is unreliable on small / turned / occluded faces — judge
visually, not only by score.
Citation
If you use this model, please cite the base model and the technique:
1@article{chen2026stdrc,
2 title = {Spatial-Temporal Decoupled Reference Conditioning for Identity-Preserving Text-to-Video Generation},
3 author = {Chen, Yuheng and Hu, Teng and Wang, Yuji and He, Qingdong and Ma, Lizhuang and Zhang, Jiangning},
4 journal = {arXiv preprint arXiv:2606.02441},
5 year = {2026}
6}
Identity loss: ArcFace (Deng et al., 2019) via InsightFace buffalo_l.
