ID-LoRA¶

Development line: project:id-lora · thread id-lora-development
Last event: 2026-03-15 · 1 dated since 2026-03-15 · Researched: 2026-09-05 · confidence: high

What it is¶

ID-LoRA is a zero-shot identity-conditioning method for LTX audio-video diffusion, built for creators who need talking-person video without training on each speaker.

• Voice identity transfer from reference audio
• Visual identity conditioning through a first-frame image
• Unified control of scene, spoken words, vocal style, and ambient sound in one prompt

Development line¶

• 2026-03-15 — ID-LoRA public project resources linked. On 2026-03-15, ID-LoRA published links to its project site, source repository, and model profiles. These links point readers to primary resources, but establish no specific release, version, capability, or license.

What changed¶

2026-03-15 — ID-LoRA appeared as a research project for unified identity-conditioned audio-video generation on the LTX-2 backbone, linking project, code, and model destinations on 2026-03-15. The initial arXiv paper from 2026-03-10 introduces Identity-Driven In-Context LoRA, negative RoPE temporal positions for reference tokens, and identity guidance. It reports a 73% preference over Kling 2.6 Pro for voice similarity, a 24% cross-environment speaker-similarity gain, and about 3,000 training pairs on one GPU.

How to use this¶

As of 2026-03-15, use the linked website and repository to explore ID-LoRA, but verify the version, license, and model availability before adoption.

1. Clone the official repository and install locked dependencies with uv sync --frozen; use Python 3.11+ and CUDA 12.x. — https://github.com/ID-LoRA/ID-LoRA/tree/main
2. Download required base-model components and the ID-LoRA checkpoint with the repository download script. — https://github.com/ID-LoRA/ID-LoRA/tree/main
3. Provide a first-frame image, a roughly five-second reference-audio file, and a prompt split into [VISUAL], [SPEECH], and [SOUNDS] sections. — https://github.com/ID-LoRA/ID-LoRA/tree/main
4. For current LTX-2.3 use, switch the uv workspace to ID-LoRA-2.3/packages/*, then run the two-stage HQ script with the LTX-2.3 checkpoint. — https://github.com/ID-LoRA/ID-LoRA/blob/main/ID-LoRA-2.3/README.md

Best practices¶

• Use reference voice audio close to five seconds; shorter or longer clips can reduce speaker-identity transfer. — https://github.com/ID-LoRA/ID-LoRA/tree/main
• Put the exact desired transcript in [SPEECH]; place speaking style and environmental sound in [SOUNDS], not in the visual description. — https://github.com/ID-LoRA/ID-LoRA/tree/main
• Use two-stage HQ on LTX-2.3 for fidelity; choose one-stage with quantization when VRAM or turnaround time is the constraint. — https://github.com/ID-LoRA/ID-LoRA/blob/main/ID-LoRA-2.3/README.md
• Use only reference image and voice material for which you have authorization. — https://github.com/ID-LoRA/ID-LoRA/tree/main

Superseded by this¶

• 2026-03-24 — native upstream ComfyUI support for LTX2 reference-audio identity transfer partly supersedes the separate ID-LoRA LTX-2.3 ComfyUI custom-node route; original weights need no conversion.

Still unknown¶

• We could not retrieve the supplied hf.ru short link, so we do not treat it as evidence.
• Public sources provide no first-party publication date for the LTX-2.3 checkpoint release, so we do not assign it a dated development event.

Sources¶

Agent brief¶

• Subject: project:id-lora, thread id-lora-development, 1 dated events 2026-03-15 → 2026-03-15.
• Practical note: As of 2026-03-15, practitioners can use the linked website and source repository as ID-LoRA discovery points, while independently verifying the applicable version, license, and model availability before adoption.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.