One Photo + Voice Recording: A Local Workflow for Identity-Locked Talking Video Generation

From a Static Photo to a Talking Video

A developer recently shared on Reddit a local workflow built on LTX-2.3 that achieves something quite imaginative: input a single static photo, output an identity-locked talking video. The entire process requires no face swap, no driving video — just a reference image and a text prompt (or your own voice recording).

reddit source

What makes this approach stand out is its "identity-locked" characteristic — the generated face strictly matches the person in the reference photo, maintaining consistent appearance and voice across multiple generations. More importantly, it runs entirely locally; the author demonstrated it on an M5 Pro laptop.

Core Technology: Face-ID and Joint Audio-Video Denoising

Two Voice Generation Modes

The workflow includes a key toggle that offers two distinct voice generation approaches:

Lip-sync mode (toggle ON): The user's voice recording is "frozen" into the audio latent space, then through joint audio-video denoising, the generated face speaks in sync with this locked audio. Lip movements precisely synchronize with the recording, and the voice remains highly consistent across different generation runs.
Model-generated voice mode (toggle OFF): No external recording is used — the model synthesizes speech automatically based on the prompt. Simply include a He says: "..." phrase in your prompt, and the model will "invent" a voice and speak the specified line.

Traditional lip-sync approaches (like Wav2Lip) use a two-stage pipeline: generate the video first, then apply lip-sync post-processing with a separate model. This decoupled approach often results in unnatural lip and facial dynamics. Joint audio-video denoising, by contrast, processes audio latents and video latents in parallel within the same diffusion process. The audio signal acts as a conditioning constraint that directly participates in every denoising iteration, causing facial movement, head pose, and speech rhythm to form a physically coupled relationship at the generation stage — significantly improving synchronization naturalness. The elegance of this design is that the author achieved real voice injection into the generation pipeline using just a "4-node audio trick," avoiding the complex pipelines of traditional face-swap or lip-transfer approaches.

Underlying Open-Source Components

The entire solution is built on the following open-source components:

Lightricks LTX-2.3: The core video generation model. LTX-2.3 is based on a Diffusion Transformer (DiT) architecture, which outperforms earlier U-Net-based video generation models in temporal consistency and detail preservation, while being easier to quantize and compress — designed specifically for efficient local inference.
Alissonerdx's Best-Face-ID LoRA: Responsible for locking and maintaining facial identity. LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning technique that injects low-rank decomposition matrices alongside the original model weight matrices to achieve targeted capability enhancement. The Face-ID LoRA is specifically fine-tuned for facial identity preservation: during the diffusion sampling process, it continuously reinforces the facial feature embeddings from the reference image, ensuring consistent facial structure, skin tone, and proportions across different frames.
BFSNodes: Provides custom node support required by the workflow.

The solution is also compatible with the Eros model, giving it reasonable versatility.

Cross-Platform Deployment: CUDA and Apple Silicon Support

For users focused on local deployment, this workflow provides two separate pipelines for CUDA (NVIDIA) and Apple Silicon (Mac), covering mainstream consumer hardware.

Real-world performance benchmarks:

NVIDIA GPU: Generating a 4-second video clip takes under 2 minutes;
Apple Silicon: Running via the GGUF Q4 quantized version, M-series chips generate the same 4-second clip in approximately 10 minutes.

GGUF is a model serialization format introduced by the llama.cpp project, supporting multiple quantization levels such as Q4 and Q8. Q4 quantization compresses model weights from 16-bit floating point to 4-bit integers, reducing VRAM usage by approximately 75% — enabling large video models that would normally require high-end GPUs to run on Apple Silicon's unified memory architecture. The M-series chips' CPU and GPU sharing a memory pool makes them more flexible when handling quantized models compared to traditional NVIDIA discrete GPU architectures, though generation speed still lags by several times due to lower GPU compute. While Mac performance is noticeably slower than dedicated graphics cards, the ability to run identity-locked talking video generation entirely locally on a consumer laptop reflects the maturity of model quantization and local inference ecosystems. For users who value privacy and don't want to upload personal photos and voice recordings to the cloud, this is a pragmatic choice.

Usage Tips: Reference Image and Prompt Are Everything

The author specifically highlighted several key factors affecting generation quality — highly useful for practical use.

Choosing a Reference Image

The quality of the reference image "makes a huge difference." The author recommends:

A tight frontal chest-up crop;
The face should take up a large portion of the frame.

The clearer and more centered the face, the better the identity-locking effect. This follows the general pattern for Face-ID approaches — the more complete the reference features, the more accurately the model can reconstruct the identity.

Prompt-Driven Identity

Another easily overlooked detail: identity is largely driven by the prompt. The author notes that with cfg 1 settings, you need to use the ref_t2v: prefix in the prompt to describe the person's characteristics. A reference image alone is not enough — you also need to accurately describe the target person in text. The combination of both achieves the best identity consistency.

Technical Value and Ethical Considerations

This workflow demonstrates an important trend in the open-source video generation space: moving beyond pure text-to-video toward controllable, identity-lockable multimodal generation. The ability to inject real voice into the generation pipeline allows content creators to produce talking-head videos using their own voice and likeness, with applicable scenarios ranging from AI digital avatars and educational content to personalized short videos.

The joint audio-video denoising approach also deserves attention — rather than treating lip synchronization as a post-processing step, it allows audio and video to mutually constrain each other during the denoising process, achieving more natural synchronization.

Of course, the widespread availability of such technology comes with ethical risks that cannot be ignored. The ability to generate highly realistic talking videos from just one photo and a voice recording undeniably lowers the barrier to creating deepfakes. How to advance creative tools while establishing effective provenance tracking and abuse-prevention mechanisms will be a challenge the entire industry must continue to address.

For readers who want to try it hands-on, the author has published the complete workflow, examples, and README documentation on GitHub, with workflow files also available on CivitAI.