What does the CelebVHQ ID-LoRA do here?

It helps LTX 2.3 preserve the subject’s identity and facial features across frames, especially in dynamic, in-the-wild shots. It works best with a clear, frontal image and a well-structured prompt that separates [VISUAL], [SPEECH], and [SOUNDS].

How is lip-sync achieved without a separate lip-sync model?

The workflow aligns perceived mouth movements by making the [SPEECH] text explicit and pairing it with matching TTS audio. While not frame-perfect lip-sync, using concise lines, clear punctuation, and expression cues improves alignment. You can retime in post if you need exact phoneme matching.

Can I use a different ID-LoRA (e.g., TalkVid) or my own?

Yes. Swap the LoRA in the LTX 2.3 node. CelebVHQ is tuned for in-the-wild motion; TalkVid is typically stronger for close-up, studio-like talking head shots. For custom LoRAs, keep the same structured prompt format and test different strengths.

Do I need the LTX x2 upscaler model?

It’s optional. If you place ltx-2.3-spatial-upscaler-x2-1.1.safetensors in latent_upscale_models, the workflow can render crisper outputs at higher resolution. If VRAM is limited, disable upscaling or render at a lower base resolution.

LTX 2.3_ID_lora_CelebVHQ - ComfyUI Workflow

This ComfyUI tutorial workflow turns a single image into a short, voice-driven video with strong visual identity and perceived lip alignment. It combines Lightricks LTX 2.3 for video generation with an ID-LoRA (CelebVHQ) to keep the subject’s face and styling consistent across frames, while Qwen 3 TTS produces a matching voice track. The pipeline is organized into clear groups: Load Image, Create Prompts, Prompt Extraction, Create Voice, and Create Video.

Under the hood, GeminiNode analyzes your uploaded image (via LoadImage) and drafts two complementary texts: a speech line with expressive cues and a scene/appearance description. RegexExtract then parses these into an ID-LoRA-ready format with three tagged sections: [VISUAL], [SPEECH], and [SOUNDS]. FB_Qwen3TTSCustomVoice uses the [SPEECH] and style hints from [SOUNDS] to synthesize a voice track, saved with SaveAudioMP3. The LTX 2.3 video node (the custom node in this graph) consumes the structured prompt plus your reference image to render a talking video, with the CelebVHQ ID-LoRA improving identity retention in dynamic, in-the-wild scenarios. PreviewAny lets you iterate quickly, and SaveVideo writes the final clip. Local users can optionally load the LTX spatial upscaler (ltx-2.3-spatial-upscaler-x2-1.1.safetensors) for cleaner details at higher resolution.

FAQ

Frequently Asked Questions

View all workflows