Descriptions:
This tutorial from the Veteran AI channel tackles two common frustrations in AI video generation with LTX Video 2.3: character identity drift over time and resolution ceilings around 720p. The presenter demonstrates a ComfyUI-based two-stage workflow using custom nodes that address both problems without requiring a new model download or LoRA — only a cloned repository added to the custom_nodes folder.
For identity preservation, the central node is LTX Likeness Guide, which converts a reference portrait into a persistent identity anchor at the latent and conditioning levels — maintaining the constraint across all frames rather than just seeding the first one. A companion node, LTXV Img To Video Inplace KJ from the KJ Nodes extension, initializes the latent from a reference-tied state instead of pure noise. Three key parameters are highlighted: `silent_reference` (face identity only, without forcing composition), `bbox_only` (restricting to the facial bounding box to eliminate background interference), and manual face region override for cases where automatic detection drifts. A second identity anchor node continuously re-reads the reference info signal throughout sampling, countering the model’s tendency to revert toward its training distribution — producing wider cheekbones, shiny skin, and racially homogenizing facial features in vanilla generation.
For resolution, the LTX Latent Upsampler Tiled node processes the upscale in tiled mode, achieving 1920×1024 and 1024×1920 outputs that direct high-resolution generation cannot reach on typical consumer GPUs. The full workflow is also available on RunningHub for cloud-based ComfyUI access.
📺 Source: Veteran AI · Published July 10, 2026
🏷️ Format: Tutorial Demo
