Summary
This video from Veteran AI provides a thorough hands-on evaluation of Best Face ID, a character-consistency LoRA designed for the LTX 2.3 video generation model, testing it across four distinct ComfyUI workflows to answer practical questions about when and how to use it.
The tutorial explains the two core problems Best Face ID addresses: preserving a character’s facial features across new scenes using a single reference image, and supporting Character Sheets — multi-view reference composites that add hairstyle, clothing, body shape, and back/side-view consistency. The host walks through the full ComfyUI setup, including required extensions (ComfyUI-BFSNodes), key parameters like source_id, phase_scale, and the critically important ref_resize_mode setting, and the specific four-layer prompt structure the model requires (starting with the ref_t2v token).
A central finding from the comparison tests is that aspect ratio pairing matters significantly: mismatching a vertical reference image with a horizontal output video noticeably degrades facial similarity, while matching orientations — or defaulting to 1024×1024 square output — produces the most stable results. The Character Sheet workflow, which requires a horizontal four-column image (face close-up, front body, 90-degree side, and back view on a white background) and the native_resolution resize mode, is shown to produce stronger full-body consistency beyond just face matching. The video also covers post-generation tiled upscaling for high-resolution output and the edge cases where problems persist even after a clean generation.
📺 Source: Veteran AI · Published July 29, 2026
🏷️ Format: Benchmark Test
