Summary
Veteran AI delivers a complete, reproducible workflow for LTX 2.3’s Ingredients LoRA — an IC-LoRA for reference-based video generation that can maintain consistent characters, props, and scenes from a single composite reference image. The video goes beyond the official documentation to address a real pain point: naively combining images and standard prompts produces mixed-up characters, misplaced props, and model improvisation. The solution is a disciplined reference sheet layout and a two-part Markdown prompt.
The full ComfyUI pipeline is covered step by step on the RunningHub platform: loading the LTX-2.3-22B checkpoint, an acceleration LoRA (CFG set to 1 for single-stage sampling), and the Ingredients LoRA at weight 1.0. Key technical details include resizing the reference sheet’s shorter side to 512px — which sets the final output resolution — and duplicating it to 241 frames for approximately 10 seconds of video. Two critical nodes are explained: LTXV Img To Video Condition Only and LTX Add Video IC LoRA Guide.
The tutorial also provides two recommended reference sheet layouts: a single-character template placing face close-ups upper-left, full-body views upper-right, props lower-left, and a clean scene lower-right; and a two-character variant that gives each character their own reference image in the upper half. Results shown include two distinct characters generated in the same shot, demonstrating strong identity consistency when the workflow is followed correctly.
📺 Source: Veteran AI · Published June 25, 2026
🏷️ Format: Tutorial Demo
