Descriptions:
LTX 2.3 from Lightricks is the latest milestone in open-source video generation, upgrading from 19B to 22B parameters with improved vertical video support, cleaner audio, more stable image-to-video conversion, and sharper in-video text rendering. The Veteran AI channel presents three ComfyUI workflows designed for different VRAM budgets, each tested with identical prompts, seeds, and reference images for a fair comparison.
The first and most detailed workflow uses the official FP8 version of LTX 2.3 (approximately 30GB versus the full 46GB model) paired with a distilled LoRA at weight 0.5. The architecture is complex: it uses Gemma 3 as one text encoder alongside a dedicated LTX 2.3 encoder, requires separate video and audio VAEs, and runs a two-stage sampling pipeline — generating video at half the target resolution before spatial upscaling with either a 1.5x or 2x upscaler model. A temporal upscaler for frame interpolation is also available. This workflow completed in 2 minutes 40 seconds but demands nearly 30GB of VRAM. For lower-VRAM setups, a second workflow by community creator Kijai is covered, along with a third pure-distilled-model path using CFG 1.0 and just 8 sampling steps.
Important quirks are flagged throughout: LTX 2.3 requires an input image even for text-to-video generation, frame counts must follow a “multiple of 8 plus 1” rule, and the Checkpoint Loader (not the standard Diffusion Model Loader) must be used since the model file bundles the video VAE internally. The standardized test format lets viewers directly compare output quality and speed across all three approaches.
📺 Source: Veteran AI · Published March 09, 2026
🏷️ Format: Tutorial Demo
