Descriptions:
Nerdy Rodent tests the newly released LTX 2.3 video generation model across multiple configurations in ComfyUI, using a deliberately challenging and consistent test prompt throughout — a woman shifting timelines while a dolphin transforms into an orca and a hexagonal window morphs into a corridor — to expose meaningful differences between variants under identical conditions. Checkpoints tested include the full 22B dev BF16, the official FP8 version, and community files from Ki with split components and BF16 VAE.
Key comparisons cover the BF16 vs FP8 VRAM trade-off, the impact of enabling or disabling first-and-last-image anchoring on prompt adherence, and the difference between the new 1.5x and 2x upscaler options introduced in version 2.3. The video also demonstrates integrating the LTX2 AI prompter for automatic prompt enhancement using both the default clip-based method and a local Ollama instance running Qwen 3VL — the latter notably fixing spelling inconsistencies (such as “armore” vs “armor”) that appear in the default enhancer’s output. A final test at 1920×1080 using the chunk feed-forward node illustrates how to manage VRAM on consumer GPUs at higher resolutions by adjusting the chunk count.
LTX 2.3 is available in multiple file sizes suitable for a range of modern home hardware. The overall assessment is that the model handles stylized prompts well but still struggles with complex multi-element transitions, with prompt wording having an outsized effect on final output quality.
📺 Source: Nerdy Rodent · Published March 06, 2026
🏷️ Format: Tutorial Demo
