Descriptions:
The Nerdy Rodent channel delivers a hands-on comparison of four LTX Video 2.3 model variants running in ComfyUI, benchmarked under identical conditions to give viewers an honest picture of the tradeoffs between quality, speed, and resource requirements. Models tested include the community-built V3 checkpoint from Kijy (which retains BF16 precision in the first and last transformer blocks for improved audio quality), Lightricks’ official distilled FP8 checkpoint, the Unsloth Q4 GGUF quantization, and NVIDIA’s new NVFP4 version of the dev model. Every run uses the same prompt, seed, and output settings—121 frames at 1280×740 resolution—on an RTX 3090, with stage-by-stage generation times recorded.
The FP8 distilled checkpoint and Kijy’s V3 perform similarly in speed (roughly 18–19 seconds for stage one, 25–30 seconds for stage two), while the GGUF Q4 runs significantly slower at 44 seconds per stage despite a smaller file size. The NVFP4 variant offers faster generation with lower VRAM demands, though a distilled version has not yet been released. The video also explores sampler configuration—CFGP sampler variants and custom sigma schedules for stage two—showing how dramatically output changes even with a fixed seed.
The second half covers face swapping in LTX 2.3 using LoRA files within ComfyUI, walking through the additional nodes required and explaining why the video-describer step was omitted in favor of a direct OLLAMA integration. The tutorial is aimed at ComfyUI users already comfortable with node-based workflows who want to evaluate the new model options before committing to a download.
📺 Source: Nerdy Rodent · Published March 22, 2026
🏷️ Format: Benchmark Test
