LTX-Video 2.3 Benchmark: 11 Models Tested for AI Video Generation

LTX-Video 2.3 Models Reveal New Frontiers in AI Video Generation

LTX-Video 2.3 is rapidly becoming a benchmark for open-source AI video generation, with a comprehensive test of 11 models revealing critical insights into performance, memory usage, and workflow efficiency. According to a detailed community analysis posted on Reddit, the latest iteration from Lightricks and its ecosystem of contributors demonstrates significant advancements in latent space video synthesis, particularly when paired with ComfyUI and optimized transformer architectures.

Model Variants and Hardware Performance Trade-offs

The tested models span three primary sources: Lightricks’ official releases, Kijai’s ComfyUI-optimized variants, and unsloth’s GGUF quantized versions. The full-size 22B-parameter models, such as ltx-2.3-22b-dev.safetensors (43GB), deliver the highest visual fidelity but demand substantial VRAM—making them inaccessible to many users. In contrast, the FP8 and GGUF quantized versions reduce memory footprints to under 22GB while retaining usable output quality, enabling deployment on mid-tier GPUs.

Notably, the GGUF models exhibited unexpected performance degradation during upscaling, with iteration times inflating significantly. This anomaly suggests potential inefficiencies in the quantization pipeline or compatibility issues with the latent upscaler component. Meanwhile, distilled models—designed for faster inference—achieved comparable results at half the sampling steps (15 vs. 35), making them ideal for rapid prototyping and batch generation.

Additional components, including text encoders like gemma_3_12B_it_fpmixed (12.8GB) and specialized VAEs for audio and video, further underscore the modular complexity of the system. LoRAs such as the 1.1GB ID adapter (ltx-2.3-id-lora-celebvhq-3k.safetensors) enable consistent character preservation across frames, a crucial feature for narrative video generation.

The workflow optimization by community contributor princepainter, using a dual KSampler approach in ComfyUI-PainterLTXV2, emerged as the most practical solution. Despite the availability of official workflows from Lightricks and RuneXX, users reported them as overly complex. The streamlined setup reduced cognitive load and improved reproducibility, highlighting a growing trend toward user-centric tooling in the AI video space.

Benchmark results at 1280x720 resolution and 10-second duration revealed that distilled-FP8 models delivered the best balance of speed and quality. When paired with the spatial upscaler (ltx-2.3-spatial-upscaler-x2-1.1.safetensors), they achieved near-4K output (1920x1080) with minimal artifacts. However, the 15.93GB VRAM of the RTX 5060 Ti—despite its high specification—was still pushed to its limits during concurrent model loading, suggesting that future iterations may benefit from dynamic memory offloading or multi-GPU support.

One standout innovation was the developer’s custom Aligned Text Overlay Video node, which dynamically injects prompt metadata into generated videos. This tool, now open-sourced on GitHub, addresses a critical gap in production pipelines by automating captioning and version tracking—features previously handled manually.

LTX-Video 2.3 is not just a technical leap—it’s reshaping how creators approach AI video generation. From quantized GGUF models enabling desktop deployment to LoRAs preserving identity across sequences, the ecosystem is maturing rapidly. The community’s emphasis on transparency, reproducible workflows, and accessible tooling signals a shift away from closed black boxes toward open, collaborative innovation.

As LTX-Video 2.3 continues to evolve, its 11 model variants provide a roadmap for users to balance quality, speed, and hardware constraints—making AI video generation more accessible than ever before. The future of synthetic video lies not in one monolithic model, but in a flexible, modular ecosystem like this one.