A reinforcement-learning LoRA fine-tune for Lightricks' LTX-2.3 video model reduces artifacts and improves temporal coherence in generated clips, now available on HuggingFace.
Practitioners working with Lightricks' LTX-2.3 video model are reporting sharper motion boundaries and fewer visual glitches after applying a new reinforcement-learning LoRA released this week.
The LoRA, trained using the OmniNFT framework and released on HuggingFace by Kijai, applies reinforcement-learning fine-tuning to the 22-billion-parameter LTX-2.3 base model. OmniNFT is a research project focused on using RL to align video diffusion models with human preference feedback. The weights ship in bf16 format at roughly 1.7 GB and are designed to work with ComfyUI workflows. According to ablation studies on the OmniNFT project page, the full training pipeline shows preference-score improvements of 12–18 percent over base LTX-2.3 on human-eval benchmarks, though those numbers reflect the broader approach rather than this specific checkpoint.
Users report the best results when pairing the LoRA with tiled sampling passes after upscaling. The LTX Tiled Sampler custom node from TenStrip's 10S-Comfy-nodes repository splits high-resolution frames into overlapping tiles during denoising, reducing memory spikes and improving detail consistency at the cost of longer render times. The RL LoRA adds roughly 10 percent to inference time in typical ComfyUI setups but does not require additional VRAM beyond what the base model already demands.
LTX-2.3 is Lightricks' open-weight video synthesis model, released in November 2024 with native support for up to 257 frames at 24 fps and a 128-frame context window. The model runs locally on consumer GPUs with 24 GB VRAM or more, depending on resolution and batch size. The HuggingFace card does not list a formal license; users should assume it inherits LTX-2.3's original terms, which permit commercial use with attribution.
