Releasing Many New Inferencing Improvement Nodes Focused on LTX2.3 - comfyui-zld

Developer Z-L-D has launched a significant update to the comfyui-zld extension, introducing a suite of new inference nodes designed to solve persistent quality issues in the Lightricks LTX2.3 video generation model. The release, culminating months of research, specifically targets LTX2.3's 'linearity collapse'—where rapidly moving vertical lines degrade into a squiggly mess—and its general noise blur. Key nodes include EMAG (based on Yadav et al., 2025), FDTG, and the SA-RF-Solver, which work together to minimize these artifacts. The developer provides four pre-configured 'LTX Cinema' workflows (High, Medium, Low, Fast) that balance output quality against generation time, which ranges from ~6 to ~25 minutes for a 5-second clip on an RTX 3090.

The most immediately impactful nodes are EMAG, FDTG, and SA-RF-Solver, with EMASync building on EMAG for higher quality at a greater time cost. The workflows detail specific configurations for guiders and samplers across different generation stages (S2, S3/S4), allowing users to choose between advanced techniques like SyncCFG and simpler CFG guidance. This release represents applied research, integrating several recent academic techniques (like RF-Solver and SA-Solver) into a practical, usable toolset for the ComfyUI community, significantly improving the stability and clarity of videos generated with the powerful but artifact-prone LTX2.3 model.