Do More Steps Improve LTX 2.5? 8 vs 20 vs 30 on RTX 5090

Do more sampling steps actually improve LTX 2.5 video quality? This controlled local ComfyUI comparison tests 8, 20 and 30 first-stage sampling steps across three difficult image-to-video scenes: molten glass and hands, fast fur and snow, and fine loom threads with mechanical motion. Test setup: • LTX 2.5 22B Distilled Transformer, INT8 ConvRot • Gemma 4 12B text encoder, INT8 ConvRot • No LoRA and no prompt rewriter • Euler ancestral, CFG 1.0 • Stage 1: 8 vs 20 vs 30 steps • Stage 2 spatial upscaler: fixed at 3 steps • 960×544 latent generation → 1920×1088 output • 121 frames, 5.04 seconds, 24 FPS • NVIDIA GeForce RTX 5090 32 GB • Peak whole-GPU use: approximately 31.0 GiB Measured loom workflow times: • Cold 8-step run: 291.6 seconds • Warm 20-step run: 79.8 seconds • Warm 30-step run: 89.0 seconds These are end-to-end workflow measurements including model loading, offload, sampling, VAE decode and saving. They are not isolated sampler-speed benchmarks. Conclusion: more steps change motion trajectories, structure and particles, but they do not act as a guaranteed quality slider. Judge the complete moving result rather than a single frame. The video and narration were produced with generative-AI tools and edited locally. No sponsored placement. #LTXVideo #ComfyUI #RTX5090 #AIVideo #LocalAI