Comparative Analysis of Euler Ancestral and Res Multistep Samplers in NVIDIA Cosmos for Text-to-Video Generation

This report presents a comparative study of two samplers, Euler ancestral and Res multistep, within the NVIDIA Cosmos diffusion model for text-to-video generation. Under fixed generation conditions (constant diffusion steps, classifier-free guidance scale, resolution, video length, and frame rate) and using identical positive/negative prompts, an equal number of videos are generated per sampler. Quality is assessed using the Peak Signal-to-Noise Ratio (PSNR), the Structural Similarity Index (SSIM) and Video Multimethod Assessment Fusion (VMAF). The results reveal that although both samplers yield low pixel-level fidelity and nearly zero perceptual quality scores, the Res multistep sampler achieves markedly higher structural similarity, indicating superior structural fidelity. These findings are discussed in the context of prior work on diffusion models and NVIDIA Cosmos, and it is concluded that Res multistep is preferable when structural coherence is critical.