**Project Overview** I’m building a high-performance video upscaling pipeline (2× and 4×) where final quality is evaluated by PieAPP on a random sample of frames (10 frames per video). Lower PieAPP is better. I need an expert who can improve both: 1. The upscaling model/training strategy for perceptual quality, and 2. The FFmpeg HEVC encoding strategy (libx265 and/or hevc_nvenc), especially adaptive CRF / maxrate selection per video. ** Input / Output Characteristics ** - Reference content is mostly 4K. - Inputs are typically downscaled versions: - 2× input: ~1080p → output 4K - 4× input: ~540p → output 4K - Video content is real-world: sports, night scenes, animals, travel, etc. - Performance target: handle ~300 frames in 20–30 seconds end-to-end (≈10–15 FPS minimum; higher is better). ** What I’m Looking For ** - Strong FFmpeg knowledge (libx265 and/or hevc_nvenc), rate-control, VBV (maxrate/bufsize), AQ/psy tuning, etc. - Experience with perceptual metrics and “optimize for the metric” workflows (PieAPP preferred; VMAF/LPIPS/SSIM also relevant). - Real experience in video SR/upscaling training + inference optimization (speed/quality tradeoffs). - Ability to create reproducible experiments and deliver “drop-in” configs/scripts. How to Apply (required) Please include: - 1–2 examples of similar work (GitHub, paper, blog, or a short write-up). - Your answers to the two screening questions below. Screening Questions (must answer) 1.Upscaling model & training: What model would you use for 2×/4× upscaling under strict speed constraints, and how would you design the loss/configs to optimize specifically for PieAPP-style perceptual error? 2.HEVC encoding & adaptive parameters: Encoding quality is not monotonic: in most cases, lower CRF (higher quality) helps PieAPP, but some clips paradoxically get worse PieAPP at very low CRF (e.g., CRF 10) while higher CRF (e.g., CRF 25) performs better—same with VBR maxrate choices. I need a reliable method to select CRF/maxrate per video (or per segment) that minimizes PieAPP. How would you reliably choose the best CRF and/or maxrate per video (or per segment) to minimize perceptual error, given that “lower CRF is better” sometimes fails? If you have any questions related to the project, I would be happy to address them.