Full open-source pipeline for realistic 4K video: 3.7MP base, tiled latent pass, DLSS 5
Grinderius · reddit · 2026-11-23
A Reddit user shares what they claim is the most realistic open-source 4K video pipeline (RTX 5090), built on the principle of never doing everything in one pass:
- Step 1 — Native 3.7MP base gen: generate directly at 2560x1440 (or 2880x1280 for 21:9) with an INT8 pruned MiniMax H3 checkpoint, euler/simple, 20 steps. Skip '8k hyperrealistic' prompt spam — describe normal lens/lighting and texture emerges naturally.
- Step 2 — 4K tiled latent pass: Lanczos upscale to 3840x2160, then a light 3-step V2V pass with the same INT8 H3 + Open-H3-Regenerate-2K-IC (must be the In-Context version, 0.85-1.0) to keep faces locked. Enable Tiled VAE (512/64) and Tiled Diffusion (1024x1024/256) to avoid seams. Crucially use ManualSigmas 0.62/0.48/0.30/0 instead of the standard 0.909 start — 0.909 ruins faces when upscaling from a clean 3.7MP clip, while 0.62 adds pores, hair and fabric grain without breaking the acting.
- Step 3 — DLSS 5 polish: ComfyUI-DLSS5-Enhancer at 1x DLAA, NR Intensity 0.80, Local/Skin Structure Strength 2.00, auto mask on, scene change threshold 0.24 to kill AI micro-shimmer.
Full settings and reasoning provided for every parameter.
More from Multimodal
- Dev builds interactive 3D human eye with Gemini Neon to teach how vision works — Scobleizer · 2026-11-23
- Princeton study: diffusion models barely extrapolate basic visual skills, ~50% on counts — NielsRogge · 2026-11-23
- AI drops Goku into real life with convincing physics and lighting, blurring what's real — heypearlai · 2026-11-23
- FaceRefine Plus v0.2.0 speeds up MiniMax H3 video face refining by 2.7x — AnniCodex · 2026-11-23
- Anima Artifacts demos a 'living magazine' you converse with instead of read — alexcovo_eth · 2026-11-23
- Turning real F1 telemetry into an AI video: 5 nights of failed experiments — BittiAI · 2026-11-23