Testing AI image-to-HDR: disabling diffusion cache removes banding; 16-bit barely matters
robertwellesley · reddit · 2026-10-19
The author systematically tested optimal HDR video output from AI-generated images: generating with Z-Image Turbo through a regular VAE and an HDR VAE, saving each as 8-bit and 16-bit PNGs, then converting SDR→HDR and pushing contrast in DaVinci to observe banding.
Key findings:
- 8-bit vs 16-bit PNG made almost no visible difference; the scope barely moved
- All four outputs showed banding regardless of VAE; the HDR VAE only tolerated harder pushes before banding appeared
- Counterintuitive: turning off all caching in Z-Image Turbo (including Sage Attention and Easy Cache) eliminated banding entirely — caching was corrupting the generated data
Best workflow: generate without caching + HDR VAE + final SDR→HDR conversion, yielding completely band-free output with a wider, more robust dynamic range.
More from Multimodal
- MiniMax-H3 hailed as favorite open release of the year, runs fully local — AdinaYakup · 2026-10-23
- Michal Malewicz experiments with depth maps in his creative workflow — michalmalewicz · 2026-10-23
- One-Shot Tunable Style Transfer on Flux3 Via Reference Latent + Node — Winter_unmuted · 2026-10-23
- PhysVidBench accepted at EMNLP 2026: do video gen models grasp physical commonsense? — aykuterdemml · 2026-10-23
- EMNLP 2026 paper teaches multimodal LLMs to get cartoons via incongruity-resolution supervision — aykuterdemml · 2026-10-23
- A gorgeous Midjourney style: sref code 820221253 for dreamy lavender watercolor looks — umesh_ai · 2026-10-23