H3 image comparisons · 2026-11-03
Tests from the V100 server: the PrunaAI p-h3-image LoRA on H3, then edit speed and quality for Qwen 2.1 base, H3 chroma and H3 Pruna. Click a picture on either page to open the viewer: ← / → flips between versions, Z or click toggles 1:1 pixels. Pictures are WebP at quality 95.
Outcome: Pruna came out slightly better than H3 chroma, but only barely and not on every case. It was also a little faster, mostly because the 32 GB of RAM makes chroma re-read about 9.5 GiB from the SSD on every job. Pruna was removed. The bot stays on chroma (h3chroma), and /edit keeps Qwen 2.1 base.
Pages
p-h3-image LoRA test
Pruna's LoRA on H3: LoRA only vs with Pruna's encoder and VAE, Pruna VAE vs SOL VAE, against H3 Pika and Pika + chroma. Text to image and edits at 2 MP, seed 42.
Edit speed test
4 edits on Qwen 2.1 base, H3 chroma and H3 Pruna at 1, 2 and 4 MP, with load and diffusion times and RAM→VRAM traffic.
Contact sheets (edit speed test)
