Sulphur 3, the biggest full fine-tune of a known video model, is now public on HF
Diabolicor · reddit · 2026-11-21
Sulphur 3, reportedly the largest full fine-tune of a well-known video model, is now public on Hugging Face (SulphurAI/Sulphur-3), with notably better generalization and adaptability to input images and poses.
- Recommended inference: TenStrip's 10Eros-Max or LokkenJP's optimized builds, which bake Sulphur 3 as a base LoRA plus useful LoRAs; int8/bf16 pruned versions available; LCM/simple sampling at 8-12 steps with 10% VSA or 13% SLA sparsity.
- Two-pass workflow with an upscale pass remains optimal, though single-pass quality beats beta7; first pass at 0.4-0.6 MP gives better quality.
- Caveats: high-rank standard H3 attention LoRAs degrade output, and some knowledge is reduced.
- Training can use the official s3rawint8 quantized base with Fizgig; sample workflows and a LoRA collection are published.
More from Multimodal
- isometric-cities: coding agents turn real neighborhoods into pixel-art maps on Mac — zsakib_ · 2026-11-21
- AI turns a simple story into an entertaining video — HeyZoyaKhan · 2026-11-21
- Alchemical Circuit Board prompt fuses medieval mysticism with electronics — LudovicCreator · 2026-11-21
- Solo creator makes Dungeon Crawler Carl short film locally with MiniMax H3 on RTX 4090 — SkeletonKeys91 · 2026-11-21
- Leaked samples of Nano Banana Pro 2 surface, users impressed if real — Angaisb_ · 2026-11-21
- Asked Claude Opus 5.5 to make a film about everything it finds beautiful — prasenx · 2026-11-21