https://huggingface.co/ByteDance/Hyper-SD/
Recommending the 8-step-CFG preserve version as it's much easier and nicer to use.
These are just suggestions. Experimentation is required and you will waste some credits to understand this LoRA.
Weight: 0.8 - 1.0
Sampling method: DDIM, Euler, anything! BUT NOT Euler A or DPM 2S a Karras. Because of how ancestral samplers work, you get bland boring outputs below 8 steps.
Sampling steps: 7 minimum, 8 recommended.
CFG is still not currently known or explored yet. You may try between 4.0 - 7.0 CFG
TODO: rentry articles going in-depth about hyper-sd capabilities and shortcomings.
Cover artwork: https://pixai.art/artwork/1785900648729422093
![[PONY] Hakuryuu | Azur Lane - AI Model cover image](https://images-ng.pixai.art/images/thumb/bb5bc51d-8137-4b13-97fa-5e85ce0450b2)
