TURBO Merge of new fp8 Chroma1-HD quant and Flash Heun Rank 256 LoRa.

This model runs on CFG 1-1.5 and is much faster than conventional Chroma1-HD

v 0.3: Full LoRa weight, recommended CFG of 1

v 0.1: LoRa weight is at 0.7, so you can go a bit higher with CFG (allows more pale colors at CFG 1 in creation mode since you can't go below CFG 1 in settings)

Settings I tested:

CFG: 1

Sampler: Heun

Scheduler: Beta

Steps: 8-13

CFG !

Sampler: Res multistep

Scheduler: Beta

Steps: 15-26 (26 gave me the most stable photorealism I believe)

You can try to go much lower with steps. Results may differ from subject. Realism more steps than 2d. You may increase CFG slightly to 1.2-1.5 at a lower step count. Just experiment!

This is a merge of silveroxides newest fp8 scaled mixed precision Chroma1-HD quant that seems to be significantly faster than his previous fp8 quants and Rank 256 distill LoRa.

LoRa combined with this model runs really fast and only requires a CFG of 1 (this is primarly the reason its so fast). On the downside negative prompts don't work anymore but honestly who is really using negative prompt in Chroma?

This combo of new fp8 quant and lora brought my local generation time down from about 50s to 18s per image. Really impressive. For context Z-Image Turbo needs 10s with my hardware.

silveroxides quant:

https://huggingface.co/silveroxides/Chroma1-HD-fp8-scaled/blob/main/Chroma1-HD-fp8mixed-final.safetensors

LoRa:

https://civitai.com/models/2032955?modelVersionId=2301229