chroma-unlocked-v4x-hyper-turbo-flash-r64-bf16
I spent some time experimenting with chroma-unlocked-v4x-hyper-turbo-flash-r64-bf16 lora with the v47_detail_calibrated checkpoint. I used a strength 1.0, cfg 1.0, steps anywhere from 5-15, dpmpp_2m/sgm_uniform. The results looked pretty good and was nearly as fast as SDXL. If you really need the negative prompt and don't mind the speed hit you can do a cfg of up to 1.5. Any higher and it gets pretty burned. Step count looks good as low as 5. Any lower and you start seeing a vignette around the image.
The only drawback is that this lora seems to kill realism, though adding more loras might fix that.
Anyway, I'm glad there's a way to speed up Chroma, which I feel was its only drawback. Looking forward to v50 and whatever happens after that.
The workflow was nothing special: Load model -> Lora -> TorchCompile -> Sage -> KSampler -> VAE -> Save
Apparently this is another flux few-step lora? https://huggingface.co/yresearch/swd_flux No details in the readme, and I haven't tested it - but figured it might be another candidate for a few step frankenmerge recipe.