Trash multicolor result when using my Lora

[novitae]

Hello @bmaltais , I am facing an issue: every time I use a Lora I made, from the "Additional network" or directly in the prompt, I get the following result:

I trained my lora following exactly this video, using the same configuration file. The only thing I changed is the precision to fp16, since the distant linux computer I used for the task wasn't supporting the bf16

I retrained my Lora an put Mixed precision and Save precision respectively on no float, and now the result I get is a dark image:

The training is made on the runwayml/stable-diffusion-v1-5 model, and I run its Lora result on the v1-5-pruned.ckpt from runwayml/stable-diffusion-v1-5 from hugging faces

I am running the Lora on a Mac m2, which might cause issues, but I tried other Loras (that were probably generated with bf16), and I works with them. Where am I going wrong ? Thanks !

[rushuna86]

have you tried just not loading his config file? i remember that video being very old and the old config files from that version not working with the new version, ie naming change etc. hell even my config file going from pre multires/noise offset to version after ddint work let alone config file that old.