amazing work, really
I managed to get better results with your lora on 3 steps than accvideo and causvideo on 12 steps! amazing job.
can you please, if you got time, write a little list on best settings (lora strength, Guidance Scale , shift scale) to each purpose, you should have a deep knowledge of that.
my current settings using your lora (Wan2.1_I2V_14B_FusionX_LoRA.safetensors)
for prototyping:
Guidance Scale : 1
Shift Scale : 2 for realism, 4 for cartoonish things
steps: 3
lora strength: 2
for good results:
Guidance Scale : 1
Shift Scale : 2 for realism, 4 for cartoonish things
steps: 12
lora strength: 1
I haven't tested a setting to produce premium results, can you share some tips?
I honestly don't use image to video as much as text to video because I personally find it hard to get good results because I think I'm just prompting wrong. I create the i2v merge and lora because I had so many requests, otherwise this was mainly for text to video. It ended up blowing up into everything though and I had no idea it would. Was just a small personal project that ended up getting way more attention then expected. :)
the default settings in the workflow I shared would be what I would call best but again, I'm still not sure I even know the best settings for image to video.