This is my LoRA extraction of the Cosmos-Predict 2.5-2B. Using the Pre-trained and the Distilled models. Converted from pt to safetensor then cleaned during extraction. These give no errors or warnings that I could find. I used adaptive rank and no clamping. Ready to merge.

I have included 3 variations. 32 / 48 / 64. Up to you which you prefer. I would recommend the 32. Its what i used for all the image previews.

CFG: 1.0 - 1.7 Depending what style you want. I used 1.0-1.4 in the previews
Steps: 8 - 20 Around 12 is good
Strength: ~1.0
Sampler: er_sde / euler_ancsetral
Scheduler: sgm_uniform / kl_optimal

Basic Stats

I do this for fun and to learn. Likes, comments and reactions are nice to see if my efforts are worth while and posting. I dont really care about the buzz or whatever but drop me a comment and like if you enjoy what Ive done. <3