{
"$type": "site.standard.document",
"bskyPostRef": {
"cid": "bafyreie7zya2tqehag4scxc6lnelxorc7usmvatqvuigf5cuw66wtaaiba",
"uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mhfhyxkgx532"
},
"path": "/t/about-traning-lora-for-z-image-turbo/173911?page=2#post_25",
"publishedAt": "2026-03-19T01:55:13.000Z",
"site": "https://discuss.huggingface.co",
"textContent": "Congratulations!\n\n> rank 16 and around 40 images will be insufficient. What are your thoughts on this?\n\nIf I were in your shoes, I’d set aside the standard Z-Image Turbo for now. I’d look for a base model derived from Z-Image Turbo that’s suited to your specific purpose (someone’s probably already built one…) and apply your LoRA to that.\n\nThe reason is that when fine-tuning Z-Image Turbo itself for a domain where it lacks prior knowledge, even thousands—or in some cases, tens of thousands—of data points might not be enough… I think leveraging someone else’s work is the most practical solution.",
"title": "About traning LoRa for Z Image Turbo"
}