[SD 1.5][LORA][LuisaP] ⚠️LORA don't need to be bigger!!!⚠️ *New Recipe test 2
<p>Stable diffusion = 2GB, Trained on 5B images.<br />Lora = 128mb, trained on 10/100/300?????</p><p><strong>this image, for example, was trained in 1 dim, 1 alpha, yes, 1 mb of filesize.</strong><br />and also, trained with only 3 images.</p><p><strong><em>a portrait of a girl on red kimono, underwater, bubbles</em></strong></p><img src="https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/9ebcdaaa-5905-4a2b-d786-cf57c27aa000/width=525" /><p>and this too, the style is identical and it's changes with prompt.</p><p><strong><em>a portrait of a girl</em></strong></p><img src="https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/f1ce33b9-c37a-42e0-e6da-f9748a024500/width=525" /><p><strong><em>a portrait of elon musk</em></strong></p><img src="https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/2de4f508-3976-47b5-b74d-9d7f9d003e00/width=525" /><p><strong><s>unet_lr: 2e3</s></strong><s><br /></s><strong><s>network_train_on: unet_only [ for styles ]</s></strong></p><p><strong><s>100 repeats 5 epochs because uses low number of images.</s></strong><br /><strong>//////////////// New training setup</strong><br /><strong>my new training recipe is 1e3, unet only, dim and alpha 1.</strong></p><p><strong>cosine with restart / 12 cycles.</strong></p><p><strong>10 repeats / 20 epochs.</strong></p><p><em>⚠️was trained on anime vae, so it's need anime vae or will look fried ⚠️</em><br /></p>
页:
[1]