<h2 id="for-other-lora-lab-tests-ronz3yo2e">For other LoRa Lab tests</h2><h3 id="goto-:-https:civitai.commodels1366934-y21iyeb37">goto : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1366934">https://civarchive.com/models/1366934</a></h3><h2 id="resource-intro.-gpr7uquv7">Resource intro.</h2>Please check the name :<ul><li>lora_lab_{lab round}_{test number}_p{picture number}_{short description}</li></ul><ul><li>dataset_{lab round}_{test number}</li><li>reg_{lab round}_{test number}</li></ul><h2 id="lora-lab-@-4-try-less-image-to-see-if-it-works-ywlcam60u">LoRa Lab @ 4 - try less image to see if it works</h2>Goal : as @3 shows, trigger token &amp; tagged reg images did work. So, will it also do the trick on less dataset?<ul><li>#1 : 8 pics; bad reg images (with lower body images, but tagged); trigger token : nsfwish, golden panyhose<ul><li>BAD -- it kinda gives me lower body image even +full body is prompted</li></ul></li><li>#2 : 8 pics; better reg images (tagged; but with some strange non-girl reg images -- 1girl, 1boy)<ul><li>OK -- but images might lose some logical composition at complex prompt</li></ul></li><li>#3 : 12 pics; better reg images (tagged; but with some strange non-girl reg images -- 1girl, 1boy)<ul><li>OK, better than #2 -- but images still might lose some logical composition at complex prompt</li></ul></li><li>#4 : on WAI v11 !! 12 pics; only full body reg images (tagged);<ul><li>BETTER than #3 but on WAI v11 -- only problem is "black sheer pantyhose could shows"</li></ul></li><li>#5 : back on WAI v12!! 12 pics; only full body reg images (tagged);<ul><li>findings :<ul><li>GOOD on WAI v12 !! but JUST OK on WAI v11</li><li>also, #4 is good on v11, but just ok on WAI v12 ( it seems compatibility is a thing )</li></ul></li></ul></li><li>#6 : 12 pics; full body reg images; trigger token : nsfwish, sheer golden pantyhose<ul><li>BETTER, it seems that sheer natural pantyhose is easier to re-produce</li></ul></li></ul><h3 id="lora-lab-@-4-findings-r5fy7viyw">LoRa Lab @ 4 -- Findings</h3><ul><li>on 3090 24G<ul><li>p8 training<ul><li>TIME : ~ 16 min</li></ul></li><li>p12 training<ul><li>TIME : ~ 21 min</li></ul></li></ul></li><li>Interestingly, less dataset still gives promising result<ul><li>AND trigger token &amp; proper reg images with tags DID help</li><li>and this is really great finding, as dataset experiments can really SPEED UP without a full dataset test run !!</li><li>HOWEVER, the trigger word <code>sheer golden pantyhose</code> seems to also afffect other part of the image, e.g. the tone, the hair, shoes, ...<ul><li>Next time could try some trigger token which is color neutral one</li></ul></li></ul></li><li>LoRa trained on different base model (v11, v12), works better on their own base model</li><li>WAI v12 is actually better than v11, indeed<ul><li>v11 still tends to give me lower body, when <code>1girl, solo, pantyhose</code> is given</li></ul></li></ul>

Lab @4 : trying on less images to see if it works

lora_lab_4_2_p8_wr_tagged.safetensors

lora_lab_4_2_p8

lora_lab_4_3_p12

lora_lab_4_4_p12

lora_lab_4_5_p12

reg_4_4-5-6

reg_4_2-3

dataset_4_6

dataset_4_3-4-5

dataset_4_2

lora_lab_4_6_p12

Files with SHA256: 4ca248a7895e794c76a01de9497d46208d5e930b42c865606db4be3fc4befe63