[SD 1.5][Hypernetwork][LuisaP] Tutorial Hypernetwork - Monkeypatch method none
<p><strong><em>clip 2, vae on, hypernetwork strenght 1.</em></strong></p><p>1-Install <strong>Monkeypatch Extension </strong>and reload the ui</p><p><a target="_blank" rel="ugc" href="https://github.com/aria1th/Hypernetwork-MonkeyPatch-Extension">https://github.com/aria1th/Hypernetwork-MonkeyPatch-Extension</a></p><p>2-Go to <strong>create Beta hypernetwork</strong> in your <strong>train </strong>section.<br />3-Place this layer structure <strong>1,0.1,0.1,1 </strong>//thanks queria!, i personally like this so much.</p><p>4-Select activation function of hypernetwork:<strong>tanh</strong></p><p>5-Select Layer weights initialization:<strong>xavier normal</strong></p><p>6-and finally, create the hypernetwork.</p><p>7-now in <strong>Train_Gamma,</strong> select your new hypernetwork.</p><p>8-Hypernetwork Learning rate: <strong>6.5e-3 </strong>"this is for the math" so is perfectly normal ,also, <strong>6.5e-4 </strong>will cause less damage to original image.</p><p>9-enable <strong>Show advanced learn rate scheduler options(for Hypernetworks) </strong>and <strong>Uses CosineAnnealingWarmupRestarts Scheduler.</strong></p><p>10-Steps for cycle = number of images in your dataset.</p><p>11-Step multiplier per cycle: 1.1 or 1.2</p><p>12-Warmup step per cycle = the half of number of images.</p><p>13-Minimum learning rate for beta scheduler = 1e-5 [ or <strong>6.5e-7 </strong>, will get less style from dataset, but more control<strong> ]</strong></p><p>14-Decays learning rate every cycle = 0.9 or 1</p><p><strong><em>15a-batchsize 2, grad 1, steps 1000.</em></strong></p><p>15b-you can also do this [ batchsize 2, grad(number of image in dataset divided by two) but for that you only will need something like 250 steps, but personally i don't like it.</p><p><br />16- your prompt file need to be style.txt.</p><p>17- you can also try to "Read parameters (prompt, etc...) from txt2img tab when making previews" to see results with the style in your prompt, for example, mine is "girl in a red kimono".</p><p>Note: i train with<strong> 2 clip skip</strong>, <strong>none hypernetwork,</strong> and <strong>1 hypernetwork strength.</strong></p><p>18- and i'ts that! 5 MB of hypernetwork trained in under 10/20 minutes.</p>
页:
[1]