Just another shitty experiments.<hr />Trained on: illustriousXL_v01;Artist: <a target="_blank" rel="ugc" href="https://danbooru.donmai.us/posts?tags=gaku&amp;z=5">gaku</a>.<hr />Recommendations: <ul><li>str: ~0.8;</li><li>cfg: 5;</li><li>sampler: euler a, 28 steps; </li><li>hires fix: 1.5x, 14 steps, 0.4 denoise; </li><li>trigger word: gaku artstyle;</li><li>have a bias/overtrain/bad-captions towards sweat and hair intakes so add it in the negative for scene that don't need them.</li></ul><hr />The idea is to put higher rank on q and v (in this case 128 rank) and train all layers possible ("module type table: {'LoConModule': 1058, 'NormModule': 256}" reported by kohyass) with lower rank, then using rs_lora to scale the different in rank. Settings: <pre><code>dim 4, alpha 1, rs_lora=True, train_norm=true,</code></pre>target name: <pre><code>enable_conv = true

unet_target_module = [
]
unet_target_name = [
 "^.*to.(q|v).*$",
 "^.*to.(k|out).*$",
 "^.*ff.*$",
 "^.*proj.(in|out).*$",
 "^(?!.*(proj.(in|out)|attn|ff)).*$",
]
text_encoder_target_module = [
]
text_encoder_target_name = [
 "^.*(q|v).proj.*$",
 "^.*(k|out).proj.*$",
 "^.*mlp.*$",
]

[name_algo_map]
 [name_algo_map."^.*to.(q|v).*$"]
 algo = "lora"
 dim = 128
 alpha = 1
 rs_lora = true
 dora_wd = true
 wd_on_output = true
 
 [name_algo_map."^.*to.(k|out).*$"]
 algo = "lora"
 dim = 16
 alpha = 1
 rs_lora = true
 dora_wd = true
 wd_on_output = true
	
 [name_algo_map."^.*ff.*$"]
 algo = "lora"
 dim = 8
 alpha = 1
 rs_lora = true
 dora_wd = true
 wd_on_output = true
 
 [name_algo_map."^.*proj.(in|out).*$"]
 algo = "lora"
 dim = 4
 alpha = 1
 rs_lora = true
 dora_wd = true
 wd_on_output = true
 
 [name_algo_map."^(?!.*(proj.(in|out)|attn|ff)).*$"]
 algo = "lora"
 dim = 4
 alpha = 1
 rs_lora = true
 
 [name_algo_map."^.*(q|v).proj.*$"]
 algo = "lora"
 dim = 128
 alpha = 1
 rs_lora = true
 dora_wd = true
 wd_on_output = true
 
 [name_algo_map."^.*(k|out).proj.*$"]
 algo = "lora"
 dim = 16
 alpha = 1
 rs_lora = true
 dora_wd = true
 wd_on_output = true
 
 [name_algo_map."^.*mlp.*$"]
 algo = "lora"
 dim = 8
 alpha = 1
 rs_lora = true
 dora_wd = true
 wd_on_output = true
 </code></pre>

Files with SHA256: 23133afff54963ed86f6ca2c00d09e498c3c4dc27b5096fff5fce4d002429c22