[SD 1.5][LORA][LoRA] Nagayama Yunon style for Anything v4.5 lora-v1.0
<ul><li><p>checkpoint: <a target="_blank" rel="ugc" href="https://huggingface.co/andite/anything-v4.0">Anything v4.5</a> (<strong>DO NOT Forget VAE</strong>)</p></li><li><p>no trigger words required</p></li><li><p>recommond <strong>CFG Scale = 7.0</strong></p></li><li><p>recommond weight:<strong> 0.6</strong></p></li><li><p>recommond use hires.fix</p></li></ul><p></p><h2>Training params</h2><pre><code>$resolution = "768,1024"$batch_size = 3
$max_train_epoches = 8
$save_every_n_epochs = 1
$network_dim = 128
$network_alpha = 128
$clip_skip = 2
$lr = "1e-4"
$unet_lr = "1e-4"
$text_encoder_lr = "1e-5"
$lr_scheduler = "cosine_with_restarts"
$lr_warmup_steps = 0
$lr_restart_cycles = 1
$noise_offset = 0.05
$keep_tokens = 3
$min_bucket_reso = 256
$max_bucket_reso = 1024
$use_lion = 1</code></pre><p></p>
页:
[1]