For Trigger Word, Use only one at a time, multiple use is experimental. Trigger Word: LolPaintArt SemiRealPaint PolishedStylisedAnime TexturedAnime 3dArt PolishedManga PolishedAnime PolishedSketch PolishedPencilArt PolishedGameArt elsword style shading Photograph Key Training changes: - Removed images with styles that are vastly different (too simple, too sketchy, contains extra body parts on edges/corners, face appear too small) - Use different trigger word for different style of training images. - LR Scheduler cosine for more stable DoRA. - Noise offset 0.1 for more aggressive learning of deep dark and bright light contrast. - Multires noise discount 0.3 - Repeat 1. Epoch 1. Steps 145. Faster Experiment of DoRAs to explore and optimise results. Parameter Settings Network Module DoRA Use Base Model WAI-illustrious - v16 Trigger words - Image Processing Parameters Repeat 1 Epoch 1 Save Every N Epochs 1 Training Parameters Seed - Clip Skip 2 Text Encoder learning rate 1 Unet/DiT learning rate 1 LR Scheduler cosine Optimizer Prodigy Network Dim 64 Network Alpha 32 Gradient Accumulation Steps - Label Parameters Shuffle caption - Keep n tokens - Advanced Parameters Noise offset 0.1 Multires noise discount 0.3 Multires noise iterations 10 conv_dim - conv_alpha - Batch Size 1 Sample Image Settings Prompt SemiRealPaint, 1girl, sfw, green eyes, green hair, short hair, green turtleneck, ribbed sweater, necklace, turtleneck, lips, white background, jewelry, looking at viewer, female solo, sweater, upper body, fur_collar, fur trim, simple background masterpiece, best quality, amazing quality, high quality, Sampler euler Steps: 145 Loss: 0.0904