Overview

After toying around with various models and prompts with no luck (even inpaint), I decided it was best to just create my own in honor of Buckethead. I'm planning to refine it and release v2, v3, etc. but happy with the experiment.

Recommendations

The LoRA was trained on base SDXL 1.0 but has been tested with a variety of custom SDXL-based models with varying levels of accuracy and uniqueness. All capture the character with the bucket, mask, guitar, and typical clothing, but may alter more subtle aspects (e.g., eye emphasis, stringy guitars, mask looking too "human-like").

Models that work well with minimal prompting (so far)

• SDXL 1.0 base

• SDXL_DarkArts

• GhostXL

• BetterThanWords

• StarlightXL Animated (v3)

Negative prompt (refine as needed)

lips, woman, girl, teeth, neck beard, chin hair, ((beard))

Example prompts

<lora:Buckethead-10:1>, Buckethead, solo, white mask hiding face, electric guitar, majestic cave concert, blue purple orange cave lighting, crowd in foreground, underground, cinematic lighting, intricate cavern wall, emotionless, WLOP, shadowy, ornate, detailed

• Negative prompt: lips, woman, girl, teeth, neck beard, chin hair, ((beard)), Deformity v6 (Neg), Bad v5 (Neg), UnrealisticDream (Neg)

• Settings: Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 5, Seed: 1441473974, Face restoration: CodeFormer, Size: 1152x896, Model: duchaitenAiartSDXL_v33515 (1), VAE: sdxl_vae.safetensors, Clip skip: 2

<lora:Buckethead-10:1>, Dystopian style, cinematic film still, breathtaking, Buckethead, the guitarist. KFC bucket on head, Halloween 4 white mask, instrument, guitar, solo, playing instrument, electric guitar, emotionless mask, male focus, holding, music, covered face. hyperrealistic, hyperrealism, 8k, cinematic lighting, WLOP. extremely detailed. Eerie, unsettling, dark, spooky, suspenseful, grim, highly detailed . Dark, mysterious, haunting, dramatic, ornate, detailed. Dark concert in caverns, Gothic style Horror-themed, mirrorless camera equipped with a wide-angle lens and an f/4 aperture for optimal depth of field. high-end DSLR quality.

• Negative: woman, girl, simplistic, minimalist, teeth, cheery, smiling, vibrant, Deformity v6 (Neg), Bad v5 (Neg)

• Settings: Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 3.0, Seed: 263278052, Face restoration: CodeFormer, Size: 1152x896, Model: SDXL_DarkArts, VAE: sdxl_vae.safetensors, Clip skip: 2,

Training

Trained on 36 hand-picked photos of Buckethead in concert and various other settings, with the majority holding his signature Gibson Les Paul Custom. Total of 10 epochs, saving at each step to test for accuracy (total of 900 steps).

Accuracy

When mixing in certain LoRAs or SDXL model variants, it sometimes makes Buckethead more feminine or will try to place human lips on the Halloween IV mask. Adjust accordingly with negative keywords to work around the issue until the next version (i.e., adding default negatives and apply proper labeling). I did not use any text embeds/tags to train on as an experiment.