Flux:

A Flux dev version of the SDXL images and captions. Please check the version description for more information. I've included the preview wildcards as well.

I tried to mirror the SDXL training workflow more closely in an attempt to homogenize the process for both models. Captions were closer to the mixed tag/natural language format of the SDXL process rather than my previous overly verbose Flux training prompts.

Overview:

• Model consists of five of her cosplay sets, Alya, Kotori, Momo, Nagatoro, and Rin.

• Masked training was not used.

• Regularization was not used.

• Dreambooth trained using the sd3-flux.1 branch of the Kohya_ss scripts.

• Dreambooth trained on Flux.1-Dev (fp16), then LoRA extracted.

The training was tested using the MGAS comparison method to find the best extracted-epoch. The introduction article can be found here.

Usage:

1. Intended for use with Flux.1-Dev. Performs OK on other realistic focused models.

2. Size: 1024,1024, Steps: 20, Sampler: Euler-Simple, CFG/Guidance scale: 1.0/3.0-3.5, Strength: 1.0-1.1

3. [Optional] Adetailer. Was NOT used for preview images.

4. All cosplays are tied together by the common tag, {cata7pln}, which consists of two rare tokens on either side of a number.

5. Add the following tags in your prompts for the individual cosplays:

   • Alya - {Alya cosplay, long pastel pink hair, blue eyes, red ribbon in hair, red bowtie with white stripes, grey blazer, blue vest and skirt, white undershirt, white thighhighs}

   • Kotori - {Kotori cosplay, long red hair, twintails, red eyes, two-tone bow in hair, black tie, red blazer, white undershirt, red skirt, black thighhighs}

   • Momo - {Momo cosplay, short brown hair, red eyes, large green earrings, black choker, white shirt collar, pink sweater, red bowtie, blue skirt, white leg warmers}

   • Nagatoro- {Nagatoro cosplay, long black hair, hazel eyes, hair clips in hair, white button-up shirt, rolled up sleeves, black pleated skirt}

   • Rin- {Rin cosplay, long brown hair, twintails, green eyes, black bow in hair, red turtleneck sweater, white cross logo, black pleated skirt, black thighhighs}

Future Outlook:

I plan to further refine my training workflow using my favorite cosplayers as I continue to build the dataset for the eventual nobody checkpoint.

==========================================================

SDXL:

Testing out my training workflow on another fantastic cosplayer.

Overview:

• Model consists of five of her cosplay sets, Alya, Kotori, Momo, Nagatoro, and Rin.

• Masked training was used for more flexible backgrounds.

• Regularization set was captioned with GPT-4o for greater control of details.

• Original Dreambooth trained using Constant AdaFactor with Min_SNR_Gamma loss.

It was extensively tested using the MGAS comparison method (here) during both training and image generation, while an example regularization prompt can be found in the VDC article (here).

Usage:

1. Trained on, and works best with, RealVisXL V5.0. Works well with other popular realistic XL checkpoints such as JuggernautXL, jibMixRealisticXL, mklanRealisticXL, copaxTimelessXL, etc.

2. Size: 1024,1024, Steps: 50, Sampler: DPM++2m Karras, CFG scale: 7, Strength: 1.0

   • DPM++SDE/Euler-a Karras at 30 steps works as well.

   • 832x1216 and 1216x832 (i.e. 2:3 & 3:2) tend to work well too.

3. [Optional] Adetailer, with denoise around 0.4.

4. All cosplays are tied together by the common tag, {cata7pln}, which consists of two rare tokens on either side of a number.

5. Add the following tags in your prompts (see preview images for examples) for the individual cosplays:

   • Alya - {Alya cosplay, long pastel pink hair, blue eyes, red ribbon in hair, striped red bowtie, grey blazer, blue vest and skirt, white undershirt, white thighhighs}

   • Kotori - {Kotori cosplay, long red hair, twintails, red eyes, two-tone bow, black tie, red blazer, white undershirt, red skirt, black thighhighs}

   • Momo - {Momo cosplay, short brown hair, red eyes, large green earrings, black choker, white shirt collar, pink sweater (or pink windbreaker), red bowtie, blue skirt, white leg warmers}

   • Nagatoro- {Nagatoro cosplay, long black hair, hazel eyes, hair clips, white button-up shirt, rolled up sleeves, black pleated skirt}

   • Rin- {Rin cosplay, long brown hair, twintails, green eyes, black bow, red turtleneck sweater, white cross logo, black pleated skirt, black thighhighs}

==========================================================

描述:

Initial Release - Version 1.0

Dreambooth trained on 75 images, broken up into 5 sets of 15 images with one set for each cosplay (one concept), using the sd3-flux.1 branch of the Kohya_ss scripts. Training was done for 175 epochs in batches of 1. Finally, the LoRA was extracted using Koyha_ss script's utility tab.

• Total trained steps: (5 x 15) x 1 x 175/ 1 = 13,125

  - or 2,625 per cosplay/concept

• Extracted LoRA rank/alpha = 32/(dynamic)

训练词语: cata7pln