Princess Connect - Priconne Character Focus ☆3/☆6 Style版本v1-Counterfeit (ID: 19708)

This LoRA model is a bit unique in that not only does it assist in imitating the Princess Connect! Re:Dive art style, but due to the relatively consistent composition and posing in the reference works it is very good at creating scenes with a centrally located character as the focus with high level of detail for and around the character.

Main features

• Good at wide aspect ratios

• Good at keeping characters in frame ("feet out of frame", "thighs", etc. prompts won't cause super zoom in nearly as often)

• Good at enhancing visual elements around the subject character

• Priconne art style

Other features

• Good for highres fix (all images bucket to 960x512, trained at 960 res)

• A bit better than usual at background continuity across areas interrupted by foreground (just a little, don't expect miracles)

• Helps somewhat with Counterfeit-V2.5 LoRA compatibility by reducing the multiple characters appearing issue.

• Can assist with composing scenes for use with ControlNet, img2img to achieve detailed compositions of different styles

Use & Prompting

• No activation tokens are required.

• (At the moment) For use with Counterfeit-V2.5. I can't guarantee it'll work well on any other model, as LoRA trained on Counterfeit-V2.5 tend to behave oddly when used elsewhere.

• Weights between 0.6-1.0 recommended.

  • If using with character LoRA, try starting on the lower end

  • Other style LoRAs will tend to clash. I recommend low or no weight of the accompanying LoRA for initial gen and then img2img afterwards

• Sometimes, faces will appear on the right side (due to ☆6 Rima...). Add (solo) or other such emphasis/negatives to suppress this.

Dataset

All ☆3 and ☆6 character art up to Café Tamaki. Six star artworks were repeated 4x per epoch, three star 1x per.

Tagging was done via wd14-swinv2-v2 with threshold 0.15. All character tags were then removed.

描述:

Trained on Counterfeit-V2.5_pruned.

Example images

Generated using Counterfeit-V2.5_pruned. When used alone weight 0.9-1.0 was used. When used with character LoRA weight 0.6 was used. All images were processed with txt2img + highres fix. For the Ichiso-style image, after initial generation, img2img Ultimate SD Upscale was used to restyle the image using only the target style LoRA.

Training Parameters

ss_clip_skip: "2"

ss_color_aug: "False"

ss_dataset_dirs: "{"1_threestar": {"n_repeats": 1, "img_count": 223}, "4_sixstar": {"n_repeats": 4, "img_count": 52}}"

ss_epoch: "60"

ss_flip_aug: "False"

ss_full_fp16: "False"

ss_gradient_accumulation_steps: "1"

ss_keep_tokens: "1"

ss_learning_rate: "0.0001"

ss_lr_scheduler: "cosine_with_restarts"

ss_max_bucket_reso: "960"

ss_max_token_length: "150"

ss_mixed_precision: "fp16"

ss_network_alpha: "128.0"

ss_network_dim: "128"

ss_new_sd_model_hash: "a074b8864e31b8681e40db3dfde0005df7b5309fd2a2f592a2caee59e4591cae"

ss_num_reg_images: "0"

ss_random_crop: "False"

ss_resolution: "(960, 960)"

ss_sd_model_name: "Counterfeit-V2.5_pruned.safetensors"

ss_shuffle_caption: "True"

ss_text_encoder_lr: "1e-05"

ss_total_batch_size: "2"

ss_unet_lr: "0.0001"

训练词语: