Civitai is too cumbersome to use—please download the other LoRAs mentioned in the post from Tensor instead. <a target="_blank" rel="ugc" href="https://tensor.art/models/871854896876390370/Lagoon-Engine-and-doujin-ver-1.0-ver-1.0">https://tensor.art/models/871854896876390370/Lagoon-Engine-and-doujin-ver-1.0-ver-1.0</a>Previously on <a target="_blank" rel="ugc" href="http://Tensor.art">Tensor.art</a>: <a target="_blank" rel="ugc" href="https://tensor.art/models/859228586379351423/Lagoon-Engine-and-doujin-ver.beta-beta1">https://tensor.art/models/859228586379351423/Lagoon-Engine-and-doujin-ver.beta-beta1</a> The beta version was released earlier. Now, version 1.0 brings several improvements over the beta: Hair rendering has improved—no more “black brown” hair.The characters now look a bit older, growing from shota to shonen.Skin tone has been deepened and is no longer too pale.These changes slightly shift the overall art style.The biggest change of all is the addition of two more characters I really love. When generating solo images with 1boy, characters are almost always correct.With 2boys, any two characters can be paired, though there's still some chance of feature blending.Due to the randomness introduced by LoRA training, and the influence of CFG, resolution, sampler, scheduler, prompt, and checkpoint choices during generation, it seems that character mixing can't be fully avoided.As for 3boys or 4boys group images—don’t even try.At best, hairstyles might remain accurate, but eyes and other features often get jumbled together. If you feel the characters in version 1.0 don’t quite resemble the originals, you can use the provided En &amp; Jin LoRA ( <a target="_blank" rel="ugc" href="https://tensor.art/models/859228586379351423/Lagoon-Engine-and-doujin-ver.beta-beta1">https://tensor.art/models/859228586379351423/Lagoon-Engine-and-doujin-ver.beta-beta1</a><a target="_blank" rel="ugc" href="https://tensor.art/models/859228586379351423/Lagoon-Engine-and-donjin-ver.beta-beta1"> </a>) to generate images of the brothers together, and use the Narama &amp; Suguru LoRA for single-character images.These LoRAs are meant to be used individually—please don’t try to load them together to generate multi-character images. ⚠️ Note: The trigger words in version 1.0 are different from previous versions.Be careful not to copy the wrong ones! Version 1.0 Trigger Words 等軍焔: ragunennii, brown-short hair, purple eyes等軍陣: ragunjinotouto, green-mid hair, blue eyes早河枹馬: hayakawanaramatomo, two-tone-blue hair, darkblue eyes御上卓: mikamisuguruitoko, grey-red hair, green eyes If too many errors occur with 2boys, try adding these optional assist triggers (not required):ragun_kyoudaien_x_naramaen_x_sugurujin_x_naramajin_x_sugurusuguru_x_narama En &amp; Jin 2boys LoRA ( the ver.beta )等軍焔: ragun_en_nii, brown hair, purple eyes等軍陣: ragun_jin_otouto, green hair, blue eyes Narama &amp; Suguru lora：!!Important!! This LoRA can only generate single-character images with 1boy. If you try 2boys, the characters will merge.早河枹馬: hayakawa_narama_tomo, two-tone blue hair, darkblue eyes,御上卓: mikami_suguru_itoko, grey red hair, green eyes, Model weight suggestion:Recommendation 1. Keep it no lower than 0.7. CFG recommendation:High CFG tends to cause more issues. Stick to 4–6, or follow what works best with your checkpoint. Sampler &amp; scheduler:There’s no definitive best combo—tried many, and the error rates were similar.Try dpm++2s_a and beta first. Recommended resolutions:Use 3:4 as first choice, 2:3 as second. Training settings summary (can be checked in SD/webui after downloading the LoRA): resolution = "1024,1024"enable_bucket = truemin_bucket_reso = 256max_bucket_reso = 2048bucket_reso_steps = 64bucket_no_upscale = true save_precision = "bf16"save_every_n_epochs = 1 train_batch_size = 1gradient_checkpointing = truenetwork_train_unet_only = falsenetwork_train_text_encoder_only = false learning_rate = 1unet_lr = 1text_encoder_lr = 1lr_scheduler = "cosine"lr_warmup_steps = 300 optimizer_type = "Prodigy"network_dim = 128network_alpha = 64 randomly_choice_prompt = falsemultires_noise_iterations = 6multires_noise_discount = 0.3 Training a multi-character LoRA is incredibly challenging and still far from stable, but it’s now barely usable. With future updates to the base model, I’ll continue working on new versions. See you next time~

ラグーンエンジン, Lagoon Engine, 封魔少年焰&陣 - doujin ver 1.0

lagoon_engine_ver1.safetensors

ver 1.0

Previously on <a href="http://Tensor.art" rel="ugc noopener noreferrer nofollow" target="_blank">Tensor.art</a>: <a href="https://tensor.art/models/859228586379351423/Lagoon-Engine-and-donjin-ver.beta-beta1" rel="noopener noreferrer nofollow">https://tensor.art/models/859228586379351423/Lagoon-Engine-and-donjin-ver.beta-beta1</a>&nbsp;The beta version was released earlier. Now, version 1.0 brings several improvements over the beta:&nbsp;Hair rendering has improved—no more “black brown” hair.The characters now look a bit older, growing from shota to shonen.Skin tone has been deepened and is no longer too pale.These changes slightly shift the overall art style.The biggest change of all is the addition of two more characters I really love.&nbsp;When generating solo images with 1boy, characters are almost always correct.With 2boys, any two characters can be paired, though there's still some chance of feature blending.Due to the randomness introduced by LoRA training, and the influence of CFG, resolution, sampler, scheduler, prompt, and checkpoint choices during generation, it seems that character mixing can't be fully avoided.As for 3boys or 4boys group images—don’t even try.At best, hairstyles might remain accurate, but eyes and other features often get jumbled together.&nbsp;If you feel the characters in version 1.0 don’t quite resemble the originals, you can use the provided En &amp; Jin LoRA ( <a href="https://tensor.art/models/859228586379351423/Lagoon-Engine-and-donjin-ver.beta-beta1" rel="noopener noreferrer nofollow">https://tensor.art/models/859228586379351423/Lagoon-Engine-and-donjin-ver.beta-beta1</a> ) to generate images of the brothers together, and use the Narama &amp; Suguru LoRA for single-character images.These LoRAs are meant to be used individually—please don’t try to load them together to generate multi-character images.&nbsp;⚠️ Note: The trigger words in version 1.0 are different from previous versions.Be careful not to copy the wrong ones!&nbsp;Version 1.0 Trigger Words&nbsp;等軍焔: ragunennii, brown-short hair, purple eyes等軍陣: ragunjinotouto, green-mid hair, blue eyes早河枹馬: hayakawanaramatomo, two-tone-blue hair, darkblue eyes御上卓: mikamisuguruitoko, grey-red hair, green eyes&nbsp;If too many errors occur with 2boys, try adding these optional assist triggers (not required):ragun_kyoudaien_x_naramaen_x_sugurujin_x_naramajin_x_sugurusuguru_x_narama&nbsp;En &amp; Jin 2boys LoRA等軍焔: ragun_en_nii, brown hair, purple eyes等軍陣: ragun_jin_otouto, green hair, blue eyes&nbsp;Narama &amp; Suguru lora：!!Important!! This LoRA can only generate single-character images with 1boy. If you try 2boys, the characters will merge.早河枹馬: hayakawa_narama_tomo, two-tone blue hair, darkblue eyes,御上卓: mikami_suguru_itoko, grey red hair, green eyes,&nbsp;Model weight suggestion:Recommendation 1. Keep it no lower than 0.7.&nbsp;CFG recommendation:High CFG tends to cause more issues. Stick to 4–6, or follow what works best with your checkpoint.&nbsp;Sampler &amp; scheduler:There’s no definitive best combo—tried many, and the error rates were similar.Try dpm++2s_a and beta first.&nbsp;Recommended resolutions:Use 3:4 as first choice, 2:3 as second.&nbsp;Training settings summary (can be checked in SD/webui after downloading the LoRA):&nbsp;resolution = "1024,1024"enable_bucket = truemin_bucket_reso = 256max_bucket_reso = 2048bucket_reso_steps = 64bucket_no_upscale = true&nbsp;save_precision = "bf16"save_every_n_epochs = 1&nbsp;train_batch_size = 1gradient_checkpointing = truenetwork_train_unet_only = falsenetwork_train_text_encoder_only = false&nbsp;learning_rate = 1unet_lr = 1text_encoder_lr = 1lr_scheduler = "cosine"lr_warmup_steps = 300&nbsp;optimizer_type = "Prodigy"network_dim = 128network_alpha = 64&nbsp;randomly_choice_prompt = falsemultires_noise_iterations = 6multires_noise_discount = 0.3&nbsp;Training a multi-character LoRA is incredibly challenging and still far from stable, but it’s now barely usable. With future updates to the base model, I’ll continue working on new versions.&nbsp;See you next time~

Files with SHA256: c65e4638eedeb6b3fb9e862aee8ac57dcf0d666a38940dfabc95d6fa6679ab1f