LoRa 与 LyCORIS 融合

在约 217,650 张 MJ 图像上训练

LyCoris 在 211,924 张 MJ 图像上训练
LoRa 在 217,650 张 MJ 图像上训练

在 di.FFUSION.ai-stage-I.safetensorsdim 中合并

dim128 | conv32

• UNet 注意力权重平均幅值：5.714574005824907

• UNet 注意力权重平均强度：0.014448476696414102

• UNet 卷积权重平均幅值：3.623326885917904

• UNet 卷积权重平均强度：0.00653153534681701

• 文本编码器权重平均幅值：5.401999128837191

• 文本编码器权重平均强度：0.010774937006810953

在配备 8 块 A100（80GB）和 8 块 40GB 显卡的设备上，训练时间均不足一天

样本生成使用：SD 2.1 768 NoEma & Ema / realismEngine_v10 / classicNegativeSD21_classicNegative768px / aZovyaRPGArtistTools_sd21768V1

负向提示词：

(deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, red eyes, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers:1.4), disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation (oversaturated:1.5),(saturated:1.2),big contrast,contrast white burn,white spots overexposed,over saturated,burnt image,ugly,hideous,deformed,mutant,butchered,gore,sloppy,artifacts,mutilated,poorly drawn,poorly detailed,smudged,sketch,pencil,glossy skin,doll,plastic,lowres,overexposed,broken,fuzzy,aliasing,cheap,oldschool,poor quality,pixelated,sleepy,closed-eyes,crippled,crooked,weird,odd,distorted,erased,cut,hideous,pixelated,text,logo,watermark,deformed eyes,close up,((disfigured)),((bad art)),((deformed)),((extra limbs)),(((duplicate))),((morbid)),((mutilated)),out of frame,extra fingers,mutated hands,poorly drawn eyes,((poorly drawn hands)),((poorly drawn face)),(((mutation))),blurry,((bad anatomy)),(((bad proportions))),cloned face,body out of frame,bad anatomy,gross proportions,(malformed limbs),((missing arms)),((missing legs)),(((extra arms))),(((extra legs))),(fused fingers),(too many fingers),(((long neck))),tiling,mutated,cross-eye

训练数据：

FFusion SM - LyCoris

"ss_sd_model_name": "stabilityai/stable-diffusion-2-1-base",

"ss_resolution": "(768, 768)",

"ss_clip_skip": "2",

"ss_num_train_images": "211924",

"ss_network_args":

"conv_dim": "16",

"conv_alpha": "16",

"algo": "loha"

"ss_network_module": "lycoris.kohya",

"ss_optimizer": "torch.optim.adamw.AdamW(weight_decay=0.01,betas=(0.9, 0.99))",

FFusion SMB 3 - LoRa

"ss_resolution": "(512, 512)",

"ss_clip_skip": "2",

"ss_num_train_images": "217650",

"ss_network_alpha": "128.0",

"ss_network_args": {

"conv_dim": "32",

"conv_alpha": "32"

"ss_network_dim": "128",

"ss_network_module": "networks.lora",

"ss_num_batches_per_epoch": "54422",

"ss_optimizer": "dadaptation.dadapt_adam.DAdaptAdam(decouple=True,weight_decay=4.0,betas=(0.9, 0.98),use_bias_correction=True,log_every=100)"

仅使用 Kohya S 与 Shih-Ying Yeh (Kohaku-BlueLeaf) 工具完成完整训练 🤟 🥃