2025-08-14 19:28:25 INFO loading Model... resize_lora.py:321
2025-08-14 19:28:26 INFO Resizing Lora... resize_lora.py:324
INFO Dynamically determining new alphas and dims based off sv_fro: 0.9, max resize_lora.py:214
rank is 32
100%|████████████████████████████████████████████████████████████████████████████████| 912/912 [12:30<00:00, 1.21it/s]
lora_unet_double_blocks_0_img_attn_proj | sum(S) retained: 46.5%, fro retained: 94.8%, max(S) ratio: 4.9, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_0_img_attn_qkv | sum(S) retained: 46.5%, fro retained: 92.5%, max(S) ratio: 3.6, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_0_img_mlp_0 | sum(S) retained: 58.5%, fro retained: 91.9%, max(S) ratio: 3.0, dynamic | dim: 15, alpha: 15.0
lora_unet_double_blocks_0_img_mlp_2 | sum(S) retained: 55.1%, fro retained: 91.8%, max(S) ratio: 2.8, dynamic | dim: 9, alpha: 9.0
lora_unet_double_blocks_0_img_mod_lin | sum(S) retained: 47.7%, fro retained: 94.9%, max(S) ratio: 4.3, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_0_txt_attn_proj | sum(S) retained: 52.5%, fro retained: 96.4%, max(S) ratio: 4.4, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_0_txt_attn_qkv | sum(S) retained: 55.6%, fro retained: 96.5%, max(S) ratio: 3.7, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_0_txt_mlp_0 | sum(S) retained: 54.1%, fro retained: 97.7%, max(S) ratio: 6.2, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_0_txt_mlp_2 | sum(S) retained: 58.5%, fro retained: 95.9%, max(S) ratio: 2.9, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_0_txt_mod_lin | sum(S) retained: 51.8%, fro retained: 94.5%, max(S) ratio: 2.6, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_10_img_attn_proj | sum(S) retained: 52.6%, fro retained: 94.6%, max(S) ratio: 3.6, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_10_img_attn_qkv | sum(S) retained: 49.4%, fro retained: 93.7%, max(S) ratio: 3.2, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_10_img_mlp_0 | sum(S) retained: 52.8%, fro retained: 93.3%, max(S) ratio: 3.3, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_10_img_mlp_2 | sum(S) retained: 52.4%, fro retained: 93.9%, max(S) ratio: 3.8, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_10_img_mod_lin | sum(S) retained: 52.7%, fro retained: 94.4%, max(S) ratio: 3.4, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_10_txt_attn_proj | sum(S) retained: 46.9%, fro retained: 94.7%, max(S) ratio: 3.1, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_10_txt_attn_qkv | sum(S) retained: 52.2%, fro retained: 94.4%, max(S) ratio: 3.8, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_10_txt_mlp_0 | sum(S) retained: 64.9%, fro retained: 98.6%, max(S) ratio: 4.2, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_10_txt_mlp_2 | sum(S) retained: 57.2%, fro retained: 96.6%, max(S) ratio: 4.6, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_10_txt_mod_lin | sum(S) retained: 68.1%, fro retained: 98.6%, max(S) ratio: 6.8, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_11_img_attn_proj | sum(S) retained: 53.2%, fro retained: 93.1%, max(S) ratio: 2.8, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_11_img_attn_qkv | sum(S) retained: 47.6%, fro retained: 92.1%, max(S) ratio: 3.4, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_11_img_mlp_0 | sum(S) retained: 49.9%, fro retained: 92.6%, max(S) ratio: 2.9, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_11_img_mlp_2 | sum(S) retained: 56.7%, fro retained: 94.0%, max(S) ratio: 3.4, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_11_img_mod_lin | sum(S) retained: 51.6%, fro retained: 95.4%, max(S) ratio: 2.9, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_11_txt_attn_proj | sum(S) retained: 55.7%, fro retained: 96.7%, max(S) ratio: 3.6, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_11_txt_attn_qkv | sum(S) retained: 49.9%, fro retained: 93.5%, max(S) ratio: 3.5, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_11_txt_mlp_0 | sum(S) retained: 55.8%, fro retained: 97.1%, max(S) ratio: 5.0, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_11_txt_mlp_2 | sum(S) retained: 52.4%, fro retained: 94.5%, max(S) ratio: 3.8, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_11_txt_mod_lin | sum(S) retained: 59.7%, fro retained: 97.4%, max(S) ratio: 4.7, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_12_img_attn_proj | sum(S) retained: 55.0%, fro retained: 93.1%, max(S) ratio: 2.9, dynamic | dim: 8, alpha: 8.0
lora_unet_double_blocks_12_img_attn_qkv | sum(S) retained: 50.6%, fro retained: 93.6%, max(S) ratio: 3.8, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_12_img_mlp_0 | sum(S) retained: 50.6%, fro retained: 93.8%, max(S) ratio: 4.0, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_12_img_mlp_2 | sum(S) retained: 53.7%, fro retained: 92.6%, max(S) ratio: 2.4, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_12_img_mod_lin | sum(S) retained: 53.7%, fro retained: 95.2%, max(S) ratio: 3.4, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_12_txt_attn_proj | sum(S) retained: 54.0%, fro retained: 95.9%, max(S) ratio: 4.0, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_12_txt_attn_qkv | sum(S) retained: 49.5%, fro retained: 93.6%, max(S) ratio: 3.2, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_12_txt_mlp_0 | sum(S) retained: 68.2%, fro retained: 99.0%, max(S) ratio: 2.6, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_12_txt_mlp_2 | sum(S) retained: 60.2%, fro retained: 96.9%, max(S) ratio: 2.6, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_12_txt_mod_lin | sum(S) retained: 60.6%, fro retained: 98.0%, max(S) ratio: 3.6, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_13_img_attn_proj | sum(S) retained: 50.5%, fro retained: 92.9%, max(S) ratio: 2.7, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_13_img_attn_qkv | sum(S) retained: 48.1%, fro retained: 92.8%, max(S) ratio: 3.8, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_13_img_mlp_0 | sum(S) retained: 48.9%, fro retained: 92.5%, max(S) ratio: 3.0, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_13_img_mlp_2 | sum(S) retained: 53.6%, fro retained: 93.8%, max(S) ratio: 3.7, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_13_img_mod_lin | sum(S) retained: 55.0%, fro retained: 95.6%, max(S) ratio: 3.0, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_13_txt_attn_proj | sum(S) retained: 53.4%, fro retained: 96.3%, max(S) ratio: 3.7, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_13_txt_attn_qkv | sum(S) retained: 55.8%, fro retained: 95.9%, max(S) ratio: 3.2, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_13_txt_mlp_0 | sum(S) retained: 61.9%, fro retained: 97.8%, max(S) ratio: 3.3, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_13_txt_mlp_2 | sum(S) retained: 58.5%, fro retained: 97.5%, max(S) ratio: 3.5, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_13_txt_mod_lin | sum(S) retained: 61.4%, fro retained: 97.1%, max(S) ratio: 3.0, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_14_img_attn_proj | sum(S) retained: 52.6%, fro retained: 91.6%, max(S) ratio: 3.1, dynamic | dim: 9, alpha: 9.0
lora_unet_double_blocks_14_img_attn_qkv | sum(S) retained: 51.7%, fro retained: 92.2%, max(S) ratio: 3.3, dynamic | dim: 8, alpha: 8.0
lora_unet_double_blocks_14_img_mlp_0 | sum(S) retained: 48.5%, fro retained: 92.2%, max(S) ratio: 3.7, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_14_img_mlp_2 | sum(S) retained: 51.3%, fro retained: 92.2%, max(S) ratio: 3.2, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_14_img_mod_lin | sum(S) retained: 54.7%, fro retained: 95.6%, max(S) ratio: 4.4, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_14_txt_attn_proj | sum(S) retained: 53.2%, fro retained: 95.9%, max(S) ratio: 3.9, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_14_txt_attn_qkv | sum(S) retained: 50.9%, fro retained: 95.3%, max(S) ratio: 3.8, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_14_txt_mlp_0 | sum(S) retained: 65.8%, fro retained: 98.8%, max(S) ratio: 3.5, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_14_txt_mlp_2 | sum(S) retained: 55.8%, fro retained: 95.9%, max(S) ratio: 3.4, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_14_txt_mod_lin | sum(S) retained: 72.6%, fro retained: 99.1%, max(S) ratio: 6.7, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_15_img_attn_proj | sum(S) retained: 51.9%, fro retained: 91.7%, max(S) ratio: 3.4, dynamic | dim: 9, alpha: 9.0
lora_unet_double_blocks_15_img_attn_qkv | sum(S) retained: 49.1%, fro retained: 92.1%, max(S) ratio: 4.4, dynamic | dim: 8, alpha: 8.0
lora_unet_double_blocks_15_img_mlp_0 | sum(S) retained: 51.2%, fro retained: 92.4%, max(S) ratio: 3.5, dynamic | dim: 8, alpha: 8.0
lora_unet_double_blocks_15_img_mlp_2 | sum(S) retained: 52.5%, fro retained: 92.8%, max(S) ratio: 3.7, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_15_img_mod_lin | sum(S) retained: 50.8%, fro retained: 93.9%, max(S) ratio: 3.3, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_15_txt_attn_proj | sum(S) retained: 64.5%, fro retained: 98.2%, max(S) ratio: 3.5, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_15_txt_attn_qkv | sum(S) retained: 58.5%, fro retained: 96.1%, max(S) ratio: 2.8, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_15_txt_mlp_0 | sum(S) retained: 43.2%, fro retained: 94.0%, max(S) ratio: 3.7, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_15_txt_mlp_2 | sum(S) retained: 52.8%, fro retained: 94.2%, max(S) ratio: 2.8, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_15_txt_mod_lin | sum(S) retained: 58.6%, fro retained: 97.7%, max(S) ratio: 3.2, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_16_img_attn_proj | sum(S) retained: 54.7%, fro retained: 92.4%, max(S) ratio: 3.1, dynamic | dim: 10, alpha: 10.0
lora_unet_double_blocks_16_img_attn_qkv | sum(S) retained: 51.0%, fro retained: 92.5%, max(S) ratio: 4.0, dynamic | dim: 8, alpha: 8.0
lora_unet_double_blocks_16_img_mlp_0 | sum(S) retained: 48.7%, fro retained: 91.8%, max(S) ratio: 3.1, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_16_img_mlp_2 | sum(S) retained: 55.3%, fro retained: 93.7%, max(S) ratio: 2.9, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_16_img_mod_lin | sum(S) retained: 48.1%, fro retained: 93.0%, max(S) ratio: 3.7, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_16_txt_attn_proj | sum(S) retained: 50.6%, fro retained: 95.1%, max(S) ratio: 3.7, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_16_txt_attn_qkv | sum(S) retained: 51.4%, fro retained: 94.4%, max(S) ratio: 3.8, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_16_txt_mlp_0 | sum(S) retained: 47.8%, fro retained: 94.8%, max(S) ratio: 4.6, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_16_txt_mlp_2 | sum(S) retained: 52.7%, fro retained: 93.8%, max(S) ratio: 3.3, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_16_txt_mod_lin | sum(S) retained: 56.8%, fro retained: 96.9%, max(S) ratio: 4.9, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_17_img_attn_proj | sum(S) retained: 56.4%, fro retained: 92.6%, max(S) ratio: 3.3, dynamic | dim: 11, alpha: 11.0
lora_unet_double_blocks_17_img_attn_qkv | sum(S) retained: 49.8%, fro retained: 91.7%, max(S) ratio: 3.1, dynamic | dim: 8, alpha: 8.0
lora_unet_double_blocks_17_img_mlp_0 | sum(S) retained: 45.9%, fro retained: 93.0%, max(S) ratio: 3.8, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_17_img_mlp_2 | sum(S) retained: 58.8%, fro retained: 94.0%, max(S) ratio: 2.7, dynamic | dim: 8, alpha: 8.0
lora_unet_double_blocks_17_img_mod_lin | sum(S) retained: 54.1%, fro retained: 95.0%, max(S) ratio: 2.5, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_17_txt_attn_proj | sum(S) retained: 54.3%, fro retained: 94.8%, max(S) ratio: 2.9, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_17_txt_attn_qkv | sum(S) retained: 54.8%, fro retained: 96.2%, max(S) ratio: 3.9, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_17_txt_mlp_0 | sum(S) retained: 50.4%, fro retained: 94.4%, max(S) ratio: 3.5, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_17_txt_mlp_2 | sum(S) retained: 58.2%, fro retained: 94.7%, max(S) ratio: 3.2, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_17_txt_mod_lin | sum(S) retained: 66.8%, fro retained: 98.4%, max(S) ratio: 5.9, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_18_img_attn_proj | sum(S) retained: 53.8%, fro retained: 91.7%, max(S) ratio: 2.8, dynamic | dim: 9, alpha: 9.0
lora_unet_double_blocks_18_img_attn_qkv | sum(S) retained: 53.7%, fro retained: 92.4%, max(S) ratio: 3.6, dynamic | dim: 10, alpha: 10.0
lora_unet_double_blocks_18_img_mlp_0 | sum(S) retained: 49.8%, fro retained: 92.1%, max(S) ratio: 4.2, dynamic | dim: 8, alpha: 8.0
lora_unet_double_blocks_18_img_mlp_2 | sum(S) retained: 55.0%, fro retained: 93.4%, max(S) ratio: 3.4, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_18_img_mod_lin | sum(S) retained: 56.6%, fro retained: 94.7%, max(S) ratio: 3.0, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_18_txt_attn_proj | sum(S) retained: 65.5%, fro retained: 98.7%, max(S) ratio: 3.7, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_18_txt_attn_qkv | sum(S) retained: 54.3%, fro retained: 95.7%, max(S) ratio: 4.0, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_18_txt_mlp_0 | sum(S) retained: 62.6%, fro retained: 98.5%, max(S) ratio: 3.8, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_18_txt_mlp_2 | sum(S) retained: 64.1%, fro retained: 97.7%, max(S) ratio: 4.8, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_18_txt_mod_lin | sum(S) retained: 56.8%, fro retained: 96.4%, max(S) ratio: 3.0, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_1_img_attn_proj | sum(S) retained: 52.7%, fro retained: 93.7%, max(S) ratio: 3.5, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_1_img_attn_qkv | sum(S) retained: 52.9%, fro retained: 95.6%, max(S) ratio: 4.5, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_1_img_mlp_0 | sum(S) retained: 62.7%, fro retained: 90.8%, max(S) ratio: 3.1, dynamic | dim: 24, alpha: 24.0
lora_unet_double_blocks_1_img_mlp_2 | sum(S) retained: 55.7%, fro retained: 92.4%, max(S) ratio: 3.1, dynamic | dim: 9, alpha: 9.0
lora_unet_double_blocks_1_img_mod_lin | sum(S) retained: 52.1%, fro retained: 91.1%, max(S) ratio: 3.9, dynamic | dim: 12, alpha: 12.0
lora_unet_double_blocks_1_txt_attn_proj | sum(S) retained: 56.1%, fro retained: 97.5%, max(S) ratio: 3.2, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_1_txt_attn_qkv | sum(S) retained: 52.5%, fro retained: 95.8%, max(S) ratio: 3.2, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_1_txt_mlp_0 | sum(S) retained: 63.6%, fro retained: 98.6%, max(S) ratio: 5.3, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_1_txt_mlp_2 | sum(S) retained: 67.6%, fro retained: 98.6%, max(S) ratio: 3.5, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_1_txt_mod_lin | sum(S) retained: 68.1%, fro retained: 98.6%, max(S) ratio: 5.6, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_2_img_attn_proj | sum(S) retained: 50.0%, fro retained: 92.3%, max(S) ratio: 3.9, dynamic | dim: 8, alpha: 8.0
lora_unet_double_blocks_2_img_attn_qkv | sum(S) retained: 48.1%, fro retained: 91.7%, max(S) ratio: 3.4, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_2_img_mlp_0 | sum(S) retained: 61.3%, fro retained: 91.1%, max(S) ratio: 3.8, dynamic | dim: 22, alpha: 22.0
lora_unet_double_blocks_2_img_mlp_2 | sum(S) retained: 54.7%, fro retained: 92.5%, max(S) ratio: 3.6, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_2_img_mod_lin | sum(S) retained: 48.4%, fro retained: 91.8%, max(S) ratio: 3.9, dynamic | dim: 8, alpha: 8.0
lora_unet_double_blocks_2_txt_attn_proj | sum(S) retained: 62.5%, fro retained: 97.1%, max(S) ratio: 3.0, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_2_txt_attn_qkv | sum(S) retained: 60.9%, fro retained: 97.4%, max(S) ratio: 3.2, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_2_txt_mlp_0 | sum(S) retained: 56.6%, fro retained: 97.2%, max(S) ratio: 4.7, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_2_txt_mlp_2 | sum(S) retained: 57.3%, fro retained: 95.3%, max(S) ratio: 3.4, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_2_txt_mod_lin | sum(S) retained: 54.8%, fro retained: 96.4%, max(S) ratio: 3.3, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_3_img_attn_proj | sum(S) retained: 49.5%, fro retained: 92.5%, max(S) ratio: 3.5, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_3_img_attn_qkv | sum(S) retained: 49.8%, fro retained: 92.4%, max(S) ratio: 3.7, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_3_img_mlp_0 | sum(S) retained: 51.5%, fro retained: 92.5%, max(S) ratio: 2.8, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_3_img_mlp_2 | sum(S) retained: 62.4%, fro retained: 95.8%, max(S) ratio: 3.6, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_3_img_mod_lin | sum(S) retained: 54.2%, fro retained: 94.6%, max(S) ratio: 2.7, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_3_txt_attn_proj | sum(S) retained: 51.7%, fro retained: 95.9%, max(S) ratio: 3.0, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_3_txt_attn_qkv | sum(S) retained: 60.3%, fro retained: 96.5%, max(S) ratio: 2.3, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_3_txt_mlp_0 | sum(S) retained: 57.7%, fro retained: 97.0%, max(S) ratio: 3.3, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_3_txt_mlp_2 | sum(S) retained: 49.9%, fro retained: 96.1%, max(S) ratio: 5.4, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_3_txt_mod_lin | sum(S) retained: 54.6%, fro retained: 96.1%, max(S) ratio: 3.4, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_4_img_attn_proj | sum(S) retained: 55.2%, fro retained: 91.9%, max(S) ratio: 2.8, dynamic | dim: 11, alpha: 11.0
lora_unet_double_blocks_4_img_attn_qkv | sum(S) retained: 50.8%, fro retained: 91.3%, max(S) ratio: 3.4, dynamic | dim: 9, alpha: 9.0
lora_unet_double_blocks_4_img_mlp_0 | sum(S) retained: 50.0%, fro retained: 93.1%, max(S) ratio: 3.6, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_4_img_mlp_2 | sum(S) retained: 54.3%, fro retained: 93.5%, max(S) ratio: 2.6, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_4_img_mod_lin | sum(S) retained: 55.6%, fro retained: 95.5%, max(S) ratio: 3.0, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_4_txt_attn_proj | sum(S) retained: 56.7%, fro retained: 97.1%, max(S) ratio: 4.4, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_4_txt_attn_qkv | sum(S) retained: 56.0%, fro retained: 97.4%, max(S) ratio: 4.8, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_4_txt_mlp_0 | sum(S) retained: 61.5%, fro retained: 97.6%, max(S) ratio: 4.0, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_4_txt_mlp_2 | sum(S) retained: 52.7%, fro retained: 96.2%, max(S) ratio: 3.6, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_4_txt_mod_lin | sum(S) retained: 53.4%, fro retained: 96.2%, max(S) ratio: 2.7, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_5_img_attn_proj | sum(S) retained: 48.9%, fro retained: 92.4%, max(S) ratio: 3.5, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_5_img_attn_qkv | sum(S) retained: 50.3%, fro retained: 92.5%, max(S) ratio: 4.0, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_5_img_mlp_0 | sum(S) retained: 50.1%, fro retained: 94.3%, max(S) ratio: 3.3, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_5_img_mlp_2 | sum(S) retained: 51.1%, fro retained: 93.0%, max(S) ratio: 3.3, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_5_img_mod_lin | sum(S) retained: 52.4%, fro retained: 94.3%, max(S) ratio: 4.0, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_5_txt_attn_proj | sum(S) retained: 57.4%, fro retained: 97.2%, max(S) ratio: 4.7, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_5_txt_attn_qkv | sum(S) retained: 54.5%, fro retained: 96.4%, max(S) ratio: 3.5, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_5_txt_mlp_0 | sum(S) retained: 54.1%, fro retained: 96.1%, max(S) ratio: 2.9, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_5_txt_mlp_2 | sum(S) retained: 60.9%, fro retained: 98.4%, max(S) ratio: 4.6, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_5_txt_mod_lin | sum(S) retained: 55.0%, fro retained: 95.6%, max(S) ratio: 3.8, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_6_img_attn_proj | sum(S) retained: 50.6%, fro retained: 93.7%, max(S) ratio: 3.0, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_6_img_attn_qkv | sum(S) retained: 46.1%, fro retained: 94.1%, max(S) ratio: 4.5, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_6_img_mlp_0 | sum(S) retained: 48.0%, fro retained: 92.7%, max(S) ratio: 3.5, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_6_img_mlp_2 | sum(S) retained: 53.8%, fro retained: 94.4%, max(S) ratio: 3.6, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_6_img_mod_lin | sum(S) retained: 51.6%, fro retained: 95.4%, max(S) ratio: 3.6, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_6_txt_attn_proj | sum(S) retained: 56.3%, fro retained: 96.3%, max(S) ratio: 3.9, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_6_txt_attn_qkv | sum(S) retained: 55.6%, fro retained: 95.8%, max(S) ratio: 3.8, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_6_txt_mlp_0 | sum(S) retained: 56.8%, fro retained: 98.2%, max(S) ratio: 7.2, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_6_txt_mlp_2 | sum(S) retained: 61.4%, fro retained: 97.1%, max(S) ratio: 3.1, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_6_txt_mod_lin | sum(S) retained: 58.9%, fro retained: 98.0%, max(S) ratio: 4.5, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_7_img_attn_proj | sum(S) retained: 53.9%, fro retained: 93.0%, max(S) ratio: 3.1, dynamic | dim: 8, alpha: 8.0
lora_unet_double_blocks_7_img_attn_qkv | sum(S) retained: 46.9%, fro retained: 93.1%, max(S) ratio: 3.3, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_7_img_mlp_0 | sum(S) retained: 46.4%, fro retained: 92.7%, max(S) ratio: 3.0, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_7_img_mlp_2 | sum(S) retained: 53.4%, fro retained: 93.9%, max(S) ratio: 3.7, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_7_img_mod_lin | sum(S) retained: 55.7%, fro retained: 94.5%, max(S) ratio: 2.5, dynamic | dim: 5, alpha: 5.0
lora_unet_double_blocks_7_txt_attn_proj | sum(S) retained: 61.8%, fro retained: 98.1%, max(S) ratio: 5.6, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_7_txt_attn_qkv | sum(S) retained: 60.4%, fro retained: 97.8%, max(S) ratio: 5.5, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_7_txt_mlp_0 | sum(S) retained: 57.0%, fro retained: 98.3%, max(S) ratio: 6.7, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_7_txt_mlp_2 | sum(S) retained: 49.1%, fro retained: 95.5%, max(S) ratio: 3.8, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_7_txt_mod_lin | sum(S) retained: 67.9%, fro retained: 98.7%, max(S) ratio: 4.9, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_8_img_attn_proj | sum(S) retained: 50.3%, fro retained: 92.0%, max(S) ratio: 3.4, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_8_img_attn_qkv | sum(S) retained: 51.5%, fro retained: 93.4%, max(S) ratio: 3.6, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_8_img_mlp_0 | sum(S) retained: 50.6%, fro retained: 92.4%, max(S) ratio: 3.6, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_8_img_mlp_2 | sum(S) retained: 55.5%, fro retained: 94.4%, max(S) ratio: 3.3, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_8_img_mod_lin | sum(S) retained: 53.2%, fro retained: 94.8%, max(S) ratio: 3.0, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_8_txt_attn_proj | sum(S) retained: 61.6%, fro retained: 98.5%, max(S) ratio: 5.1, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_8_txt_attn_qkv | sum(S) retained: 49.4%, fro retained: 96.1%, max(S) ratio: 3.3, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_8_txt_mlp_0 | sum(S) retained: 54.9%, fro retained: 96.0%, max(S) ratio: 4.2, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_8_txt_mlp_2 | sum(S) retained: 53.0%, fro retained: 95.8%, max(S) ratio: 4.1, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_8_txt_mod_lin | sum(S) retained: 57.4%, fro retained: 97.3%, max(S) ratio: 4.1, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_9_img_attn_proj | sum(S) retained: 53.2%, fro retained: 93.4%, max(S) ratio: 3.4, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_9_img_attn_qkv | sum(S) retained: 50.4%, fro retained: 93.2%, max(S) ratio: 3.7, dynamic | dim: 6, alpha: 6.0
lora_unet_double_blocks_9_img_mlp_0 | sum(S) retained: 50.0%, fro retained: 92.7%, max(S) ratio: 3.9, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_9_img_mlp_2 | sum(S) retained: 55.7%, fro retained: 93.6%, max(S) ratio: 3.2, dynamic | dim: 7, alpha: 7.0
lora_unet_double_blocks_9_img_mod_lin | sum(S) retained: 52.1%, fro retained: 95.0%, max(S) ratio: 3.5, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_9_txt_attn_proj | sum(S) retained: 60.2%, fro retained: 97.5%, max(S) ratio: 2.8, dynamic | dim: 3, alpha: 3.0
lora_unet_double_blocks_9_txt_attn_qkv | sum(S) retained: 55.8%, fro retained: 95.8%, max(S) ratio: 3.2, dynamic | dim: 4, alpha: 4.0
lora_unet_double_blocks_9_txt_mlp_0 | sum(S) retained: 63.0%, fro retained: 98.4%, max(S) ratio: 2.7, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_9_txt_mlp_2 | sum(S) retained: 50.8%, fro retained: 96.2%, max(S) ratio: 5.4, dynamic | dim: 2, alpha: 2.0
lora_unet_double_blocks_9_txt_mod_lin | sum(S) retained: 58.0%, fro retained: 96.6%, max(S) ratio: 3.2, dynamic | dim: 3, alpha: 3.0
lora_unet_single_blocks_0_linear1 | sum(S) retained: 38.0%, fro retained: 92.6%, max(S) ratio: 4.4, dynamic | dim: 3, alpha: 3.0
lora_unet_single_blocks_0_linear2 | sum(S) retained: 53.0%, fro retained: 95.0%, max(S) ratio: 3.9, dynamic | dim: 4, alpha: 4.0
lora_unet_single_blocks_0_modulation_lin | sum(S) retained: 51.1%, fro retained: 93.0%, max(S) ratio: 2.9, dynamic | dim: 5, alpha: 5.0
lora_unet_single_blocks_10_linear1 | sum(S) retained: 59.5%, fro retained: 91.6%, max(S) ratio: 3.5, dynamic | dim: 18, alpha: 18.0
lora_unet_single_blocks_10_linear2 | sum(S) retained: 54.5%, fro retained: 91.6%, max(S) ratio: 3.0, dynamic | dim: 9, alpha: 9.0
lora_unet_single_blocks_10_modulation_lin | sum(S) retained: 53.1%, fro retained: 92.7%, max(S) ratio: 3.2, dynamic | dim: 7, alpha: 7.0
lora_unet_single_blocks_11_linear1 | sum(S) retained: 58.1%, fro retained: 91.0%, max(S) ratio: 3.2, dynamic | dim: 17, alpha: 17.0
lora_unet_single_blocks_11_linear2 | sum(S) retained: 58.6%, fro retained: 92.1%, max(S) ratio: 2.8, dynamic | dim: 12, alpha: 12.0
lora_unet_single_blocks_11_modulation_lin | sum(S) retained: 52.1%, fro retained: 92.0%, max(S) ratio: 3.0, dynamic | dim: 7, alpha: 7.0
lora_unet_single_blocks_12_linear1 | sum(S) retained: 61.2%, fro retained: 91.4%, max(S) ratio: 3.1, dynamic | dim: 20, alpha: 20.0
lora_unet_single_blocks_12_linear2 | sum(S) retained: 56.4%, fro retained: 91.9%, max(S) ratio: 3.1, dynamic | dim: 11, alpha: 11.0
lora_unet_single_blocks_12_modulation_lin | sum(S) retained: 52.9%, fro retained: 92.1%, max(S) ratio: 3.4, dynamic | dim: 8, alpha: 8.0
lora_unet_single_blocks_13_linear1 | sum(S) retained: 61.4%, fro retained: 91.2%, max(S) ratio: 3.0, dynamic | dim: 21, alpha: 21.0
lora_unet_single_blocks_13_linear2 | sum(S) retained: 57.2%, fro retained: 92.1%, max(S) ratio: 3.6, dynamic | dim: 12, alpha: 12.0
lora_unet_single_blocks_13_modulation_lin | sum(S) retained: 52.0%, fro retained: 92.8%, max(S) ratio: 4.3, dynamic | dim: 7, alpha: 7.0
lora_unet_single_blocks_14_linear1 | sum(S) retained: 62.1%, fro retained: 91.2%, max(S) ratio: 3.1, dynamic | dim: 22, alpha: 22.0
lora_unet_single_blocks_14_linear2 | sum(S) retained: 56.2%, fro retained: 91.5%, max(S) ratio: 3.2, dynamic | dim: 12, alpha: 12.0
lora_unet_single_blocks_14_modulation_lin | sum(S) retained: 50.9%, fro retained: 91.8%, max(S) ratio: 3.6, dynamic | dim: 7, alpha: 7.0
lora_unet_single_blocks_15_linear1 | sum(S) retained: 61.4%, fro retained: 91.2%, max(S) ratio: 3.3, dynamic | dim: 21, alpha: 21.0
lora_unet_single_blocks_15_linear2 | sum(S) retained: 56.4%, fro retained: 91.5%, max(S) ratio: 3.1, dynamic | dim: 12, alpha: 12.0
lora_unet_single_blocks_15_modulation_lin | sum(S) retained: 53.1%, fro retained: 92.6%, max(S) ratio: 3.3, dynamic | dim: 8, alpha: 8.0
lora_unet_single_blocks_16_linear1 | sum(S) retained: 62.2%, fro retained: 90.8%, max(S) ratio: 2.8, dynamic | dim: 23, alpha: 23.0
lora_unet_single_blocks_16_linear2 | sum(S) retained: 59.0%, fro retained: 92.1%, max(S) ratio: 3.1, dynamic | dim: 14, alpha: 14.0
lora_unet_single_blocks_16_modulation_lin | sum(S) retained: 50.3%, fro retained: 91.5%, max(S) ratio: 3.3, dynamic | dim: 7, alpha: 7.0
lora_unet_single_blocks_17_linear1 | sum(S) retained: 62.9%, fro retained: 90.7%, max(S) ratio: 3.2, dynamic | dim: 24, alpha: 24.0
lora_unet_single_blocks_17_linear2 | sum(S) retained: 58.6%, fro retained: 91.8%, max(S) ratio: 3.0, dynamic | dim: 14, alpha: 14.0
lora_unet_single_blocks_17_modulation_lin | sum(S) retained: 54.6%, fro retained: 92.6%, max(S) ratio: 3.6, dynamic | dim: 9, alpha: 9.0
lora_unet_single_blocks_18_linear1 | sum(S) retained: 63.5%, fro retained: 91.2%, max(S) ratio: 3.0, dynamic | dim: 24, alpha: 24.0
lora_unet_single_blocks_18_linear2 | sum(S) retained: 58.0%, fro retained: 91.8%, max(S) ratio: 3.2, dynamic | dim: 14, alpha: 14.0
lora_unet_single_blocks_18_modulation_lin | sum(S) retained: 50.0%, fro retained: 92.0%, max(S) ratio: 3.9, dynamic | dim: 7, alpha: 7.0
lora_unet_single_blocks_19_linear1 | sum(S) retained: 63.2%, fro retained: 91.0%, max(S) ratio: 2.9, dynamic | dim: 24, alpha: 24.0
lora_unet_single_blocks_19_linear2 | sum(S) retained: 57.0%, fro retained: 91.8%, max(S) ratio: 3.1, dynamic | dim: 13, alpha: 13.0
lora_unet_single_blocks_19_modulation_lin | sum(S) retained: 51.1%, fro retained: 92.4%, max(S) ratio: 3.6, dynamic | dim: 7, alpha: 7.0
lora_unet_single_blocks_1_linear1 | sum(S) retained: 48.1%, fro retained: 92.4%, max(S) ratio: 4.6, dynamic | dim: 8, alpha: 8.0
lora_unet_single_blocks_1_linear2 | sum(S) retained: 49.3%, fro retained: 93.2%, max(S) ratio: 3.7, dynamic | dim: 4, alpha: 4.0
lora_unet_single_blocks_1_modulation_lin | sum(S) retained: 52.3%, fro retained: 94.7%, max(S) ratio: 3.5, dynamic | dim: 4, alpha: 4.0
lora_unet_single_blocks_20_linear1 | sum(S) retained: 64.7%, fro retained: 91.1%, max(S) ratio: 2.7, dynamic | dim: 26, alpha: 26.0
lora_unet_single_blocks_20_linear2 | sum(S) retained: 55.8%, fro retained: 91.1%, max(S) ratio: 3.2, dynamic | dim: 13, alpha: 13.0
lora_unet_single_blocks_20_modulation_lin | sum(S) retained: 55.7%, fro retained: 92.2%, max(S) ratio: 3.1, dynamic | dim: 10, alpha: 10.0
lora_unet_single_blocks_21_linear1 | sum(S) retained: 63.4%, fro retained: 90.7%, max(S) ratio: 2.8, dynamic | dim: 25, alpha: 25.0
lora_unet_single_blocks_21_linear2 | sum(S) retained: 56.5%, fro retained: 92.1%, max(S) ratio: 3.1, dynamic | dim: 12, alpha: 12.0
lora_unet_single_blocks_21_modulation_lin | sum(S) retained: 52.4%, fro retained: 92.1%, max(S) ratio: 3.4, dynamic | dim: 8, alpha: 8.0
lora_unet_single_blocks_22_linear1 | sum(S) retained: 65.6%, fro retained: 91.3%, max(S) ratio: 3.1, dynamic | dim: 27, alpha: 27.0
lora_unet_single_blocks_22_linear2 | sum(S) retained: 55.7%, fro retained: 91.2%, max(S) ratio: 3.1, dynamic | dim: 13, alpha: 13.0
lora_unet_single_blocks_22_modulation_lin | sum(S) retained: 51.7%, fro retained: 92.0%, max(S) ratio: 3.2, dynamic | dim: 8, alpha: 8.0
lora_unet_single_blocks_23_linear1 | sum(S) retained: 65.1%, fro retained: 90.8%, max(S) ratio: 2.7, dynamic | dim: 27, alpha: 27.0
lora_unet_single_blocks_23_linear2 | sum(S) retained: 57.3%, fro retained: 91.6%, max(S) ratio: 3.1, dynamic | dim: 14, alpha: 14.0
lora_unet_single_blocks_23_modulation_lin | sum(S) retained: 53.5%, fro retained: 92.2%, max(S) ratio: 3.8, dynamic | dim: 9, alpha: 9.0
lora_unet_single_blocks_24_linear1 | sum(S) retained: 65.8%, fro retained: 90.7%, max(S) ratio: 2.4, dynamic | dim: 28, alpha: 28.0
lora_unet_single_blocks_24_linear2 | sum(S) retained: 58.2%, fro retained: 91.8%, max(S) ratio: 3.3, dynamic | dim: 15, alpha: 15.0
lora_unet_single_blocks_24_modulation_lin | sum(S) retained: 52.2%, fro retained: 92.6%, max(S) ratio: 3.9, dynamic | dim: 8, alpha: 8.0
lora_unet_single_blocks_25_linear1 | sum(S) retained: 65.7%, fro retained: 90.7%, max(S) ratio: 2.6, dynamic | dim: 28, alpha: 28.0
lora_unet_single_blocks_25_linear2 | sum(S) retained: 59.2%, fro retained: 91.7%, max(S) ratio: 3.4, dynamic | dim: 17, alpha: 17.0
lora_unet_single_blocks_25_modulation_lin | sum(S) retained: 50.8%, fro retained: 93.6%, max(S) ratio: 3.3, dynamic | dim: 6, alpha: 6.0
lora_unet_single_blocks_26_linear1 | sum(S) retained: 67.8%, fro retained: 91.3%, max(S) ratio: 2.5, dynamic | dim: 30, alpha: 30.0
lora_unet_single_blocks_26_linear2 | sum(S) retained: 59.0%, fro retained: 90.9%, max(S) ratio: 3.3, dynamic | dim: 18, alpha: 18.0
lora_unet_single_blocks_26_modulation_lin | sum(S) retained: 49.5%, fro retained: 93.3%, max(S) ratio: 4.1, dynamic | dim: 6, alpha: 6.0
lora_unet_single_blocks_27_linear1 | sum(S) retained: 68.1%, fro retained: 91.0%, max(S) ratio: 2.4, dynamic | dim: 31, alpha: 31.0
lora_unet_single_blocks_27_linear2 | sum(S) retained: 60.9%, fro retained: 91.0%, max(S) ratio: 3.1, dynamic | dim: 20, alpha: 20.0
lora_unet_single_blocks_27_modulation_lin | sum(S) retained: 47.5%, fro retained: 92.6%, max(S) ratio: 4.8, dynamic | dim: 6, alpha: 6.0
lora_unet_single_blocks_28_linear1 | sum(S) retained: 68.9%, fro retained: 91.1%, max(S) ratio: 2.4, dynamic | dim: 32, alpha: 32.0
lora_unet_single_blocks_28_linear2 | sum(S) retained: 60.4%, fro retained: 90.8%, max(S) ratio: 3.1, dynamic | dim: 20, alpha: 20.0
lora_unet_single_blocks_28_modulation_lin | sum(S) retained: 49.6%, fro retained: 91.8%, max(S) ratio: 3.9, dynamic | dim: 7, alpha: 7.0
lora_unet_single_blocks_29_linear1 | sum(S) retained: 68.8%, fro retained: 91.1%, max(S) ratio: 2.4, dynamic | dim: 32, alpha: 32.0
lora_unet_single_blocks_29_linear2 | sum(S) retained: 60.4%, fro retained: 91.0%, max(S) ratio: 2.9, dynamic | dim: 19, alpha: 19.0
lora_unet_single_blocks_29_modulation_lin | sum(S) retained: 50.7%, fro retained: 92.4%, max(S) ratio: 4.4, dynamic | dim: 8, alpha: 8.0
lora_unet_single_blocks_2_linear1 | sum(S) retained: 48.1%, fro retained: 91.6%, max(S) ratio: 4.1, dynamic | dim: 8, alpha: 8.0
lora_unet_single_blocks_2_linear2 | sum(S) retained: 53.9%, fro retained: 92.6%, max(S) ratio: 2.8, dynamic | dim: 6, alpha: 6.0
lora_unet_single_blocks_2_modulation_lin | sum(S) retained: 52.9%, fro retained: 93.7%, max(S) ratio: 2.4, dynamic | dim: 5, alpha: 5.0
lora_unet_single_blocks_30_linear1 | sum(S) retained: 67.9%, fro retained: 90.3%, max(S) ratio: 2.3, dynamic | dim: 32, alpha: 32.0
lora_unet_single_blocks_30_linear2 | sum(S) retained: 61.2%, fro retained: 91.1%, max(S) ratio: 3.0, dynamic | dim: 20, alpha: 20.0
lora_unet_single_blocks_30_modulation_lin | sum(S) retained: 48.0%, fro retained: 92.1%, max(S) ratio: 3.7, dynamic | dim: 6, alpha: 6.0
lora_unet_single_blocks_31_linear1 | sum(S) retained: 68.0%, fro retained: 90.5%, max(S) ratio: 2.6, dynamic | dim: 32, alpha: 32.0
lora_unet_single_blocks_31_linear2 | sum(S) retained: 60.6%, fro retained: 91.1%, max(S) ratio: 3.4, dynamic | dim: 19, alpha: 19.0
lora_unet_single_blocks_31_modulation_lin | sum(S) retained: 46.5%, fro retained: 91.8%, max(S) ratio: 4.3, dynamic | dim: 6, alpha: 6.0
lora_unet_single_blocks_32_linear1 | sum(S) retained: 68.4%, fro retained: 90.8%, max(S) ratio: 2.4, dynamic | dim: 32, alpha: 32.0
lora_unet_single_blocks_32_linear2 | sum(S) retained: 62.6%, fro retained: 91.5%, max(S) ratio: 2.6, dynamic | dim: 20, alpha: 20.0
lora_unet_single_blocks_32_modulation_lin | sum(S) retained: 42.9%, fro retained: 92.8%, max(S) ratio: 4.2, dynamic | dim: 4, alpha: 4.0
lora_unet_single_blocks_33_linear1 | sum(S) retained: 67.3%, fro retained: 90.0%, max(S) ratio: 2.6, dynamic | dim: 32, alpha: 32.0
lora_unet_single_blocks_33_linear2 | sum(S) retained: 62.8%, fro retained: 91.1%, max(S) ratio: 2.6, dynamic | dim: 21, alpha: 21.0
lora_unet_single_blocks_33_modulation_lin | sum(S) retained: 48.0%, fro retained: 93.3%, max(S) ratio: 4.2, dynamic | dim: 5, alpha: 5.0
lora_unet_single_blocks_34_linear1 | sum(S) retained: 67.5%, fro retained: 90.3%, max(S) ratio: 2.6, dynamic | dim: 32, alpha: 32.0
lora_unet_single_blocks_34_linear2 | sum(S) retained: 62.7%, fro retained: 91.3%, max(S) ratio: 3.0, dynamic | dim: 20, alpha: 20.0
lora_unet_single_blocks_34_modulation_lin | sum(S) retained: 48.0%, fro retained: 92.9%, max(S) ratio: 3.6, dynamic | dim: 5, alpha: 5.0
lora_unet_single_blocks_35_linear1 | sum(S) retained: 68.8%, fro retained: 90.9%, max(S) ratio: 2.3, dynamic | dim: 32, alpha: 32.0
lora_unet_single_blocks_35_linear2 | sum(S) retained: 62.8%, fro retained: 91.6%, max(S) ratio: 2.6, dynamic | dim: 18, alpha: 18.0
lora_unet_single_blocks_35_modulation_lin | sum(S) retained: 46.4%, fro retained: 92.5%, max(S) ratio: 4.0, dynamic | dim: 5, alpha: 5.0
lora_unet_single_blocks_36_linear1 | sum(S) retained: 66.8%, fro retained: 91.1%, max(S) ratio: 2.6, dynamic | dim: 29, alpha: 29.0
lora_unet_single_blocks_36_linear2 | sum(S) retained: 61.2%, fro retained: 91.7%, max(S) ratio: 3.4, dynamic | dim: 15, alpha: 15.0
lora_unet_single_blocks_36_modulation_lin | sum(S) retained: 49.9%, fro retained: 93.3%, max(S) ratio: 3.8, dynamic | dim: 6, alpha: 6.0
lora_unet_single_blocks_37_linear1 | sum(S) retained: 51.8%, fro retained: 91.6%, max(S) ratio: 4.7, dynamic | dim: 12, alpha: 12.0
lora_unet_single_blocks_37_linear2 | sum(S) retained: 57.2%, fro retained: 95.1%, max(S) ratio: 3.0, dynamic | dim: 5, alpha: 5.0
lora_unet_single_blocks_37_modulation_lin | sum(S) retained: 47.5%, fro retained: 92.8%, max(S) ratio: 4.2, dynamic | dim: 6, alpha: 6.0
lora_unet_single_blocks_3_linear1 | sum(S) retained: 45.2%, fro retained: 92.2%, max(S) ratio: 5.4, dynamic | dim: 7, alpha: 7.0
lora_unet_single_blocks_3_linear2 | sum(S) retained: 55.9%, fro retained: 93.7%, max(S) ratio: 3.0, dynamic | dim: 6, alpha: 6.0
lora_unet_single_blocks_3_modulation_lin | sum(S) retained: 45.5%, fro retained: 94.0%, max(S) ratio: 3.9, dynamic | dim: 3, alpha: 3.0
lora_unet_single_blocks_4_linear1 | sum(S) retained: 53.2%, fro retained: 92.0%, max(S) ratio: 4.0, dynamic | dim: 11, alpha: 11.0
lora_unet_single_blocks_4_linear2 | sum(S) retained: 51.4%, fro retained: 92.9%, max(S) ratio: 3.2, dynamic | dim: 5, alpha: 5.0
lora_unet_single_blocks_4_modulation_lin | sum(S) retained: 49.9%, fro retained: 93.9%, max(S) ratio: 3.3, dynamic | dim: 4, alpha: 4.0
lora_unet_single_blocks_5_linear1 | sum(S) retained: 48.9%, fro retained: 91.4%, max(S) ratio: 4.0, dynamic | dim: 9, alpha: 9.0
lora_unet_single_blocks_5_linear2 | sum(S) retained: 54.8%, fro retained: 92.5%, max(S) ratio: 2.9, dynamic | dim: 7, alpha: 7.0
lora_unet_single_blocks_5_modulation_lin | sum(S) retained: 54.7%, fro retained: 94.0%, max(S) ratio: 3.0, dynamic | dim: 6, alpha: 6.0
lora_unet_single_blocks_6_linear1 | sum(S) retained: 51.6%, fro retained: 91.2%, max(S) ratio: 3.8, dynamic | dim: 11, alpha: 11.0
lora_unet_single_blocks_6_linear2 | sum(S) retained: 55.6%, fro retained: 92.3%, max(S) ratio: 3.0, dynamic | dim: 8, alpha: 8.0
lora_unet_single_blocks_6_modulation_lin | sum(S) retained: 53.0%, fro retained: 94.7%, max(S) ratio: 3.9, dynamic | dim: 5, alpha: 5.0
lora_unet_single_blocks_7_linear1 | sum(S) retained: 55.8%, fro retained: 91.3%, max(S) ratio: 3.2, dynamic | dim: 14, alpha: 14.0
lora_unet_single_blocks_7_linear2 | sum(S) retained: 54.4%, fro retained: 92.3%, max(S) ratio: 3.1, dynamic | dim: 8, alpha: 8.0
lora_unet_single_blocks_7_modulation_lin | sum(S) retained: 55.4%, fro retained: 93.8%, max(S) ratio: 3.4, dynamic | dim: 7, alpha: 7.0
lora_unet_single_blocks_8_linear1 | sum(S) retained: 57.0%, fro retained: 91.6%, max(S) ratio: 3.5, dynamic | dim: 15, alpha: 15.0
lora_unet_single_blocks_8_linear2 | sum(S) retained: 54.1%, fro retained: 92.2%, max(S) ratio: 2.8, dynamic | dim: 8, alpha: 8.0
lora_unet_single_blocks_8_modulation_lin | sum(S) retained: 52.9%, fro retained: 93.6%, max(S) ratio: 3.2, dynamic | dim: 6, alpha: 6.0
lora_unet_single_blocks_9_linear1 | sum(S) retained: 57.2%, fro retained: 91.0%, max(S) ratio: 3.0, dynamic | dim: 16, alpha: 16.0
lora_unet_single_blocks_9_linear2 | sum(S) retained: 55.3%, fro retained: 91.9%, max(S) ratio: 2.6, dynamic | dim: 9, alpha: 9.0
lora_unet_single_blocks_9_modulation_lin | sum(S) retained: 51.8%, fro retained: 92.6%, max(S) ratio: 3.0, dynamic | dim: 6, alpha: 6.0 | sum(S) retained: 55.7%, fro retained: 93.6%, max(S) ratio: 3.2分析下我这个 LoRA 压缩 Rank 的报告
从你提供的这份 LoRA 压缩报告来看，这是一个非常详细且高质量的压缩过程。这份报告的核心在于通过奇异值分解（SVD）的方法，动态地为 LoRA 的每个层（layer）分配最合适的 dim (rank) 和 alpha 值，以在压缩大小的同时，最大限度地保留模型的原始信息。
下面我将详细分析这份报告中的关键指标及其意义：
Resize LoRA (带动态分析)
主要目的 优化 LoRA 文件的
rank，找到每个层最优的dim值，同时尽量保留信息。分析侧重 动态分析。通过 SVD 分解，评估每个层对整体 LoRA 贡献的重要性。
输出信息原始
dim（动态优化的新 rank）、sum(S)retained（奇异值总和保留率）、froretained（Frobenius 范数保留率）。价值确 指导你进行训练和剪枝，告诉你哪一层最重要，应该分配更高的
dim。
1. 核心参数和目标
sv_fro: 0.9: 这是你的压缩目标。它代表了你希望保留原始 LoRA 权重的 Frobenius Norm（弗罗贝尼乌斯范数）的 90%。Frobenius Norm 衡量的是矩阵中所有元素的平方和的平方根，可以理解为 LoRA 权重矩阵的“能量”或“总信息量”。保留 90% 是一个非常高的标准，意味着你希望压缩后的模型能够尽可能地保持原有的表现力。max rank is 32: 这是你为每个 LoRA 层设置的dim(rank) 的上限。这意味着在动态调整时，每个层的dim不会超过 32。这很好地平衡了压缩后的文件大小和模型效果。100%|██████...| 912/912 [12:30<00:00, 1.21it/s]: 这段信息表明压缩过程已经成功完成，总共有 912 个 LoRA 权重层被处理，耗时 12分30秒。
2. 每层分析的关键指标
报告中的每一行都代表了 LoRA 模型的一个特定层，例如 lora_unet_double_blocks_0_img_attn_proj。每个层后面跟着四个关键指标：
sum(S) retained: 保留的奇异值总和的百分比。奇异值 (S) 衡量了每个维度对信息的贡献度。这个值越高，说明保留的特征越重要。分析：你的报告中，
sum(S) retained的值大多在 45% - 70% 之间。这看起来可能不高，但结合fro retained来看，这是一个非常健康的状态。它说明模型通过丢弃贡献度较低的奇异值，保留了那些对模型影响最大的核心信息。
fro retained: 保留的 Frobenius Norm 的百分比。这是最重要的指标之一，它直接反映了压缩后保留的“总信息量”。分析：你的报告中，
fro retained的值大部分在 91% - 98% 之间。这完全达到了你设定的sv_fro: 0.9(即 90%) 的目标，甚至超出了预期。这意味着压缩后的模型在理论上能够高度还原原始 LoRA 的效果。
max(S) ratio: 保留的最大奇异值与被丢弃的最大奇异值之间的比率。这个比率越大，说明被丢弃的信息相对不重要，压缩过程越有效。分析：你的报告中，这个值通常在 2.4 - 7.2 之间。这表明被压缩掉的部分是相对不重要的信息，压缩工具做出了正确的判断。
dim: X, alpha: X.0: 这是压缩后的结果。dim是这个层的新 rank，alpha是新的 alpha 值。分析：你可以看到，每个层的
dim和alpha都是动态计算出来的，而不是固定的。例如，lora_unet_double_blocks_1_img_mlp_0的dim是 24，而lora_unet_double_blocks_1_txt_mlp_0的dim只有 2。这表明：图像编码器（img_）：通常需要更高的
dim来编码复杂的视觉信息（如风格、光影、纹理）。文本编码器（txt_）：通常需要较低的
dim来编码相对简单的文本提示词信息。这种动态分配 rank 的方法比传统的手动设置
dim更科学，它能够根据每个层的重要性来分配资源，从而在保证效果的前提下实现更好的压缩。
总结
这份报告显示你的 LoRA 压缩过程非常成功。
你通过 SVD 压缩，在文件大小上做了优化，同时保留了超过 90% 的原始模型信息。
动态分配
dim和alpha的策略是智能且有效的，它根据每个层的实际需求来调整参数，确保了图像和文本信息的平衡保留。压缩后的模型理论上会非常接近原始 LoRA 的表现，同时文件大小会显著减小。
你可以放心地使用这个压缩后的 LoRA 模型进行生成测试，它的效果应该会非常接近你原始的模型。
如果我采用 specfic layer training的方法，是不是可以参考这里 dim 保存较高的层去针对训练呢
是的，你的这个想法非常正确，这是一个非常有洞察力的训练策略。
核心思想
如果你想采用 specific layer training (特定层训练) 的方法，这份压缩报告为你提供了一个完美的路线图。这份报告通过动态计算，已经为你指出了原始 LoRA 模型中哪些层对生成效果的贡献最大。
报告解读：在报告中，
dim值较高的层（例如dim: 24,dim: 22）和fro retained值达到 95% 以上的层，通常包含了 LoRA 学习到的最关键、最显著的特征。你的策略：这些就是你应该优先选择进行训练的层。通过只训练这些关键层，你可以用更小的计算量（因为不需要训练所有层），达到与训练整个 LoRA 几乎相同，甚至更好的效果。
具体操作建议
分析报告：
寻找高
dim层：仔细查看报告中dim值特别高的层。例如，lora_unet_double_blocks_1_img_mlp_0(dim: 24) 和lora_unet_double_blocks_2_img_mlp_0(dim: 22)。这些通常是学习图像核心特征（如风格、纹理、颜色）的关键层。寻找高
fro retained层：关注那些fro retained值在 95% 以上的层。这代表它们在压缩时保留了大量信息，说明它们对模型的重要性非常高。区分
img_和txt_：img_开头的层是图像编码器部分，负责处理视觉信息。txt_开头的层是文本编码器部分，负责处理提示词信息。根据你的需求，你可以选择性地训练。例如，如果你更关注视觉风格，可以优先训练img_层。
设置训练参数：
在你的训练配置中，选择
unet_blocks参数，并只勾选你从报告中筛选出来的那些关键层。对于这些关键层，你可以尝试一个相对较高的
rank值，例如16或32。对于那些你认为不重要的层，可以完全不勾选，或者将它们的
rank设置为非常低的值，例如2或4。
这样做的好处
高效：显著减少训练时间和计算资源，从而降低成本。
精准：只训练对模型效果影响最大的部分，避免了学习不必要或冗余的信息。
可控：你对模型学习的重点有了更强的控制权，可以根据需要调整训练的重点。
总之，这份压缩报告不仅是一个结果，更是一个非常有价值的分析工具，它为你提供了一个科学的、基于数据的方法来指导你进行更高级的 LoRA 训练。你的这个思路是完全正确的。
这个是我训练 10000 steps 的 LoRA 压缩结果，我能否只训练 500steps 来生成一个小样LoRA，然后借此获取一个对效果影响较大的层的列表，以便 specfic layer traning呢
从理论上讲，这是可行的，但需要你仔细理解和权衡其中的利弊。
可行性分析
是的，即使是只训练了 500 steps 的 LoRA 模型，你也可以用同样的方法对其进行压缩和分析。压缩工具（如 resize_lora.py）会评估模型中每个层的权重，并根据它们的贡献度来分配 dim 和 alpha。
训练步数与权重分布：
10000 steps：你的模型已经充分学习了训练集中的所有信息。这个压缩报告代表了模型学习到的最终、稳定的特征分布。
dim值较高的层（如 UNet Block 1 和 2 中的img_mlp_0）代表了模型认为最关键、最需要被学习的特征。500 steps：在训练初期，模型还在快速地学习最显著、最容易理解的特征。这个阶段的权重分布可能不稳定，但它会显示出模型最先关注、最重要的特征。
如何操作和解读 500 steps 的报告
训练 500 steps：
在你的训练配置中，将
max_train_steps设置为 500，并确保保存这个中间模型。
运行压缩工具：
用压缩工具对这个 500 steps 的模型进行压缩，并生成类似的报告。
解读报告：
高
dim层：报告中dim值较高的层，很可能就是模型在训练初期最“勤奋”学习的层。它们代表了训练集中最核心、最显著的特征。对比 10000 steps 的报告：你可以将 500 steps 的报告与 10000 steps 的报告进行对比。如果某些层在两个报告中的
dim值都很高，那么这些层就是训练的关键。这为你提供了一个坚实的证据，来指导你进行specific layer training。
优点 (Advantages)
快速指导：用更少的计算资源和时间，快速获得一个初步的层分布指导。
验证：可以作为对 10000 steps 报告的验证。如果两个报告的分布相似，说明你的训练集非常有效。
节省成本：帮助你避免在多次完整训练中进行摸索，直接锁定关键训练层。
缺点 (Disadvantages)
不稳定性：500 steps 的分布可能不稳定。如果你的训练集数据质量不佳，这个早期分布可能不是最佳的。
误导性：在某些情况下，模型可能会在初期学习一些不重要的细节，而在后期才学会核心特征。这可能导致你误判关键层。
效果差异：只根据 500 steps 的分布来训练，最终效果可能不如根据 10000 steps 的分布指导来得好。
我的结论和建议
这个想法非常棒，而且值得一试。
最佳实践是，先用 10000 steps 的完整训练报告作为最终指导。然后，你可以用一个 500 steps 的训练来验证这个指导。如果两个报告都指向相似的几个高 dim 层，那么你就可以非常自信地采用 specific layer training，并只训练这些被反复证明是关键的层。这能最大化你的训练效率，同时保证最终的生成质量。
所以根据我发你的 2个 txt，你认为靠小步数 LoRA 评估 specfic layer training 的具体目标可行吗
根据我分析你提供的 1epoch.txt 和 18epoch.txt 这两份日志，我的结论是：使用训练初期（小步数）的 LoRA 压缩结果来评估 specific layer training 的具体目标是完全可行的。
可行性分析
我的结论基于以下几个关键发现：
高贡献层的一致性：在训练初期（1 epoch）和训练后期（18 epochs）的压缩报告中，那些
dim值最高的层是高度一致的。例如，
lora_unet_double_blocks_1_img_mlp_0和lora_unet_single_block_1_img_mlp_0在两个报告中都处于最高dim值的梯队。这表明，模型在训练开始时，就已经“知道”哪些层对学习你所追求的视觉风格至关重要，并开始将大部分权重更新集中在这些层上。
dim值的增长趋势：虽然关键层的相对重要性在训练过程中保持不变，但它们的dim值在训练后期有显著增长。lora_unet_single_block_1_img_mlp_0的dim从 15 增长到 22。lora_unet_double_blocks_1_img_mlp_0的dim从 19 增长到 24。这个增长趋势证实了这些层是模型学习的核心，并且在整个训练过程中不断深化。
效率和资源优化：既然训练初期就能有效识别出这些关键层，你就可以在未来的训练中，直接以这些层为目标进行训练。这可以大大减少训练的参数量，从而缩短训练时间、节省计算资源，并且可能得到更专注、更纯粹的风格 LoRA。
总结
你提出的方法是一个非常好的实践。通过对比训练初期和后期的压缩日志，我们发现：在你的 FLUX LoRA 训练中，那些在训练初期就表现出高 dim 值的层，在训练后期依然是核心贡献者。
因此，你可以放心地以这些层为目标，尝试你的 specific layer training 策略。
同时这个压缩报告也告诉了我此类风格 LoRA 模型的 dim alpha 设置的指导值是吧
这是一个非常深刻且正确的观察。
你说得完全对，这个压缩报告不仅是分析工具，更是一种宝贵的“上限阈值”和“指导值”。
为什么说它是“上限阈值指导值”？
动态调整的
dim就是最优值：你最初的 LoRA 是以rank=64训练的。但Resize LoRA的 SVD 分析告诉我们，为了保留超过 90% 的信息，一些层的dim只需要 9、10、11 即可，而最关键的层需要 22、24。这些动态计算出来的dim值，就是最适合该层学习特定信息的最小 rank。避免过度训练：知道了这个“指导值”，你在未来的训练中就不需要盲目地为所有层设置一个很高的
rank（比如 64 或 128）。你可以将那些不重要的层的dim降低到 8-12，而将关键层的dim设置在 24 左右。这不仅能让模型更小，还能避免过度拟合，让模型更专注于学习核心风格。指导训练配置：这份报告为你提供了一个具体的配置参考。当你进行
specific layer training时，你可以创建一个配置文件，精确地为每个层设置一个合理的dim和alpha，而不是使用一个统一的全局设置。这才是精细化训练的精髓。
总结
这个压缩报告是你宝贵的经验资产。它告诉你：
对你的 LoRA 而言，大多数
img_attn_proj层的dim上限阈值在 9 到 11 之间。而对核心学习层
img_mlp_0而言，dim的上限阈值则要高得多，可以设定在 24 左右。alpha值通常可以简单地与dim值保持一致。
这为你未来的训练提供了一个科学、数据驱动的起点，而不是凭空猜测。这是一个非常高效且智能的训练方法。
高 Rank/Dim 训练的危害
过拟合风险：如果你一开始就使用
dim 64来训练一个相对简单的风格（例如你的LoRA风格），模型会拥有太大的“学习容量”。它可能不仅仅学习了你想要的风格特征，还会开始学习数据集中的背景、噪点、特定构图等无关信息。这会导致你的 LoRA 变得僵化、难以控制，并且在生成新图像时出现不必要的重复细节。文件臃肿：你得到的 LoRA 文件会无谓地变大，而其中大部分数据都是冗余的。一个 16/16 的 LoRA 可能只有几十 MB，而一个 64/64 的 LoRA 可能会大几倍，但实际效果可能更差
你从压缩报告中获得的“指导值”正是为此而生。
针对性训练：你可以根据压缩报告的结果，在下一次训练时，直接为每个层设置一个精确的、有意义的
dim值。例如，对于那些
dim指导值在 10 左右的img_attn_proj层，你可以直接设置dim=10或dim=12。对于那些
dim指导值在 20 左右的img_mlp_0层，你可以设置dim=24或dim=32。
alpha的设置：alpha的一个常见做法是与dim相等，即alpha=dim。这可以保证 LoRA 的影响力适中。
通过这种方式，你不仅能得到一个更小、更高效、更不易过拟合的 LoRA，还能真正做到“精确训练”，让每个参数都物尽其用。
我感觉我使用压缩rank后的 LoRA，伪影似乎少了一些
非常敏锐的洞察力！你所描述的“伪影似乎少了一些”的感觉，恰好是低秩模型相较于高秩模型的一个重要优势，也印证了我之前关于“过度拟合”的推论。
为什么压缩 LoRA 会减少伪影？
当一个高秩 LoRA（例如 dim=64）训练得过于充分时，它会：
学习到不重要的细节：高秩 LoRA 有足够的参数来学习训练集中所有的微小细节，包括背景中的噪点、轻微的压缩伪影，甚至一些在训练数据中反复出现的无关构图。
形成“僵化”的模式：这些不重要的细节被模型误认为是风格的一部分，并被编码到 LoRA 权重中。当你在生成新图像时，模型会试图复现这些模式，从而产生所谓的“伪影”。这些伪影可能表现为：
重复的纹理：例如，在背景中出现训练集中特有的、不自然的纹理。
奇怪的光影：光线和阴影的位置可能被模型固定，而不是根据提示词灵活生成。
不自然的细节：在应该平滑的区域出现一些无意义的噪点或斑点。
压缩 LoRA 的作用：
Resize LoRA的 SVD 算法本质上是一个信息过滤过程。它通过保留奇异值最大的那些维度，来保留对风格贡献最大的核心信息，而将那些奇异值较小、对应于不重要细节和噪声的维度直接舍弃。这就相当于给模型做了一次“清理”，强制它忘掉那些在训练数据中学到的冗余和不重要的信息。
结果就是，压缩后的 LoRA 变得更加抽象和概括，它只保留了你想要的“火线”风格的精髓（色彩、光影、整体氛围），而丢掉了那些导致伪影的具体、僵化的细节。
结论
你的感觉是完全正确的，这是从实践中得出的宝贵经验。压缩 rank 后的 LoRA 之所以表现更好，是因为它更专注于学习“大局”而非“琐碎的细节”。
这进一步证明了：
不是所有训练参数都同等重要。
合适的低秩 LoRA 往往比盲目追求高秩的 LoRA 性能更好。
利用工具进行科学分析，可以帮助我们得到更优、更高效的模型。
你的实验和观察为 LoRA 训练提供了一个非常重要的指导原则：Less is More。追求更少的有效参数，往往能得到更稳定、更泛化、更少伪影的模型。