Trigger word
@4x0styleRecommended strength is 0.6 - 1.0
V4
What's different from v3: pantyhose texture.
That was the one thing v3 weaker — the actual material: the fine mesh structure, the individual threads, the way the knit lines catch light across the fabric. This version goes after that specifically.
The approach is experimental NaViT native-resolution training: images are trained at their original resolution (up to 3000×4000) instead of being downscaled into fixed buckets. Fine fabric structure survives instead of getting smeared away in the downscale and WAN VAE— which is exactly what was killing the texture before.
This also means the LoRA holds up at large inference resolutions, and that's not a separate feature — it's the same goal. More pixels means more room for the weave to actually render. The samples were generated at 1920×1920, and I'd recommend generating at high resolution to get the most out of this version.
Recommendation to Anima Lora Trainer I am working on /ᐠ ̷ ̷𝅒 ̷‸ ̷𝅒 ̷ ᐟ\ﾉ
https://github.com/WalkingMeatAxolotl/AnimaLoraStudio
transformer_path: ~
vae_path: ~
text_encoder_path: ~
t5_tokenizer_path: ~
data_dir: ~
resolution:
- 1024
aspect_ratio_limit: 2.0
reg_data_dir: ~
reg_caption: null
reg_weight: 0.5
shuffle_caption: true
keep_tokens: 1
flip_augment: true
tag_dropout: 0.0
prefer_json: true
caption_comfy_encoding: true
cache_latents: true
vae_cache_batch_size: 0
navit_packing: true
navit_token_budget: 16384
navit_max_images_per_pack: 0
navit_text_trim_padding: false
navit_pack_strategy: next_fit
navit_pack_ffd_window: 256
navit_drop_last: false
navit_native_resolution: true
navit_native_over_budget: downscale
cache_encode_tiled: true
cache_encode_tile_px: 1024
cache_encode_tile_overlap: 128
cache_encode_max_pixels: 0
lora_type: lora
lora_rank: 32
lora_alpha: 32.0
lora_dora: false
lora_rs: false
lora_dropout: 0.0
lora_rank_dropout: 0.0
lora_module_dropout: 0.05
lora_reg_dims: null
epochs: 40
max_steps: 0
batch_size: 2
grad_checkpoint: true
grad_accum: 2
learning_rate: 1.0
lr_scheduler: none
optimizer_type: prodigy_plus_schedulefree
ppsf_d_coef: 3.0
ppsf_prodigy_steps: 0
ppsf_beta1: 0.9
ppsf_beta2: 0.99
ppsf_split_groups: true
ppsf_split_groups_mean: false
ppsf_use_speed: false
ppsf_fused_back_pass: false
ppsf_use_stableadamw: true
weight_decay: 0.0
kv_trim: false
vae_tiling: auto
noise_enhancement_type: none
timestep_sampling: uniform
timestep_schedule_shift: 0.7
timestep_shift_resolution_aware: true
infonoise_enabled: false
loss_type: mse
loss_weighting: none
leap_enabled: false
sra_enabled: false
grad_clip_max_norm: 0.0
mixed_precision: bf16
attention_backend: xformers
num_workers: 0
output_dir: ~
output_name: ~
save_every_epochs: 2
save_every_steps: 0
save_state_every_epochs: 0
save_state_every_steps: 500
seed: 42
resume_lora: null
resume_state: nullDescription
Trained on Anima Base 1.0
What's different from v3: pantyhose texture.
That was the one thing v3 is weaker — the actual material: the fine mesh structure, the individual threads, the way the knit lines catch light across the fabric. This version goes after that specifically.
The approach is experimental NaViT native-resolution training: images are trained at their original resolution (up to 3000×4000) instead of being downscaled into fixed buckets. Fine fabric structure survives instead of getting smeared away in the downscale — which is exactly what was killing the texture before.
This also means the LoRA holds up at large inference resolutions, and that's not a separate feature — it's the same goal. More pixels means more room for the weave to actually render. The samples were generated at 1920×1920, and I'd recommend generating at high resolution to get the most out of this version.