No difference with XL model

sdbds commented 7 months ago

tested with train_xl and train_xl_mem_reduce

Con6924 commented 7 months ago

Could you post your training configs for more details? Here seem to be some alterations between before/after training, you may use a smaller la_strength to give a larger weight for erasing.

sdbds commented 7 months ago

Could you post your training configs for more details? Here seem to be some alterations between before/after training, you may use a smaller la_strength to give a larger weight for erasing.

i use default config in repos pikachu...

maybe guidance_scale or la_strength problem

config.yaml

prompts_file: "configs/pikachu/prompt.yaml"

pretrained_model:
  name_or_path: "tac_anime.safetensors"
  v2: false
  v_pred: false
  clip_skip: 1

network:
  rank: 32
  alpha: 1.0

train:
  precision: bf16
  noise_scheduler: "ddim"
  iterations: 2000
  batch_size: 1
  lr: 0.0002
  unet_lr: 0.0002
  text_encoder_lr: 1e-05
  optimizer_type: "AdamW"
  lr_scheduler: "constant"
  lr_warmup_steps: 0
  lr_scheduler_num_cycles: 1
  max_denoising_steps: 50

save:
  name: "pikachu"
  path: "output/pikachu"
  per_steps: 500
  precision: bf16

logging:
  use_wandb: false
  interval: 0
  seed: 0
  generate_num: 2
  run_name: "pikachu"
  verbose: false
  prompts: ['pikachu', '', 'dog', 'mickey', 'woman']

other:
  use_xformers: true

prompt.yaml

- target: "pikachu"
  positive: "pikachu"
  unconditional: ""
  neutral: ""
  action: "erase_with_la"
  guidance_scale: 4
  resolution: 512
  batch_size: 1
  dynamic_resolution: True
  la_strength: 1000
  sampling_batch_size: 1

sdbds commented 7 months ago

Could you post your training configs for more details? Here seem to be some alterations between before/after training, you may use a smaller la_strength to give a larger weight for erasing.

Thank you for replaying,i change la_strength to 10, it works well.

sdbds commented 7 months ago

Any suggestion for more erasing effect?

if i used SPM with other object+ pikachu,it will be mixer with them instead of erasing just pikachu

Con6924 commented 7 months ago

Any suggestion for more erasing effect?

if i used SPM with other object+ pikachu,it will be mixer with them instead of erasing just pikachu

With our posted inference code, the Facilitated Transport (FT) mechanism is used to avoid this phenomenon. If you are using sd-webui for generation, you might just try to adjust the lora_weight in manual. If possible, we will develop a SPM plugin to implement the FT mechanism for webui in the future.

Con6924 / SPM

No difference with XL model #3