config/models/ace_0.6b_512.yaml

NAME: ACE_0.6B_512
IS_DEFAULT: False
DEFAULT_PARAS:
  PARAS:
  #
  INPUT:
    INPUT_IMAGE:
    INPUT_MASK:
    TASK:
    PROMPT: ""
    NEGATIVE_PROMPT: ""
    OUTPUT_HEIGHT: 512
    OUTPUT_WIDTH: 512
    SAMPLER: ddim
    SAMPLE_STEPS: 20
    GUIDE_SCALE: 4.5
    GUIDE_RESCALE: 0.5
    SEED: -1
    TAR_INDEX: 0
  OUTPUT:
    LATENT:
    IMAGES:
    SEED:
  MODULES_PARAS:
    FIRST_STAGE_MODEL:
      FUNCTION:
        - NAME: encode
          DTYPE: float16
          INPUT: ["IMAGE"]
        - NAME: decode
          DTYPE: float16
          INPUT: ["LATENT"]
    #
    DIFFUSION_MODEL:
      FUNCTION:
        - NAME: forward
          DTYPE: float16
          INPUT: ["SAMPLE_STEPS", "SAMPLE", "GUIDE_SCALE"]
    #
    COND_STAGE_MODEL:
      FUNCTION:
        - NAME: encode_list
          DTYPE: bfloat16
          INPUT: ["PROMPT"]
#
MODEL:
  NAME: LdmACE
  PRETRAINED_MODEL:
  IGNORE_KEYS: [ ]
  SCALE_FACTOR: 0.18215
  SIZE_FACTOR: 8
  DECODER_BIAS: 0.5
  DEFAULT_N_PROMPT: ""
  TEXT_IDENTIFIER: [ '{image}', '{image1}', '{image2}', '{image3}', '{image4}', '{image5}', '{image6}', '{image7}', '{image8}', '{image9}' ]
  USE_TEXT_POS_EMBEDDINGS: True
  #
  DIFFUSION:
    NAME: ACEDiffusion
    PREDICTION_TYPE: eps
    MIN_SNR_GAMMA:
    NOISE_SCHEDULER:
      NAME: LinearScheduler
      NUM_TIMESTEPS: 1000
      BETA_MIN: 0.0001
      BETA_MAX: 0.02
  #
  DIFFUSION_MODEL:
    NAME: DiTACE
    PRETRAINED_MODEL: hf://scepter-studio/ACE-0.6B-512px@models/dit/ace_0.6b_512px.pth
    IGNORE_KEYS: [ ]
    PATCH_SIZE: 2
    IN_CHANNELS: 4
    HIDDEN_SIZE: 1152
    DEPTH: 28
    NUM_HEADS: 16
    MLP_RATIO: 4.0
    PRED_SIGMA: True
    DROP_PATH: 0.0
    WINDOW_DIZE: 0
    Y_CHANNELS: 4096
    MAX_SEQ_LEN: 1024
    QK_NORM: True
    USE_GRAD_CHECKPOINT: True
    ATTENTION_BACKEND: flash_attn
  #
  FIRST_STAGE_MODEL:
    NAME: AutoencoderKL
    EMBED_DIM: 4
    PRETRAINED_MODEL: hf://scepter-studio/ACE-0.6B-512px@models/vae/vae.bin
    IGNORE_KEYS: []
    #
    ENCODER:
      NAME: Encoder
      CH: 128
      OUT_CH: 3
      NUM_RES_BLOCKS: 2
      IN_CHANNELS: 3
      ATTN_RESOLUTIONS: [ ]
      CH_MULT: [ 1, 2, 4, 4 ]
      Z_CHANNELS: 4
      DOUBLE_Z: True
      DROPOUT: 0.0
      RESAMP_WITH_CONV: True
    #
    DECODER:
      NAME: Decoder
      CH: 128
      OUT_CH: 3
      NUM_RES_BLOCKS: 2
      IN_CHANNELS: 3
      ATTN_RESOLUTIONS: [ ]
      CH_MULT: [ 1, 2, 4, 4 ]
      Z_CHANNELS: 4
      DROPOUT: 0.0
      RESAMP_WITH_CONV: True
      GIVE_PRE_END: False
      TANH_OUT: False
  #
  COND_STAGE_MODEL:
    NAME: ACETextEmbedder
    PRETRAINED_MODEL: hf://scepter-studio/ACE-0.6B-512px@models/text_encoder/t5-v1_1-xxl/
    TOKENIZER_PATH: hf://scepter-studio/ACE-0.6B-512px@models/tokenizer/t5-v1_1-xxl
    LENGTH: 120
    T5_DTYPE: bfloat16
    ADDED_IDENTIFIER: [ '{image}', '{caption}', '{mask}', '{ref_image}', '{image1}', '{image2}', '{image3}', '{image4}', '{image5}', '{image6}', '{image7}', '{image8}', '{image9}' ]
    CLEAN: whitespace
    USE_GRAD: False