Ostris ai-toolkit - default settings for Z-Image
learning rate is 0.0002, 6000 steps (Less steps = bad pussy but ok pose)

Personally found 50 good quality images. resized to 1024x768, 1024x1024 (although I found later that the training resizes them slightly smaller than that)

Couldn't find anything about masked loss like in Kohya, so blacked out the faces of the women in a paint program. (It was quicker to do it manually).

I do captioning before blacking the faces (Florence2), then manually alter each caption, to better describe the scene, making sure I add the keyword and pose information. The captions are quite detailed, once or two sentences. It works much better than simple captions.

Make sure the training caption says something like "A woman with a blacked out face kneeling on a stool, faceforwardlookback kneeling pose. On a white sofa". I never get black faces generated if I do this.

@TheGlowingGuardian Here you go.
Not the whole config file (missed out samples, etc), but the important bits. (I redacted some folder paths).

process:

    - type: "diffusion_trainer"

      training_folder: "<redacted>\ai-toolkit\\output"

      sqlite_db_path: "./aitk_db.db"

      device: "cuda"

      trigger_word: "faceforwardlookback"

      performance_log_every: 10

      network:

        type: "lora"

        linear: 32

        linear_alpha: 32

        conv: 16

        conv_alpha: 16

        lokr_full_rank: true

        lokr_factor: -1

        network_kwargs:

          ignore_if_contains: []

      save:

        dtype: "bf16"

        save_every: 250

        max_step_saves_to_keep: 2

        save_format: "diffusers"

        push_to_hub: false

      datasets:

        - folder_path: "<redacted>\ai-toolkit\\datasets/fflb2"

          mask_path: null

          mask_min_value: 0.1

          default_caption: "faceforwardlookback"

          caption_ext: "txt"

          caption_dropout_rate: 0.05

          cache_latents_to_disk: true

          is_reg: false

          network_weight: 1

          resolution:

            - 512

            - 768

            - 1024

          controls: []

          shrink_video_to_frames: true

          num_frames: 1

          do_i2v: true

          flip_x: false

          flip_y: false

      train:

        batch_size: 1

        bypass_guidance_embedding: false

        steps: 6000

        gradient_accumulation: 1

        train_unet: true

        train_text_encoder: false

        gradient_checkpointing: true

        noise_scheduler: "flowmatch"

        optimizer: "adamw8bit"

        timestep_type: "weighted"

        content_or_style: "balanced"

        optimizer_params:

          weight_decay: 0.0001

        unload_text_encoder: false

        cache_text_embeddings: true

        lr: 0.0002

        ema_config:

          use_ema: false

          ema_decay: 0.99

        skip_first_sample: true

        force_first_sample: false

        disable_sampling: false

        dtype: "bf16"

        diff_output_preservation: false

        diff_output_preservation_multiplier: 1

        diff_output_preservation_class: "person"

        switch_boundary_every: 1

        loss_type: "mse"

      model:

        name_or_path: "<redacted>\\Z-Image-Turbo"

        quantize: true

        qtype: "qfloat8"

        quantize_te: true

        qtype_te: "qfloat8"

        arch: "zimage:turbo"

        low_vram: true

        model_kwargs: {}

        layer_offloading: false

        layer_offloading_text_encoder_percent: 1

        layer_offloading_transformer_percent: 1

        assistant_lora_path: "<redacted>\\zimage_turbo_training_adapter_v1.safetensors"

@GlowingGuardianGirl Glad to help. 😊