Upload refiner_local_upsample_lazy_preserve_ratio_rearange_ablation_lora_videomama_binary_mix_ratio_polyhole_rerange_dinoh_rank64_20261101-111822 steps 34000, 39000

refiner_local_upsample_lazy_preserve_ratio_rearange_ablation_lora_videomama_binary_mix_ratio_polyhole_rerange_dinoh_rank64_20261101-111822/checkpoints/epoch_43_step_34000.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e0a923dd4adc6a9614863c75a2c71a8b25e5b49c2e54e051330380ade7f7156
+size 5897182226

refiner_local_upsample_lazy_preserve_ratio_rearange_ablation_lora_videomama_binary_mix_ratio_polyhole_rerange_dinoh_rank64_20261101-111822/checkpoints/epoch_49_step_39000.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f6516add748353731b78c9614a193ff259d8a5e9236bed4107f1ac6b6b76bffa
+size 5897182226

refiner_local_upsample_lazy_preserve_ratio_rearange_ablation_lora_videomama_binary_mix_ratio_polyhole_rerange_dinoh_rank64_20261101-111822/config.yaml

@@ -0,0 +1,376 @@
+data:
+    data_dir: []
+    caption_proportion:
+        prompt: 1
+    external_caption_suffixes: []
+    external_clipscore_suffixes: []
+    clip_thr_temperature: 1.0
+    clip_thr: 0.0
+    del_img_clip_thr: 0.0
+    sort_dataset: false
+    load_text_feat: false
+    transform: default_train
+    type: VideoMattingCompositingDataset
+    image_size: 4096
+    hq_only: false
+    valid_num: 0
+    data: null
+    extra: null
+    s3_bucket: ''
+    s3_prefix: ''
+    s3_metadata_key: ''
+    s3_region: us-west-2
+    dataset_repeat: 1
+    frame_sampling: random
+    mask_augment_method: downsample,polygon
+    downsample_scales:
+    - 24
+    - 32
+    - 48
+    polygon_tolerances:
+    - 0.004
+    - 0.003
+    - 0.002
+    polygon_keep_holes: true
+    binarize_threshold: 127
+    preserve_prob: 0.0
+    preserve_noise_std: 0.0
+    guide_mode: binary
+    guide_lowres: 512
+    guide_lowres_blur_sigma: 0.0
+    guide_lowres_noise_std: 0.0
+    include_guide_mask: true
+    clip_len: 3
+    low_res: 256
+    sources: null
+    polarizer_fgr_prefixes: []
+    greenscreen_fgr_prefixes: []
+    bg_prefix: ''
+    s3_unsigned: false
+    fg_fit_range:
+    - -0.2
+    - 1.2
+    fg_flip_prob: 0.5
+    bg_fit_range:
+    - 1.0
+    - 1.0
+    bg_flip_prob: 0.5
+    foreground_sources:
+    -   type: video
+        fg_path: VideoMatte240K_JPEG_HD/train/fgr
+        alpha_path: VideoMatte240K_JPEG_HD/train/pha
+        weight: 1.0
+        alpha_black_level: 1
+    -   type: video
+        fg_path: CRGNN/fg
+        alpha_path: CRGNN/alpha
+        weight: 1.0
+    -   type: image
+        fg_path: AM-2k/train/fg
+        alpha_path: AM-2k/train/mask
+        weight: 1.0
+    -   type: image
+        fg_path: P3M-10k/train/blurred_image
+        alpha_path: P3M-10k/train/mask
+        weight: 1.0
+    -   type: image
+        fg_path: AIM-500/original
+        alpha_path: AIM-500/mask
+        weight: 1.0
+    -   type: image
+        fg_path: Distinctions-646/Train/FG
+        alpha_path: Distinctions-646/Train/GT
+        weight: 1.0
+    background_sources:
+    - BG-20k/train
+    data_root: /home/ubuntu/data1/samuel/data
+    s3_root: ''
+    fg_instances:
+    - 1
+    - 2
+    matte_target: subject
+    occluder_front_prob: 0.5
+    min_subject_visible: 0.1
+    max_fg_upscale: 1.0
+    min_fg_resolution: 768
+    datasets: null
+model:
+    model: PixDiTRefinerTrainer
+    teacher: null
+    image_size: 1024
+    mixed_precision: bf16
+    fp32_attention: true
+    load_from: /home/ubuntu/data1/samuel/data/pretrained_models/PixelDiT-1300M-1024px/pixeldit_t2i_v1.pth
+    discriminator_model: null
+    teacher_model: null
+    teacher_model_weight_dtype: null
+    resume_from: null
+    aspect_ratio_type: ASPECT_RATIO_1024
+    multi_scale: false
+    pe_interpolation: 1.0
+    micro_condition: false
+    attn_type: linear
+    autocast_linear_attn: false
+    ffn_type: glumbconv
+    mlp_acts:
+    - silu
+    - silu
+    - null
+    mlp_ratio: 2.5
+    use_pe: false
+    pos_embed_type: sincos
+    qk_norm: false
+    class_dropout_prob: 0.0
+    linear_head_dim: 32
+    cross_norm: false
+    cross_attn_type: flash
+    logvar: false
+    cfg_scale: 4
+    cfg_embed: false
+    cfg_embed_scale: 1.0
+    guidance_type: classifier-free
+    ladd_multi_scale: true
+    head_block_ids: null
+    extra:
+        patch_size: 16
+        num_groups: 24
+        hidden_size: 1536
+        pixel_hidden_size: 16
+        pixel_attn_hidden_size: 1152
+        pixel_num_groups: 16
+        patch_depth: 14
+        pixel_depth: 2
+        num_text_blocks: 4
+        txt_embed_dim: 2304
+        txt_max_length: 300
+        use_text_rope: true
+        text_rope_theta: 10000.0
+        repa_encoder_index: -1
+        cond_image_channels: 6
+        feature_source: dinov3
+        dino_model: facebook/dinov3-vith16plus-pretrain-lvd1689m
+        dino_resolution: 1024
+        dino_resolution_ratio: 1
+        dino_layer: -1
+        dino_feature_dim: 1280
+        zero_init_lq: true
+        pit_lq_inject: false
+        use_preserve_embed: true
+        lq_gate_type: residual
+        lq_inject_blocks:
+        - 0
+        - 2
+        - 4
+        guide_source: videomama
+        binary_guide_prob: 0.5
+        videomama_res: 512
+        videomama_res_ratio: 0.25
+        videomama_base_model: /home/ubuntu/data1/samuel/data/pretrained_models/stable-video-diffusion-img2vid-xt
+        videomama_unet_ckpt: /home/ubuntu/data1/samuel/data/pretrained_models/VideoMaMa
+        videomama_svd_root: /home/ubuntu/data1/samuel/data/code/VideoMaMa/samuel_svd
+        videomama_mask_cond_mode: vae
+vae:
+    vae_type: none
+    vae_pretrained: none
+    weight_dtype: float32
+    scale_factor: 1.0
+    vae_latent_dim: 3
+    vae_downsample_rate: 1
+    sample_posterior: false
+    extra: null
+text_encoder:
+    text_encoder_name: zero
+    caption_channels: 2304
+    y_norm: true
+    y_norm_scale_factor: 0.01
+    model_max_length: 300
+    chi_prompt: []
+    extra: null
+scheduler:
+    train_sampling_steps: 1000
+    predict_flow_v: true
+    noise_schedule: linear_flow
+    pred_sigma: false
+    learn_sigma: false
+    vis_sampler: flow_dpm-solver
+    flow_shift: 4.0
+    weighting_scheme: logit_normal
+    weighting_scheme_discriminator: logit_normal_trigflow
+    add_noise_timesteps:
+    - 1.5708
+    logit_mean: 0.0
+    logit_std: 1.0
+    logit_mean_discriminator: 0.0
+    logit_std_discriminator: 1.0
+    sigma_data: 0.5
+    timestep_norm_scale_factor: 1.0
+    extra: null
+train:
+    num_workers: 4
+    seed: 1
+    train_batch_size: 1
+    num_epochs: 100
+    gradient_accumulation_steps: 4
+    effective_batch_size: 16
+    train_resolutions:
+    -   - 1536
+        - 1536
+    -   - 2048
+        - 2048
+    -   - 2560
+        - 2560
+    grad_checkpointing: true
+    gradient_clip: 0.1
+    gc_step: 1
+    optimizer:
+        betas:
+        - 0.9
+        - 0.999
+        - 0.9999
+        eps:
+        - 1.0e-30
+        - 1.0e-16
+        lr: 0.0001
+        type: CAMEWrapper
+        weight_decay: 0.0
+    optimizer_D:
+        eps: 1.0e-10
+        lr: 0.0001
+        type: AdamW
+        weight_decay: 0.03
+    load_from_optimizer: false
+    load_from_lr_scheduler: false
+    resume_lr_scheduler: true
+    lr_schedule: constant
+    lr_schedule_args:
+        num_warmup_steps: 1000
+    override_lr_on_resume: false
+    auto_lr:
+        rule: sqrt
+    eval_batch_size: 16
+    use_fsdp: false
+    use_flash_attn: false
+    eval_sampling_steps: 500
+    log_interval: 20
+    mask_type: 'null'
+    mask_loss_coef: 0.0
+    load_mask_index: false
+    snr_loss: false
+    real_prompt_ratio: 1.0
+    early_stop_hours: 10000.0
+    save_image_epochs: 1
+    save_model_epochs: 5
+    save_model_steps: 1000
+    visualize: true
+    null_embed_root: /home/ubuntu/data1/samuel/data/training_models/refiner_local_upsample_lazy_preserve_ratio_rearange_ablation_lora_videomama_binary_mix_ratio_polyhole_rerange_dinoh_rank64_20261101-111822/pretrained_models
+    valid_prompt_embed_root: output/tmp_embed/
+    validation_prompts:
+    - dog
+    - portrait photo of a girl, photograph, highly detailed face, depth of field
+    - Self-portrait oil painting, a beautiful cyborg with golden hair, 8k
+    - Astronaut in a jungle, cold color palette, muted colors, detailed, 8k
+    - A photo of beautiful mountain with realistic sunset and blue lake, highly detailed,
+        masterpiece
+    local_save_vis: true
+    deterministic_validation: true
+    online_metric: false
+    eval_metric_step: 5000
+    online_metric_dir: metric_helper
+    work_dir: output/pixeldit_refiner_upsample_lazy_preserve_binary_videomama_dyncanvas_ratio_polyhole_local
+    skip_step: 0
+    loss_type: huber
+    huber_c: 0.001
+    num_ddim_timesteps: 50
+    ema_decay: 0.95
+    ema_update: false
+    ema_rate: 0.9999
+    repa_loss_weight: 0.0
+    repa_encoder_index: 14
+    mse_weight: 1.0
+    video_shared_noise: false
+    matting_alpha_weight: 0.0
+    matting_grad_weight: 0.0
+    matting_lap_weight: 0.0
+    matting_lap_levels: 5
+    matting_snr_weight: true
+    benchmark_eval: false
+    benchmark_data_dir: ''
+    benchmark_image_subdir: blurred_image
+    benchmark_mask_subdir: mask
+    benchmark_trimap_subdir: trimap
+    benchmark_step: 4000
+    benchmark_max_samples: 30
+    benchmark_steps: 20
+    benchmark_resize_mode: resize
+    val_image_dir: /home/ubuntu/data1/samuel/data/youtube_curated_4k/image
+    val_mask_dir: /home/ubuntu/data1/samuel/data/youtube_curated_4k/mask
+    val_min_resolution: 1536
+    patch_shuffle_unit: 0
+    patch_shuffle_prob: 1.0
+    pack_fg_train: true
+    pack_fg_unit: 128
+    pack_fg_thresh: 0.02
+    pack_fg_dilate: 1
+    pack_fg_slice: random
+    pack_fg_canvas_resolutions:
+    - 1024
+    - 1536
+    sync_rank_cost_draws: true
+    pack_fg_preserve_enable: true
+    pack_fg_refine_prob: 0.7
+    validation_num: 4
+    refiner_val_pack: ''
+    pipeline_val_num: 3
+    lora_enable: true
+    lora_rank: 64
+    lora_alpha: 64.0
+    lora_dropout: 0.0
+    lora_target_modules:
+    - all_linear
+    lora_exclude_modules:
+    - _repa_projector
+    - y_embedder
+    lora_full_train_modules:
+    - s_embedder
+    - pixel_embedder
+    - final_layer
+    - lq_proj
+    - pit_lq_gate
+    - refine_embed
+    - preserve_embed
+refiner:
+    enable: false
+    vae_path: /home/colligo/sensei/Neurips26/weights/DiffSynth-Studio/Wan-Series-Converted-Safetensors/Wan2.1_VAE.safetensors
+    low_res: 256
+    z_dim: 16
+    min_detail_ratio: 0.25
+    max_blur_sigma: 1.0
+    max_noise_std: 0.04
+    enable_morph: false
+    max_morph_kernel: 7
+    min_severity: 0.0
+    max_severity: 1.0
+    degrade_mode: detail
+    cutoff_res:
+    - 32
+    cutoff_blur_sigma: 0.0
+    cutoff_noise_std: 0.0
+    use_real_coarse: false
+    coarse_num_frames: 3
+    lq_gate_type: sigma_aware_per_token_per_dim
+    coarse_steps: 1
+    coarse_ckpt: ''
+    coarse_base_wan: /home/colligo/sensei/Neurips26/weights/Wan-AI/Wan2.1-T2V-1.3B/diffusion_pytorch_model.safetensors
+    coarse_text_encoder: /home/colligo/sensei/Neurips26/weights/DiffSynth-Studio/Wan-Series-Converted-Safetensors/models_t5_umt5-xxl-enc-bf16.safetensors
+    coarse_tokenizer: /home/colligo/sensei/Neurips26/weights/Wan-AI/Wan2.1-T2V-1.3B/google/umt5-xxl
+    coarse_context: /home/colligo/data/pretrained_models/coarse_native/matting_context_umt5.pt
+    coarse_lora_alpha: 1.0
+work_dir: /home/ubuntu/data1/samuel/data/training_models/refiner_local_upsample_lazy_preserve_ratio_rearange_ablation_lora_videomama_binary_mix_ratio_polyhole_rerange_dinoh_rank64_20261101-111822
+resume_from: null
+load_from: /home/ubuntu/data1/samuel/data/pretrained_models/PixelDiT-1300M-1024px/pixeldit_t2i_v1.pth
+debug: false
+caching: false
+report_to: wandb
+tracker_project_name: pixeldit_refiner_local
+name: refiner_local_upsample_lazy_preserve_ratio_rearange_ablation_lora_videomama_binary_mix_ratio_polyhole_rerange_dinoh_rank64_20261101-111822
+loss_report_name: loss