stack.yaml

# SceneMachine model stack manifest
#
# Lists the exact upstream weight repos this version of SceneMachine
# depends on. All weights are hosted in the WindstormLabs HF org
# (Windstorm Labs is SceneMachine's parent organization; the same
# weight mirrors are shared by other Windstorm sub-projects).
#
# Update this file when the application code requires a new model
# version. The application reads this manifest at boot to know which
# WindstormLabs/* repos to download from.

manifest_version: 1
generated: 2026-05-13
scenemachine_min_version: "0.1"

# The canonical mirror org for SceneMachine's model dependencies.
# If you fork SceneMachine and want a different mirror, override this.
mirror_org: WindstormLabs

stacks:

  # Wan 2.2 family — the primary video generation stack.
  wan22:
    description: |
      Alibaba's Wan 2.2 14B family. Three sibling models share the
      same VAE / text encoder / CLIP vision encoders.

    text_to_video:
      hf_repo: WindstormLabs/wan22-t2v-fp8
      files:
        - wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors
        - wan2.2_t2v_low_noise_14B_fp8_scaled.safetensors
      vram_gb: 22
      use_case: |
        Establishing shots, prompt-only scenes, anything without a
        character reference or prior-frame continuity.

    image_to_video:
      hf_repo: WindstormLabs/wan22-i2v-fp8
      files:
        - wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors
        - wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors
      vram_gb: 24
      use_case: |
        Shot-to-shot continuity. Feed the last frame of the prior shot
        as the seed image; produces a video that flows visually from it.

    animate:
      hf_repo: WindstormLabs/wan22-animate-bf16
      files:
        - wan2.2_animate_14B_bf16.safetensors
      vram_gb: 32
      use_case: |
        Character-ID-preserving generation. Requires a reference image
        of the character; the model preserves their identity across the
        shot. Validated 1.7 min/shot with the Lightx2v speed LoRA.

    shared_encoders:
      hf_repo: WindstormLabs/wan22-encoders
      files:
        - wan_2.1_vae.safetensors                  # used by all 3 Wan stacks
        - umt5_xxl_bf16_from_pth.safetensors       # T5 text encoder, all 3
        - sigclip_vision_patch14_384.safetensors   # CLIP vision, I2V only
        - clip_vision_h.safetensors                # CLIP-ViT-H, Animate only (1280-dim)

    speed_loras:
      hf_repo: WindstormLabs/wan22-loras
      files:
        - Wan_2_2_I2V_A14B_HIGH_lightx2v_4step_lora_260412_rank_64_fp16.safetensors
        - wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors
      use_case: |
        Kijai's Lightx2v 4-step distillation. When enabled, drops the
        sampler from 30 steps to 4 with cfg=1.0 — 8.3× wallclock speedup
        on Wan Animate. Verified to transfer cleanly to Animate when
        paired with the correct embed chain + CLIP-ViT-H.

  # LTX-2 — alternate cinematic stack (slower, comparable quality).
  ltx2:
    description: |
      Lightricks LTX-2 19B Dev FP8 plus the Gemma text encoder.

    dev_fp8:
      hf_repo: WindstormLabs/ltx2-19b-fp8
      files:
        - ltx-2-19b-dev-fp8.safetensors
        - model-00001-of-00005.safetensors  # Gemma encoder shards
        - model-00002-of-00005.safetensors
        - model-00003-of-00005.safetensors
        - model-00004-of-00005.safetensors
        - model-00005-of-00005.safetensors
      vram_gb: 28

  # Hunyuan — reserved for Stack B implementation (not yet wired in app)
  hunyuan:
    description: |
      Tencent HunyuanVideo + HunyuanVideo-I2V + HunyuanCustom partial
      mirror. Reserved for SceneMachine Stack B (alternate character-
      consistency path via Hunyuan's built-in identity preservation,
      no LoRA needed). Provider workflow not yet implemented in the
      app — weights are mirrored for future use.
    hf_repo: WindstormLabs/hunyuan
    license_note: |
      Tencent HunyuanVideo Community License — check the upstream
      repo for current terms before any commercial use.