upload stack.yaml (license: CC0)
Browse files- stack.yaml +104 -0
stack.yaml
ADDED
|
@@ -0,0 +1,104 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SceneMachine model stack manifest
|
| 2 |
+
#
|
| 3 |
+
# Lists the exact upstream weight repos this version of SceneMachine
|
| 4 |
+
# depends on. All weights are hosted in the WindstormLabs HF org
|
| 5 |
+
# (Windstorm Labs is SceneMachine's parent organization; the same
|
| 6 |
+
# weight mirrors are shared by other Windstorm sub-projects).
|
| 7 |
+
#
|
| 8 |
+
# Update this file when the application code requires a new model
|
| 9 |
+
# version. The application reads this manifest at boot to know which
|
| 10 |
+
# WindstormLabs/* repos to download from.
|
| 11 |
+
|
| 12 |
+
manifest_version: 1
|
| 13 |
+
generated: 2026-05-13
|
| 14 |
+
scenemachine_min_version: "0.1"
|
| 15 |
+
|
| 16 |
+
# The canonical mirror org for SceneMachine's model dependencies.
|
| 17 |
+
# If you fork SceneMachine and want a different mirror, override this.
|
| 18 |
+
mirror_org: WindstormLabs
|
| 19 |
+
|
| 20 |
+
stacks:
|
| 21 |
+
|
| 22 |
+
# Wan 2.2 family — the primary video generation stack.
|
| 23 |
+
wan22:
|
| 24 |
+
description: |
|
| 25 |
+
Alibaba's Wan 2.2 14B family. Three sibling models share the
|
| 26 |
+
same VAE / text encoder / CLIP vision encoders.
|
| 27 |
+
|
| 28 |
+
text_to_video:
|
| 29 |
+
hf_repo: WindstormLabs/wan22-t2v-fp8
|
| 30 |
+
files:
|
| 31 |
+
- wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors
|
| 32 |
+
- wan2.2_t2v_low_noise_14B_fp8_scaled.safetensors
|
| 33 |
+
vram_gb: 22
|
| 34 |
+
use_case: |
|
| 35 |
+
Establishing shots, prompt-only scenes, anything without a
|
| 36 |
+
character reference or prior-frame continuity.
|
| 37 |
+
|
| 38 |
+
image_to_video:
|
| 39 |
+
hf_repo: WindstormLabs/wan22-i2v-fp8
|
| 40 |
+
files:
|
| 41 |
+
- wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors
|
| 42 |
+
- wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors
|
| 43 |
+
vram_gb: 24
|
| 44 |
+
use_case: |
|
| 45 |
+
Shot-to-shot continuity. Feed the last frame of the prior shot
|
| 46 |
+
as the seed image; produces a video that flows visually from it.
|
| 47 |
+
|
| 48 |
+
animate:
|
| 49 |
+
hf_repo: WindstormLabs/wan22-animate-bf16
|
| 50 |
+
files:
|
| 51 |
+
- wan2.2_animate_14B_bf16.safetensors
|
| 52 |
+
vram_gb: 32
|
| 53 |
+
use_case: |
|
| 54 |
+
Character-ID-preserving generation. Requires a reference image
|
| 55 |
+
of the character; the model preserves their identity across the
|
| 56 |
+
shot. Validated 1.7 min/shot with the Lightx2v speed LoRA.
|
| 57 |
+
|
| 58 |
+
shared_encoders:
|
| 59 |
+
hf_repo: WindstormLabs/wan22-encoders
|
| 60 |
+
files:
|
| 61 |
+
- wan_2.1_vae.safetensors # used by all 3 Wan stacks
|
| 62 |
+
- umt5_xxl_bf16_from_pth.safetensors # T5 text encoder, all 3
|
| 63 |
+
- sigclip_vision_patch14_384.safetensors # CLIP vision, I2V only
|
| 64 |
+
- clip_vision_h.safetensors # CLIP-ViT-H, Animate only (1280-dim)
|
| 65 |
+
|
| 66 |
+
speed_loras:
|
| 67 |
+
hf_repo: WindstormLabs/wan22-loras
|
| 68 |
+
files:
|
| 69 |
+
- Wan_2_2_I2V_A14B_HIGH_lightx2v_4step_lora_260412_rank_64_fp16.safetensors
|
| 70 |
+
- wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors
|
| 71 |
+
use_case: |
|
| 72 |
+
Kijai's Lightx2v 4-step distillation. When enabled, drops the
|
| 73 |
+
sampler from 30 steps to 4 with cfg=1.0 — 8.3× wallclock speedup
|
| 74 |
+
on Wan Animate. Verified to transfer cleanly to Animate when
|
| 75 |
+
paired with the correct embed chain + CLIP-ViT-H.
|
| 76 |
+
|
| 77 |
+
# LTX-2 — alternate cinematic stack (slower, comparable quality).
|
| 78 |
+
ltx2:
|
| 79 |
+
description: |
|
| 80 |
+
Lightricks LTX-2 19B Dev FP8 plus the Gemma text encoder.
|
| 81 |
+
|
| 82 |
+
dev_fp8:
|
| 83 |
+
hf_repo: WindstormLabs/ltx2-19b-fp8
|
| 84 |
+
files:
|
| 85 |
+
- ltx-2-19b-dev-fp8.safetensors
|
| 86 |
+
- model-00001-of-00005.safetensors # Gemma encoder shards
|
| 87 |
+
- model-00002-of-00005.safetensors
|
| 88 |
+
- model-00003-of-00005.safetensors
|
| 89 |
+
- model-00004-of-00005.safetensors
|
| 90 |
+
- model-00005-of-00005.safetensors
|
| 91 |
+
vram_gb: 28
|
| 92 |
+
|
| 93 |
+
# Hunyuan — reserved for Stack B implementation (not yet wired in app)
|
| 94 |
+
hunyuan:
|
| 95 |
+
description: |
|
| 96 |
+
Tencent HunyuanVideo + HunyuanVideo-I2V + HunyuanCustom partial
|
| 97 |
+
mirror. Reserved for SceneMachine Stack B (alternate character-
|
| 98 |
+
consistency path via Hunyuan's built-in identity preservation,
|
| 99 |
+
no LoRA needed). Provider workflow not yet implemented in the
|
| 100 |
+
app — weights are mirrored for future use.
|
| 101 |
+
hf_repo: WindstormLabs/hunyuan
|
| 102 |
+
license_note: |
|
| 103 |
+
Tencent HunyuanVideo Community License — check the upstream
|
| 104 |
+
repo for current terms before any commercial use.
|