spi / docs /model_registry.json
desiree's picture
Phase 1: cached model registry (research cache)
5f500c0 verified
Raw History Blame Contribute Delete
4.13 kB
{
"_meta": {
"purpose": "SPI research cache: model metadata so no subsystem is re-researched. Policy: check cache freshness before researching again.",
"checked_at": "2026-09-23",
"sources": ["official HF model cards", "diffusers docs (>=0.40.0)", "vendor repos"],
"diffusers_pin": ">=0.40.0"
},
"wan22": {
"family": "wan",
"provider": "Wan-AI (Alibaba)",
"role": "phase2_primary",
"license": "apache-2.0",
"variants": {
"t2v_a14b": {"hf": "Wan-AI/Wan2.2-T2V-A14B-Diffusers", "params_b": 27.0, "pipeline": "WanPipeline", "vram_gb": 80, "note": "MoE; guidance_scale=4.0, guidance_scale_2=3.0; FP8/offload for <80GB"},
"i2v_a14b": {"hf": "Wan-AI/Wan2.2-I2V-A14B-Diffusers", "pipeline": "WanImageToVideoPipeline", "vram_gb": 80},
"ti2v_5b": {"hf": "Wan-AI/Wan2.2-TI2V-5B", "params_b": 5.0, "pipeline": "WanPipeline", "vram_gb": 24, "note": "720p@24fps on RTX 4090"},
"t2v_1_3b": {"hf": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "params_b": 1.3, "pipeline": "WanPipeline", "vram_gb": 8.2},
"vace_14b": {"hf": "Wan-AI/Wan2.1-VACE-14B-diffusers", "params_b": 14.0, "pipeline": "WanVACEPipeline", "vram_gb": 40, "note": "pose/depth/mask/reference/extend/inpaint/outpaint; Wan2.1-based, not combinable with Wan2.2 weights in one call"},
"fun_control_5b": {"hf": "alibaba-pai/Wan2.2-Fun-5B-Control", "pipeline": "DiffSynth (alibaba-pai)", "note": "pose/depth/trajectory control signals"}
},
"diffusers": "stable >=0.40.0",
"finetune": {"lora": "DiffSynth-Studio (official), community diffusers scripts", "full": "DiffSynth-Studio"},
"fps": 16, "frames_per_clip": 81
},
"ltx2": {
"family": "ltx",
"provider": "Lightricks",
"role": "phase2_secondary_speed",
"license": {
"name": "LTX Open Weights License",
"hf_field": "other",
"gated": "auto",
"commercial_cap_usd": 10000000,
"note": "paid license required for Commercial Entities (>= $10M annual revenue); free for dev/eval and smaller entities"
},
"variants": {
"ltx2_19b": {"hf": "Lightricks/LTX-2", "params_b": 19.0, "pipeline": "LTX2Pipeline", "audio": true, "vram_gb_bf16": 48, "vram_gb_fp8": 24, "steps_distilled": 8, "guidance_scale_distilled": 1.0},
"ltx25": {"hf": "Lightricks/LTX-2.5-Diffusers", "params_b": 19.0, "pipeline": "LTX2Pipeline"},
"ltx_video_2b": {"hf": "Lightricks/LTX-Video", "params_b": 2.0, "pipeline": "LTXPipeline", "vram_gb": 10}
},
"control": "LTXConditionPipeline (image/video cond frames, decode_timestep=0.05); official IC-LoRAs: Ingredients, Clean-Plate, Union-Control, Slow-Motion, upscalers",
"speed": {"sec_per_step_720p_121f_h100": 1.22, "vs_wan22_14b": 22.3}
},
"cogvideox": {
"family": "cogvideox",
"provider": "zai-org (THUDM)",
"role": "fallback_only",
"license": {"cogvideox_2b": "apache-2.0", "cogvideox_5b": "custom (restrictive)"},
"variants": {
"cogvideox_5b": {"hf": "zai-org/CogVideoX-5b", "params_b": 5.57, "pipeline": "CogVideoXPipeline", "vram_gb": 12, "stale": true, "last_updated": "2024-11"},
"cogvideox_2b": {"hf": "zai-org/CogVideoX-2b", "params_b": 2.0, "pipeline": "CogVideoXPipeline", "vram_gb": 5}
},
"finetune": "diffusers examples/cogvideox LoRA"
},
"hunyuanvideo": {
"family": "hunyuanvideo",
"provider": "Tencent",
"role": "rejected",
"license": "custom (bans EU/UK/KR use)",
"note": "13B needs 45+ GB without offload; no first-class diffusers pipeline"
},
"vbench": {
"purpose": "reference evaluation suite (adapter, not the whole DirectorBench)",
"pip_package": "vbench",
"version": "0.1.5",
"license": "apache-2.0",
"repo": "Vchitect/VBench",
"dims": 16,
"note": "needs detectron2 from git for some dims; use as cheap CI regression gate"
},
"vbench2": {
"purpose": "release-time physics/interaction evals",
"pip_package": "vbench2",
"license": "apache-2.0",
"repo": "Vchitect/VBench-2.0",
"dims": 18,
"note": "heavy install (torch 2.5.1 pins, flash-attn 2.7.2, mmcv 2.2.0); reserve for release checkpoints"
}
}