DEMO_QUALITY_CONTRACT.json

{
  "schema_version": "demo_quality_contract.v1",
  "model_card_promise": "Text-to-video and video-to-video generation in a streaming fashion with ~1s time to first frame, powered by a 14B autoregressive diffusion model distilled from Wan 2.1.",
  "demo_task": "text-to-video",
  "primary_user_flow": "User enters a text prompt, adjusts blocks and steps, clicks Generate, and receives an MP4 video.",
  "examples_provided": true,
  "examples": [
    {
      "inputs": {
        "prompt": "a cat sitting on a boat",
        "num_blocks": 3,
        "num_inference_steps": 4,
        "seed": 42
      },
      "description": "Classic model-card example with conservative settings."
    },
    {
      "inputs": {
        "prompt": "a futuristic city at sunset",
        "num_blocks": 3,
        "num_inference_steps": 4,
        "seed": 123
      },
      "description": "Sci-fi scene preset."
    },
    {
      "inputs": {
        "prompt": "a panda playing guitar in a forest",
        "num_blocks": 2,
        "num_inference_steps": 4,
        "seed": 7
      },
      "description": "Lowest-cost preset suitable for a first call."
    }
  ],
  "canonical_smoke_example": {
    "api_name": "/generate",
    "inputs": {
      "prompt": "a cat sitting on a boat",
      "num_blocks": 1,
      "num_inference_steps": 2,
      "seed": 42
    },
    "expected_output_type": "video/mp4",
    "reason": "Single block (3 frames) with 2 diffusion steps is the smallest real forward-pass configuration that exercises the pipeline end-to-end while keeping duration as low as possible."
  },
  "real_inference_required": true,
  "real_inference_implemented": true,
  "fallback_or_diagnostic_only": false,
  "limitations_disclosed": true,
  "promise_fulfillment_risk": "High. The model requires ~50+ GB of VRAM for its weights alone in bf16, exceeding standard HF Space GPU tiers (48 GB large / 96 GB xlarge). Even if it loads, the advertised 11 fps and interactive streaming rely on hardware (B200 / H100 with FP8) beyond typical Space capacity. The Space honestly attempts inference and surfaces OOM/load errors instead of faking output."
}