[docs]: Open-H3-Regenerate-2K cookbook entry and a v2v cookbook workload

docs-site/lib/cookbook-pages.ts

@@ -48,6 +48,15 @@ export const recipePages = [
     tags: ["Reference to video", "Audio", "NVFP4"],
     featured: true,
   },
+  {
+    id: "open-h3-regenerate-2k",
+    family: "MiniMax H3",
+    title: "Open-H3-Regenerate-2K",
+    mark: "H3",
+    path: "/cookbook/minimax-h3/open-h3-regenerate-2k",
+    description: "Regenerate a 768P MiniMax-H3 draft at 2560x1440 with a three-step refiner.",
+    tags: ["Video to video", "2K", "3-step"],
+  },
   {
     id: "fastwan21",
     family: "Wan",

docs/assets/cookbook-recipes.json

@@ -760,6 +760,22 @@
       ],
       "limitations": ["Pass --reference-audio for an additional audio reference. The checked-in source defaults to four GPUs."]
     },
+    {
+      "id": "open-h3-regenerate-2k",
+      "family": "minimax_h3",
+      "stage": "inference",
+      "task": "Video to 2K video",
+      "label": "Open-H3-Regenerate-2K",
+      "summary": "Regenerate a 768P MiniMax-H3 draft at 2560x1440 with NVlabs' three-step refiner. The output is silent.",
+      "model": "Efficient-Large-Model/Open-H3-Regenerate-2K",
+      "source": "examples/inference/basic/basic_open_h3_regenerate_2k.py",
+      "command": ": \"${INPUT_VIDEO:?Set INPUT_VIDEO to a 768P MiniMax-H3 draft}\"\npython examples/inference/basic/basic_open_h3_regenerate_2k.py --video \"$INPUT_VIDEO\" --prompt \"A red fox trots through fresh snow in a pine forest at golden hour.\"",
+      "gpu_types": ["NVIDIA"],
+      "hardware": {"platform": "cuda", "gpu_count": 1, "evidence": "source-configured"},
+      "evidence": "Source-backed",
+      "expected_artifact": "2560x1440 MP4 at outputs/open_h3_regenerate_2k/regen2k_0.mp4",
+      "related": ["minimax-h3-t2v", "fasth3-8step-v2-cuda"]
+    },
     {
       "id": "fasth3-lora-preview",
       "family": "minimax_h3",

docs/cookbook/options.yaml

@@ -44,7 +44,7 @@ runtimes:
         host: 0.0.0.0
         port: 8000
         served_model_name: null
-    workloads: [t2v, i2v]
+    workloads: [t2v, i2v, v2v]
     options: &cuda_options
       server.served_model_name:
         anyOf:

docs/cookbook/recipes/open-h3-regenerate-2k.yaml

@@ -0,0 +1,18 @@
+title: Open-H3-Regenerate-2K
+summary: Regenerate a 768P MiniMax-H3 draft at 2560x1440 with three refine steps.
+
+deployments:
+  cuda-rest:
+    runtime: fastvideo-cuda-rest
+    guide: docs/inference/open_h3_regenerate_2k.md
+    workload: v2v
+    defaults: examples/serving/openai_open_h3_regenerate_2k.yaml
+    requirements:
+      - 'UV_TORCH_BACKEND=cu130 uv pip install -e ".[fasth3]"'
+      - "One GPU with about 150 GB free (B200 or GB200). Send the draft as video_reference; local paths must sit under server.media_roots."
+      - "Optional NVFP4 decoder (Blackwell only): hf download FastVideo/FastH3-NVFP4-VAE-Decoder --include 'quality-36L/decoder-nvfp4.safetensors' --local-dir fasth3-nvfp4-vae"
+    options:
+      - server.*
+      - generator.engine.offload.text_encoder
+      - generator.pipeline.experimental.h3_vae_nvfp4_decoder
+      - default_request.sampling.seed

docs/js/cookbook-workloads.mjs

@@ -32,6 +32,14 @@ export const WORKLOADS = Object.freeze({
     extension: "wav",
     media_type: "audio/wav",
   }),
+  v2v: Object.freeze({
+    input: "video",
+    output: "video",
+    endpoint: "/v1/videos/sync",
+    extension: "mp4",
+    media_type: "video/mp4",
+    reference: "video_reference",
+  }),
   v2a: Object.freeze({
     input: "video",
     output: "audio",

docs/tests/cookbook-config.test.mjs

@@ -488,6 +488,7 @@ test("sample requests follow shared workload media", () => {
     t2i: { endpoint: /\/v1\/images\/generations/, output: /--output output\.json/, reference: null },
     t2a: { endpoint: /\/v1\/videos\/sync/, output: /--output output\.wav/, reference: null },
     v2a: { endpoint: /\/v1\/videos\/sync/, output: /--output output\.wav/, reference: "video_reference" },
+    v2v: { endpoint: /\/v1\/videos\/sync/, output: /--output output\.mp4/, reference: "video_reference" },
   };
   for (const [workload, expected] of Object.entries(expectations)) {
     const data = catalog();