Add 3-beat 10s H3 quality ref test

custom/chris/minimax-h3-two-beat-savevideo-test/README.md

@@ -1,10 +1,10 @@
-# MiniMax H3 Two-Beat SaveVideo Test
+# MiniMax H3 SaveVideo Tests
 
 Practical test flow for `DumasH3LongVideos` that:
 
 - uses `LoadImage` refs derived from images stored in this repo
-- renders exactly 2 beats
+- renders short beat-based clips
-- stays under 10 seconds total
+- stays near or under 10 seconds total
 - converts the returned image batch with `CreateVideo`
 - saves the final file with `SaveVideo`
 
@@ -22,18 +22,17 @@ Upload those repo images into the live ComfyUI input folder with these names:
 - `default-character-reference.jpg`
 - `default-location-reference.jpg`
 
-The included `api-prompt.json` expects those names in `LoadImage`.
+The included baseline prompts expect those names in `LoadImage`.
+
+The new reference-quality prompt uses these live input names instead:
+
+- `chris-character-reference.jpg`
+- `chris-location-reference.jpg`
 
 ## Flow summary
 
 `LoadImage x2 -> DiffusionModelLoaderKJ / CLIPLoader / VAELoader x2 -> DumasH3LongVideos -> CreateVideo -> SaveVideo`
 
-## Timing
-
-- `shot_seconds = 3`
-- `2 beats`
-- total target runtime = `6s`
-
 ## Variants
 
 - `api-prompt.json`
@@ -46,7 +45,22 @@ The included `api-prompt.json` expects those names in `LoadImage`.
   - `2` beats at `4s` each
   - total target runtime = `8s`
 
+- `api-prompt-3beat-10s-480p-quality.json`
+  - reference-quality follow-up for today's H3 live test
+  - `3` beats with explicit per-beat timing: `3s`, `3s`, `4s`
+  - total target runtime = `10s`
+  - expects the provided live refs to be uploaded as:
+    - `chris-character-reference.jpg`
+    - `chris-location-reference.jpg`
+  - quality-oriented settings:
+    - `864x480` (`0.4 MP` at `16:9`)
+    - `20` steps
+    - `euler` sampler
+    - `beta` scheduler
+    - H3 turbo LoRA enabled
+
 ## Notes
 
 - This is a practical API flow, not a promoted master workflow.
-- The H3 node is configured for a cheap proof run, not final quality.
+- `api-prompt.json` and `api-prompt-realistic-16x9.json` remain cheap proof-run variants.
+- `api-prompt-3beat-10s-480p-quality.json` is the quality-biased reference test.

custom/chris/minimax-h3-two-beat-savevideo-test/api-prompt-3beat-10s-480p-quality.json

@@ -0,0 +1,175 @@
+{
+  "1": {
+    "class_type": "LoadImage",
+    "inputs": {
+      "image": "chris-character-reference.jpg"
+    },
+    "_meta": {
+      "title": "Character Ref"
+    }
+  },
+  "2": {
+    "class_type": "LoadImage",
+    "inputs": {
+      "image": "chris-location-reference.jpg"
+    },
+    "_meta": {
+      "title": "Location Ref"
+    }
+  },
+  "10": {
+    "class_type": "DiffusionModelLoaderKJ",
+    "inputs": {
+      "model_name": "minimax_h3_ref2va_pruned_int8_convrot.safetensors",
+      "weight_dtype": "default",
+      "compute_dtype": "default",
+      "patch_cublaslinear": false,
+      "sage_attention": "disabled",
+      "enable_fp16_accumulation": false
+    },
+    "_meta": {
+      "title": "Load Model"
+    }
+  },
+  "11": {
+    "class_type": "CLIPLoader",
+    "inputs": {
+      "clip_name": "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors",
+      "type": "minimax",
+      "device": "default"
+    },
+    "_meta": {
+      "title": "Load CLIP"
+    }
+  },
+  "12": {
+    "class_type": "VAELoader",
+    "inputs": {
+      "vae_name": "minimax_h3_video_vae_int8_convrot.safetensors"
+    },
+    "_meta": {
+      "title": "Video VAE"
+    }
+  },
+  "13": {
+    "class_type": "VAELoader",
+    "inputs": {
+      "vae_name": "minimax_h3_audio_vae_fp32.safetensors"
+    },
+    "_meta": {
+      "title": "Audio VAE"
+    }
+  },
+  "14": {
+    "class_type": "MiniMaxH3TurboLoRA",
+    "inputs": {
+      "model": [
+        "10",
+        0
+      ],
+      "lora_name": "minimax\\minimax_h3_ref2v_lightx2v_turbo_4step_v0.1_resized_avg_rank_20_bf16.safetensors",
+      "strength": 1,
+      "low_vram": false
+    },
+    "_meta": {
+      "title": "MiniMax H3 Turbo LoRA"
+    }
+  },
+  "20": {
+    "class_type": "DumasH3LongVideos",
+    "inputs": {
+      "model": [
+        "14",
+        0
+      ],
+      "clip": [
+        "11",
+        0
+      ],
+      "vae": [
+        "12",
+        0
+      ],
+      "audio_vae": [
+        "13",
+        0
+      ],
+      "ref_image_1": [
+        "1",
+        0
+      ],
+      "ref_image_2": [
+        "2",
+        0
+      ],
+      "prompt": "Subject 1 is the same woman shown in <Picture 1>. Preserve her facial structure, body proportions, sleek dark high ponytail, and confident editorial presence throughout. Use <Picture 1> only for Subject 1 identity and outfit information. Ignore the background, lighting, pose framing, and studio environment shown in <Picture 1>. Use <Picture 2> as the sole environment and street-layout reference for the location.\n\nseconds: 3\nIn the rain-dark street shown in <Picture 2>, Subject 1 stands still near the curve of the cobbled lane, turns her head slightly, and takes in the surroundings in a steady medium-full shot.\n\nseconds: 3\nSubject 1 begins walking forward through the wet street shown in <Picture 2> with calm, deliberate confidence while the camera tracks gently backward to keep her framed cleanly.\n\nseconds: 4\nSubject 1 continues down the curved cobbled street, glancing toward the storefronts as the camera eases into a subtle side-follow and then settles into a composed closing hold.",
+      "anchor_override": "sleek dark high ponytail, realistic proportions, polished fashion-editorial presence",
+      "character_memory": "white cropped blouse with a high sculpted collar, black high-waisted trousers with a sheer corset panel and gold chain details, gold hoop earrings",
+      "resolution": "16:9",
+      "megapixels": 0.4,
+      "steps": 20,
+      "cfg": 1,
+      "sampler_name": "euler",
+      "scheduler": "beta",
+      "seed": 17,
+      "plan_only": false,
+      "shot_seconds": 4,
+      "ref_mode": "where tagged",
+      "per_beat_length": true,
+      "mute_nonspeech_audio": true,
+      "cleanup_between_shots": true,
+      "allow_res_backoff": true,
+      "decode_tile_frames": 8,
+      "decode_tile_size": 256,
+      "auto_soundscape": "fill if blank"
+    },
+    "_meta": {
+      "title": "Dumas H3 Long Videos"
+    }
+  },
+  "30": {
+    "class_type": "CreateVideo",
+    "inputs": {
+      "images": [
+        "20",
+        0
+      ],
+      "fps": 24,
+      "audio": [
+        "20",
+        1
+      ],
+      "bit_depth": 8
+    },
+    "_meta": {
+      "title": "CreateVideo"
+    }
+  },
+  "31": {
+    "class_type": "SaveVideo",
+    "inputs": {
+      "video": [
+        "30",
+        0
+      ],
+      "filename_prefix": "MiniMax/dumas-h3-3beat-10s-480p-quality",
+      "format": "auto",
+      "codec": "auto"
+    },
+    "_meta": {
+      "title": "SaveVideo"
+    }
+  },
+  "32": {
+    "class_type": "PreviewAny",
+    "inputs": {
+      "source": [
+        "20",
+        2
+      ]
+    },
+    "_meta": {
+      "title": "Info"
+    }
+  }
+}