Add fixed hard generation suite and visual analysis

.gitattributes

@@ -48,3 +48,35 @@ examples/t2i/natural-ours.png filter=lfs diff=lfs merge=lfs -text
 examples/t2i/product-community.png filter=lfs diff=lfs merge=lfs -text
 examples/t2i/product-comparison.png filter=lfs diff=lfs merge=lfs -text
 examples/t2i/product-ours.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/edit/audio-interface-color/contact-sheet.jpg filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/edit/audio-interface-color/transition-grid.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/edit/reading-chair-material/contact-sheet.jpg filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/edit/reading-chair-material/transition-grid.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/edit-overview.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/dense-ui-dashboard/community.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/dense-ui-dashboard/comparison.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/dense-ui-dashboard/ours.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/exploded-camera-diagram/community.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/exploded-camera-diagram/comparison.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/exploded-camera-diagram/ours.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/library-interior/community.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/library-interior/comparison.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/library-interior/ours.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/rainy-neon-street/community.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/rainy-neon-street/comparison.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/rainy-neon-street/ours.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/scene-text-package/community.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/scene-text-package/comparison.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/scene-text-package/ours.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/skeleton-watch/community.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/skeleton-watch/comparison.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/skeleton-watch/ours.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/wet-kingfisher/community.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/wet-kingfisher/comparison.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/wet-kingfisher/ours.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/woven-textile-macro/community.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/woven-textile-macro/comparison.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i/woven-textile-macro/ours.png filter=lfs diff=lfs merge=lfs -text
+examples/hard-v1/t2i-overview.png filter=lfs diff=lfs merge=lfs -text
+source-fixtures-hard-v1/audio-interface.png filter=lfs diff=lfs merge=lfs -text
+source-fixtures-hard-v1/reading-room.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -97,6 +97,32 @@ and composition, lighting, material, or geometry can drift during the
 transition. A useful result may depend on searching multiple slices. That is a
 fundamental limitation compared with a native image-edit model.
 
+### Fixed hard suite: 8 T2I + 2 edits
+
+This additional suite was preregistered before generation. It uses eight fixed
+T2I prompts and seeds plus two fixed source-conditioned edits. There was no
+seed search, prompt rewrite after review, or best-frame selection, and the
+failures are intentionally retained.
+
+![Eight fixed hard T2I comparisons](examples/hard-v1/t2i-overview.png)
+
+The 500K decoder is consistently much sharper and more coherent than the
+community decoder across natural detail, technical diagrams, product detail,
+architecture, repeated texture, and dense UI. Exact text remains a clear
+failure mode: package copy, diagram labels, signs, and UI headings are often
+incorrect even when the surrounding structure is strong.
+
+![Two fixed hard edit transitions](examples/hard-v1/edit-overview.png)
+
+Both edits apply the requested color or material change. They also expose the
+main FL2VA limitation: stronger edit slices increasingly alter typography,
+lighting, controls, background objects, and other content that should have
+been preserved.
+
+See the [case-by-case visual review](examples/hard-v1/visual-review.md), the
+[fixed prompt and seed registration](hard-example-prompts-v1.json), and the
+[path-sanitized checksum manifest](examples/hard-v1/public-manifest.json).
+
 ## Fixed reconstruction benchmark
 
 The benchmark contains 288 held-out cases spanning photographs, text,

examples/hard-v1/edit/audio-interface-color/smoke-report.json

@@ -0,0 +1,137 @@
+{
+  "completed_unix": 1787198224.3000648,
+  "conclusion": "pending_visual_review",
+  "config": {
+    "comparison": "source_then_ordered_temporal_slices",
+    "decoder": {
+      "decoder_id": "published-step-00062500-6c5ff2caa8fade67",
+      "decoder_kind": "published",
+      "source_config_fingerprint": "225279cbb4cd1e2db94b5b2ae15e123a921fefada0782e5c68d35302e4b4d0a9",
+      "source_decoder_tensor_count": 585,
+      "source_identity_sha256": "6c5ff2caa8fade6769f4dd53ee244f77a06652c8cb66b2fedc93f75046d9f001",
+      "source_resume_sha256": "40e94c1bef5a8467b5d41ec7c039212c6246b67b37f61b9ae802f4f48dfb23b4",
+      "source_run_name": "minimax_h3_t1_hardwarm_scale500k_curriculum256_384_512_768_1024_20260818_v2",
+      "source_samples_seen": 500000,
+      "source_step": 62500,
+      "weights_kind": "decoder_only_safetensors"
+    },
+    "hosted_image_logging": false,
+    "inference_name": "minimax_h3_release_hard_v1_edit_audio-interface-color",
+    "latent_config_fingerprint": "1332a593b726dd9a2ba955d65d880cfc989bd19fd58eb3bb5dc2ab654d89a5e6",
+    "latent_indices": [
+      0,
+      9,
+      18,
+      27,
+      36
+    ],
+    "latent_sha256": "f3d6bb4a1af9a0b86bd232834025a4fb8133362de018da5da855e14a8d3701ba",
+    "schema_version": 2,
+    "selection_claim": "diagnostic_only_not_best_frame_selection"
+  },
+  "config_fingerprint": "4769b802a1616d4143eaeccea4c0040261eb7a02abaec85b1f528ff1c4961c79",
+  "contact_sheet": {
+    "filename": "contact-sheet.jpg",
+    "height": 306,
+    "sha256": "e00258ca8e19f32019e96358315d4241e4a4c888f1de7d9c63f6947debff8d17",
+    "width": 2880
+  },
+  "rows": [
+    {
+      "clip_fraction": 0.0027024613227695227,
+      "elapsed_seconds": 1.503443842000003,
+      "image": {
+        "bytes": 438676,
+        "filename": "latent-0000.png",
+        "height": 544,
+        "mode": "RGB",
+        "sha256": "30811ce01afa7fd900ceaea3d01764299b31918ff7083162275aa271fb93be68",
+        "width": 960
+      },
+      "latent_index": 0,
+      "source_mae": 0.01330658420920372,
+      "source_psnr": 28.693299302667114,
+      "unclamped_maximum": 1.082937479019165,
+      "unclamped_minimum": -0.105837881565094
+    },
+    {
+      "clip_fraction": 0.09712201356887817,
+      "elapsed_seconds": 0.2297019429999949,
+      "image": {
+        "bytes": 454834,
+        "filename": "latent-0009.png",
+        "height": 544,
+        "mode": "RGB",
+        "sha256": "5f1a1c92cbe5f2425becbbd6b4c9408f21575e30123a6be19bb8f30776f01412",
+        "width": 960
+      },
+      "latent_index": 9,
+      "source_mae": 0.172123521566391,
+      "source_psnr": 13.940534926087018,
+      "unclamped_maximum": 1.1424375772476196,
+      "unclamped_minimum": -0.14832812547683716
+    },
+    {
+      "clip_fraction": 0.08207210153341293,
+      "elapsed_seconds": 0.2492334559999989,
+      "image": {
+        "bytes": 478239,
+        "filename": "latent-0018.png",
+        "height": 544,
+        "mode": "RGB",
+        "sha256": "3770a48d630987a1e6ef031e1f14c93553c291255312c681433456fe866dd218",
+        "width": 960
+      },
+      "latent_index": 18,
+      "source_mae": 0.20812328159809113,
+      "source_psnr": 11.44990247106402,
+      "unclamped_maximum": 1.1087499856948853,
+      "unclamped_minimum": -0.12441989779472351
+    },
+    {
+      "clip_fraction": 0.0850362554192543,
+      "elapsed_seconds": 0.23954891700000047,
+      "image": {
+        "bytes": 489010,
+        "filename": "latent-0027.png",
+        "height": 544,
+        "mode": "RGB",
+        "sha256": "cbd5c626fca1b3cb40bec03bfd5bfcf1eec86e6e749407f13c49b5899e8c2eca",
+        "width": 960
+      },
+      "latent_index": 27,
+      "source_mae": 0.234516903758049,
+      "source_psnr": 10.281407725243305,
+      "unclamped_maximum": 1.1293398141860962,
+      "unclamped_minimum": -0.10420510172843933
+    },
+    {
+      "clip_fraction": 0.06332146376371384,
+      "elapsed_seconds": 0.24810163000000074,
+      "image": {
+        "bytes": 515665,
+        "filename": "latent-0036.png",
+        "height": 544,
+        "mode": "RGB",
+        "sha256": "cb726c112c223635843f7b4759f3a6bc9c8839cc0ca87f2a82b3c107f4491776",
+        "width": 960
+      },
+      "latent_index": 36,
+      "source_mae": 0.239557683467865,
+      "source_psnr": 10.082632193133861,
+      "unclamped_maximum": 1.1715527772903442,
+      "unclamped_minimum": -0.15825781226158142
+    }
+  ],
+  "runtime": {
+    "function_body_elapsed_seconds": 26.949949796000002,
+    "gpu_count": 1,
+    "gpu_peak_allocated_bytes": 9922627584,
+    "gpu_peak_reserved_bytes": 9950986240,
+    "gpu_total_memory_bytes": 150110011392,
+    "gpu_type": "H200",
+    "process_peak_rss_bytes": 15207927808
+  },
+  "schema_version": 2,
+  "status": "complete"
+}

examples/hard-v1/edit/reading-chair-material/smoke-report.json

@@ -0,0 +1,137 @@
+{
+  "completed_unix": 1787198224.4727833,
+  "conclusion": "pending_visual_review",
+  "config": {
+    "comparison": "source_then_ordered_temporal_slices",
+    "decoder": {
+      "decoder_id": "published-step-00062500-6c5ff2caa8fade67",
+      "decoder_kind": "published",
+      "source_config_fingerprint": "225279cbb4cd1e2db94b5b2ae15e123a921fefada0782e5c68d35302e4b4d0a9",
+      "source_decoder_tensor_count": 585,
+      "source_identity_sha256": "6c5ff2caa8fade6769f4dd53ee244f77a06652c8cb66b2fedc93f75046d9f001",
+      "source_resume_sha256": "40e94c1bef5a8467b5d41ec7c039212c6246b67b37f61b9ae802f4f48dfb23b4",
+      "source_run_name": "minimax_h3_t1_hardwarm_scale500k_curriculum256_384_512_768_1024_20260818_v2",
+      "source_samples_seen": 500000,
+      "source_step": 62500,
+      "weights_kind": "decoder_only_safetensors"
+    },
+    "hosted_image_logging": false,
+    "inference_name": "minimax_h3_release_hard_v1_edit_reading-chair-material",
+    "latent_config_fingerprint": "d626d69c17aefedae21e4be20caa73d5444ed114f8867a1cb6d6843f3046cdad",
+    "latent_indices": [
+      0,
+      9,
+      18,
+      27,
+      36
+    ],
+    "latent_sha256": "60e833c53acc9c6c4bfb0bc02d76b632ae736fdcc899562d198a45f3e62cd4ea",
+    "schema_version": 2,
+    "selection_claim": "diagnostic_only_not_best_frame_selection"
+  },
+  "config_fingerprint": "23a72e251b9ef5e290fc42c536535dfaa3f3ec73d4f4c574c76fb12b67edc777",
+  "contact_sheet": {
+    "filename": "contact-sheet.jpg",
+    "height": 306,
+    "sha256": "b806b1357f672d4c8cd0fd69e9a80f7131eb686025ee0025b2803b3172404a57",
+    "width": 2880
+  },
+  "rows": [
+    {
+      "clip_fraction": 0.011979167349636555,
+      "elapsed_seconds": 1.640508336,
+      "image": {
+        "bytes": 606063,
+        "filename": "latent-0000.png",
+        "height": 544,
+        "mode": "RGB",
+        "sha256": "58e3598dd00ccdcf46dbb2b289a7906022e672f8227fa625ebd1e8c2ca3979c1",
+        "width": 960
+      },
+      "latent_index": 0,
+      "source_mae": 0.026626599952578545,
+      "source_psnr": 26.557573561183325,
+      "unclamped_maximum": 1.0905976295471191,
+      "unclamped_minimum": -0.13356834650039673
+    },
+    {
+      "clip_fraction": 0.010850694961845875,
+      "elapsed_seconds": 0.3001589170000045,
+      "image": {
+        "bytes": 839084,
+        "filename": "latent-0009.png",
+        "height": 544,
+        "mode": "RGB",
+        "sha256": "13cdf638c80a200dc61b9159bb84cd3f179521a3852af6265f719f727be2c682",
+        "width": 960
+      },
+      "latent_index": 9,
+      "source_mae": 0.15718692541122437,
+      "source_psnr": 13.96486529387079,
+      "unclamped_maximum": 1.069574236869812,
+      "unclamped_minimum": -0.14946874976158142
+    },
+    {
+      "clip_fraction": 0.00662977434694767,
+      "elapsed_seconds": 0.3012888800000013,
+      "image": {
+        "bytes": 821429,
+        "filename": "latent-0018.png",
+        "height": 544,
+        "mode": "RGB",
+        "sha256": "a2ade1cd26c47e8d258b5d160fe503bd1507e022d86ef249434e440d362d488e",
+        "width": 960
+      },
+      "latent_index": 18,
+      "source_mae": 0.15596969425678253,
+      "source_psnr": 13.951380626080685,
+      "unclamped_maximum": 1.0543692111968994,
+      "unclamped_minimum": -0.11167576909065247
+    },
+    {
+      "clip_fraction": 0.007935687899589539,
+      "elapsed_seconds": 0.3091636890000018,
+      "image": {
+        "bytes": 795034,
+        "filename": "latent-0027.png",
+        "height": 544,
+        "mode": "RGB",
+        "sha256": "d231f8ab701e71b691b5bf8cd3146667ea6283f113ba55d47d4e9fbc70e8c17b",
+        "width": 960
+      },
+      "latent_index": 27,
+      "source_mae": 0.15370646119117737,
+      "source_psnr": 14.21605408844046,
+      "unclamped_maximum": 1.0735292434692383,
+      "unclamped_minimum": -0.16704687476158142
+    },
+    {
+      "clip_fraction": 0.008613537065684795,
+      "elapsed_seconds": 0.29245957299999503,
+      "image": {
+        "bytes": 778236,
+        "filename": "latent-0036.png",
+        "height": 544,
+        "mode": "RGB",
+        "sha256": "11be45c66111a517c2929a902e2978eaaf3cadfd84e39593cb77f54799b6183e",
+        "width": 960
+      },
+      "latent_index": 36,
+      "source_mae": 0.1412827968597412,
+      "source_psnr": 14.677192398508993,
+      "unclamped_maximum": 1.1044628620147705,
+      "unclamped_minimum": -0.14551368355751038
+    }
+  ],
+  "runtime": {
+    "function_body_elapsed_seconds": 25.251890952999997,
+    "gpu_count": 1,
+    "gpu_peak_allocated_bytes": 9922627584,
+    "gpu_peak_reserved_bytes": 9950986240,
+    "gpu_total_memory_bytes": 150110011392,
+    "gpu_type": "H200",
+    "process_peak_rss_bytes": 15210901504
+  },
+  "schema_version": 2,
+  "status": "complete"
+}

examples/hard-v1/public-manifest.json

@@ -0,0 +1,400 @@
+{
+  "decoder": {
+    "release_sha256": "6c5ff2caa8fade6769f4dd53ee244f77a06652c8cb66b2fedc93f75046d9f001",
+    "source_step": 62500,
+    "training_exposures": 500000
+  },
+  "image_edits": [
+    {
+      "decoded_latent_indices": [
+        0,
+        9,
+        18,
+        27,
+        36
+      ],
+      "files": [
+        {
+          "bytes": 186838,
+          "height": 306,
+          "mode": "RGB",
+          "path": "examples/hard-v1/edit/audio-interface-color/contact-sheet.jpg",
+          "sha256": "e00258ca8e19f32019e96358315d4241e4a4c888f1de7d9c63f6947debff8d17",
+          "width": 2880
+        },
+        {
+          "bytes": 790273,
+          "height": 632,
+          "mode": "RGB",
+          "path": "examples/hard-v1/edit/audio-interface-color/transition-grid.png",
+          "sha256": "fe5d6321cecbdab4503ad2e950d8f4c7ad0182bc7dc8783cf55dd4eaf31d6829",
+          "width": 1440
+        },
+        {
+          "bytes": 4678,
+          "path": "examples/hard-v1/edit/audio-interface-color/smoke-report.json",
+          "sha256": "45b4eca060382415e932b84818037683160d70755e3b90e8f6c4d3ab37f825db"
+        }
+      ],
+      "id": "audio-interface-color",
+      "instruction": "Change only the cobalt-blue metal enclosure to warm matte orange anodized aluminum. Preserve the black front panel, exact text including AERON 07, every knob and port, LED colors and states, ventilation holes, camera angle, crop, lighting, shadows, and background.",
+      "seed": 2026081931,
+      "source": "source-fixtures-hard-v1/audio-interface.png",
+      "source_file": {
+        "bytes": 2012461,
+        "height": 962,
+        "mode": "RGB",
+        "path": "source-fixtures-hard-v1/audio-interface.png",
+        "sha256": "108a9d4b1975d29a80d85f7c6097495677732aeae65ceb16a0e66342d5bccc3f",
+        "width": 1635
+      },
+      "steps": 25
+    },
+    {
+      "decoded_latent_indices": [
+        0,
+        9,
+        18,
+        27,
+        36
+      ],
+      "files": [
+        {
+          "bytes": 309079,
+          "height": 306,
+          "mode": "RGB",
+          "path": "examples/hard-v1/edit/reading-chair-material/contact-sheet.jpg",
+          "sha256": "b806b1357f672d4c8cd0fd69e9a80f7131eb686025ee0025b2803b3172404a57",
+          "width": 2880
+        },
+        {
+          "bytes": 1351075,
+          "height": 632,
+          "mode": "RGB",
+          "path": "examples/hard-v1/edit/reading-chair-material/transition-grid.png",
+          "sha256": "1d3d99cfb0f82e4c8c3ff624b1c04720031638f501ec961cb0aaffde2c84187d",
+          "width": 1440
+        },
+        {
+          "bytes": 4682,
+          "path": "examples/hard-v1/edit/reading-chair-material/smoke-report.json",
+          "sha256": "8ecc9bab5c7667e68ac4e10d3b26311e15039928387aba3ec335bbc1b8387181"
+        }
+      ],
+      "id": "reading-chair-material",
+      "instruction": "Change only the mustard-yellow chair upholstery to deep emerald-green herringbone velvet. Preserve the chair shape, room geometry, poster text, lamp, plant, books, window, rain, rug grid, table, glass, lighting, camera angle, and crop.",
+      "seed": 2026081932,
+      "source": "source-fixtures-hard-v1/reading-room.png",
+      "source_file": {
+        "bytes": 2874468,
+        "height": 1024,
+        "mode": "RGB",
+        "path": "source-fixtures-hard-v1/reading-room.png",
+        "sha256": "eee847c1cc5e311d3c62405820b65a0d17ae599dc440eadc8d47eba236b4ab18",
+        "width": 1536
+      },
+      "steps": 25
+    }
+  ],
+  "overview_files": [
+    {
+      "bytes": 2685110,
+      "height": 1376,
+      "mode": "RGB",
+      "path": "examples/hard-v1/t2i-overview.png",
+      "sha256": "1b009908622901ec74d53f4a0b0b1ef063426f576b4113e387be1307834b63de",
+      "width": 1920
+    },
+    {
+      "bytes": 1102872,
+      "height": 467,
+      "mode": "RGB",
+      "path": "examples/hard-v1/edit-overview.png",
+      "sha256": "5240d95375863c5b2cb038dbb7a15695c8ababc755c4e33892d892656f0b004c",
+      "width": 1920
+    }
+  ],
+  "review": {
+    "failures_retained": true,
+    "human_preference_study": false,
+    "kind": "AI visual review"
+  },
+  "schema_version": 1,
+  "selection_policy": {
+    "edit_latent_indices": [
+      0,
+      9,
+      18,
+      27,
+      36
+    ],
+    "prompts_fixed_before_generation": true,
+    "seed_search": false,
+    "show_failures": true,
+    "t2i_latent_index": 0
+  },
+  "spec": {
+    "bytes": 5004,
+    "path": "hard-example-prompts-v1.json",
+    "sha256": "a2d8d258778189f2abadf088eb1e4091e845a4d45c7ddf540b126c688c5fe129"
+  },
+  "status": "complete",
+  "text_to_image": [
+    {
+      "files": [
+        {
+          "bytes": 333778,
+          "height": 544,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/scene-text-package/community.png",
+          "sha256": "e77a4273090a97ea982c2f2d8b2427e23448a8cde39fe4d5657ee196003c1baf",
+          "width": 960
+        },
+        {
+          "bytes": 276606,
+          "height": 544,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/scene-text-package/ours.png",
+          "sha256": "c58a104b31306b2b72f4fa5680ff4faf632c698496a36c0973ce9676c45fbe4a",
+          "width": 960
+        },
+        {
+          "bytes": 567625,
+          "height": 596,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/scene-text-package/comparison.png",
+          "sha256": "0f10228a989852e956448eee8cf97afbec68a75f7584df89f4f5d62e0968bd44",
+          "width": 1920
+        }
+      ],
+      "id": "scene-text-package",
+      "prompt": "Premium studio product photograph of a small matte cream coffee package standing upright on a dark stone pedestal, centered, full package visible, crisp folded edges and subtle paper fibers. The package must display the exact large title 'NORTH / 27', the exact subtitle 'HIGH ALTITUDE COFFEE', and three small lines 'ORIGIN: KENYA', 'ROAST: LIGHT', '250 g'. Soft directional light, realistic shadows, no people, no extra text, no logos, no watermark.",
+      "seed": 2026081921,
+      "steps": 30
+    },
+    {
+      "files": [
+        {
+          "bytes": 558627,
+          "height": 544,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/wet-kingfisher/community.png",
+          "sha256": "5fd6971194b03dca3baa305a5d31ad59fb95a7a79ea27a514d2c2d893dabc5d5",
+          "width": 960
+        },
+        {
+          "bytes": 511112,
+          "height": 544,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/wet-kingfisher/ours.png",
+          "sha256": "0f2a9d35ae58903b278caefb6bf4a34fdf6bf418933d4e3edb925b85f85c146e",
+          "width": 960
+        },
+        {
+          "bytes": 1035170,
+          "height": 596,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/wet-kingfisher/comparison.png",
+          "sha256": "744a8c509fbd3368eb6addf95ae4bbe879de9520189a799abeed0506aee23095",
+          "width": 1920
+        }
+      ],
+      "id": "wet-kingfisher",
+      "prompt": "Ultra-detailed natural wildlife photograph of a wet kingfisher perched on a mossy branch immediately after rain, full bird visible, individual blue and orange feather barbs, tiny water droplets, delicate claws, realistic eye reflection, soft forest bokeh, natural overcast light, accurate anatomy, no text, no watermark.",
+      "seed": 2026081922,
+      "steps": 30
+    },
+    {
+      "files": [
+        {
+          "bytes": 377786,
+          "height": 544,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/exploded-camera-diagram/community.png",
+          "sha256": "2ec9c8d352cd4aca5c861eb77aef439fe42de4e7865af235b97047d6cb506bfc",
+          "width": 960
+        },
+        {
+          "bytes": 244023,
+          "height": 544,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/exploded-camera-diagram/ours.png",
+          "sha256": "8a01fb08c3ac963881ba2640a212eeca42d97df4af5fc3bdaf98e1c0d4f64061",
+          "width": 960
+        },
+        {
+          "bytes": 606487,
+          "height": 596,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/exploded-camera-diagram/comparison.png",
+          "sha256": "f1c488412b3cee4a0f285ddcb7e0b1b98bb858546daf28ef2c033eb44aa5cd5e",
+          "width": 1920
+        }
+      ],
+      "id": "exploded-camera-diagram",
+      "prompt": "Clean industrial-design exploded technical diagram of a compact mirrorless camera on white paper, precise black line work with pale gray shading, twelve aligned mechanical components, thin leader lines, dimension arrows, consistent perspective, and exact labels 'LENS', 'SENSOR', 'SHUTTER', 'BODY', 'BATTERY'. Entire diagram visible with generous margins, no hands, no watermark.",
+      "seed": 2026081923,
+      "steps": 30
+    },
+    {
+      "files": [
+        {
+          "bytes": 659246,
+          "height": 544,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/rainy-neon-street/community.png",
+          "sha256": "ae88cc120c2ce5fe11c647ff40fe9d21d5e658e513228896600c1e1e0e2fa459",
+          "width": 960
+        },
+        {
+          "bytes": 640689,
+          "height": 544,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/rainy-neon-street/ours.png",
+          "sha256": "41f7743c395c66bfa2f7c9582cc65b33b105999c5f42ddb36fe70f58b5c777a4",
+          "width": 960
+        },
+        {
+          "bytes": 1277624,
+          "height": 596,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/rainy-neon-street/comparison.png",
+          "sha256": "59ba60cd14395bb808150d11f9c841e1f4c9af97e38033aed458451e84b6a18c",
+          "width": 1920
+        }
+      ],
+      "id": "rainy-neon-street",
+      "prompt": "Photorealistic rainy night street in a dense modern city, wide-angle eye-level view, wet pavement with crisp neon reflections, bicycles, umbrellas, layered shop windows, fine electrical cables, realistic people at a distance. One dominant red sign must read exactly 'NIGHT MARKET' and a smaller blue sign must read exactly 'OPEN 24 HOURS'. No other legible text, no watermark.",
+      "seed": 2026081924,
+      "steps": 30
+    },
+    {
+      "files": [
+        {
+          "bytes": 753957,
+          "height": 544,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/woven-textile-macro/community.png",
+          "sha256": "be6dbd2f85ca54e6a2455a035686fdac06b5528179da198061b7304e1e9d8f87",
+          "width": 960
+        },
+        {
+          "bytes": 745570,
+          "height": 544,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/woven-textile-macro/ours.png",
+          "sha256": "6c75c3aefa8b870a6d4dd220f8b871f8a92646f8d036abd145e1b661b97f1083",
+          "width": 960
+        },
+        {
+          "bytes": 1482673,
+          "height": 596,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/woven-textile-macro/comparison.png",
+          "sha256": "6a58a07cf7ef40cb194c34095dff967119999c1586dab79f2cc9d7e2e670f173",
+          "width": 1920
+        }
+      ],
+      "id": "woven-textile-macro",
+      "prompt": "High-resolution macro product photograph of a folded navy and ivory technical textile with a tiny repeating herringbone weave, visible individual fibers, precise alternating thread pattern, one stitched orange seam, a brushed aluminum ruler for scale, sharp focus across the fabric plane, neutral studio light, no moire, no text, no watermark.",
+      "seed": 2026081925,
+      "steps": 30
+    },
+    {
+      "files": [
+        {
+          "bytes": 504724,
+          "height": 544,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/skeleton-watch/community.png",
+          "sha256": "695a9a64aef83717ca9437539985cd7d15b1248b1483faa219779f6ca54ccfc1",
+          "width": 960
+        },
+        {
+          "bytes": 431881,
+          "height": 544,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/skeleton-watch/ours.png",
+          "sha256": "69350994395dd491001c1b4bca4385dd4dced9e8f074d8b0d49fb17435746074",
+          "width": 960
+        },
+        {
+          "bytes": 900412,
+          "height": 596,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/skeleton-watch/comparison.png",
+          "sha256": "e04adb577ca209ff7966737c5ed7beffab54639171ece495c6c2967145019033",
+          "width": 1920
+        }
+      ],
+      "id": "skeleton-watch",
+      "prompt": "Extreme-detail studio photograph of a mechanical skeleton wristwatch, straight-on view, entire round watch visible, transparent dial showing coherent gears, jewels and springs, sixty evenly spaced minute tick marks, two slender hands set to 10:10, polished steel case, black leather strap, controlled reflections, dark gray background, no brand, no text, no watermark.",
+      "seed": 2026081926,
+      "steps": 30
+    },
+    {
+      "files": [
+        {
+          "bytes": 760621,
+          "height": 544,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/library-interior/community.png",
+          "sha256": "bf5fbb6c67a716b307602cd6b0d670a7f570a205ba09896a285134ca1f67bc30",
+          "width": 960
+        },
+        {
+          "bytes": 766652,
+          "height": 544,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/library-interior/ours.png",
+          "sha256": "0f6b9676e195f9bd6d8ba1c75846c0c97213ea48a9839dbec7c4c161cea3ced9",
+          "width": 960
+        },
+        {
+          "bytes": 1517633,
+          "height": 596,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/library-interior/comparison.png",
+          "sha256": "bb9ec988433f44c759c742183f7fc9d5977a15b71e4923e6ef737c833b0e4507",
+          "width": 1920
+        }
+      ],
+      "id": "library-interior",
+      "prompt": "Photorealistic contemporary library interior with strict two-point architectural perspective, full-height walnut shelves containing hundreds of distinct books, perforated brass stair rail, slatted ceiling, patterned terrazzo floor, reading tables and precise chair legs, warm indirect lighting, deep focus, straight verticals, no people, no text, no watermark.",
+      "seed": 2026081927,
+      "steps": 30
+    },
+    {
+      "files": [
+        {
+          "bytes": 405479,
+          "height": 544,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/dense-ui-dashboard/community.png",
+          "sha256": "c9fb0827c0ab24a4848ecfed62274a2776e99d9d25ba162949f3e349abaac803",
+          "width": 960
+        },
+        {
+          "bytes": 396060,
+          "height": 544,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/dense-ui-dashboard/ours.png",
+          "sha256": "5aa1adb8fce79d9e2f0a4739899887d24f2bda59522692f371c16610da77175f",
+          "width": 960
+        },
+        {
+          "bytes": 763265,
+          "height": 596,
+          "mode": "RGB",
+          "path": "examples/hard-v1/t2i/dense-ui-dashboard/comparison.png",
+          "sha256": "6b5e63dbe247918676e46554a326c9696a3cdc6f4212cc0c2f643e76c4cf458a",
+          "width": 1920
+        }
+      ],
+      "id": "dense-ui-dashboard",
+      "prompt": "High-fidelity dark desktop analytics dashboard shown perfectly front-on, 16:9 interface, crisp twelve-column grid, four line charts, one candlestick chart, a map, compact tables and aligned controls. Exact visible headings: 'SYSTEM HEALTH', 'LATENCY', 'THROUGHPUT', 'ERROR RATE'. Small values must remain sharp and logically aligned. No device frame, no perspective, no logo, no watermark.",
+      "seed": 2026081928,
+      "steps": 30
+    }
+  ]
+}

examples/hard-v1/visual-review.md

@@ -0,0 +1,55 @@
+# Fixed hard-example suite
+
+This suite contains eight text-to-image cases and two source-conditioned edit
+cases. Prompts, seeds, inference steps, the T2I latent index, and the five edit
+latent indices were fixed before generation. There was no seed search, prompt
+rewrite after viewing outputs, or best-frame selection. Failures are retained.
+
+The T2I comparisons decode the exact same frozen H3 latent slice with
+`Mamad8/MiniMax-H3-Image-VAE` and this 500K checkpoint. They isolate the
+decoder difference, not the quality of H3 denoising or prompt following.
+
+![Eight fixed T2I comparisons](t2i-overview.png)
+
+## T2I observations
+
+| Case | What works | What remains difficult |
+|---|---|---|
+| [Exact package text](t2i/scene-text-package/comparison.png) | Coherent package, folds, pedestal, and large title structure | Small required copy is incorrect; exact typography is not reliable |
+| [Wet feathers and anatomy](t2i/wet-kingfisher/comparison.png) | Strong silhouette, eye, beak, feather separation, droplets, and background depth | This is one generated case without ground truth; it does not overturn the broader photograph benchmark |
+| [Exploded technical diagram](t2i/exploded-camera-diagram/comparison.png) | Strong component separation, line work, camera geometry, and leader-line structure | Most requested labels are malformed or substituted |
+| [Rain, reflections, and signs](t2i/rainy-neon-street/comparison.png) | Coherent street perspective, bicycles, storefronts, rain, and reflections | Exact English sign text is absent or incorrect |
+| [Repeating textile weave](t2i/woven-textile-macro/comparison.png) | Fibers, folds, orange seam, and repeated weave remain sharp | The precise herringbone phase and ruler markings are not exact |
+| [Mechanical watch detail](t2i/skeleton-watch/comparison.png) | Strong circular geometry, case, hands, ticks, and mechanical detail | The movement is visually plausible rather than mechanically exact |
+| [Dense architectural interior](t2i/library-interior/comparison.png) | Strong shelves, ceiling rhythm, stairs, tables, chairs, and depth | Some small books and architectural junctions are simplified |
+| [Dense UI and small text](t2i/dense-ui-dashboard/comparison.png) | Crisp panel grid, chart structure, color coding, and control density | Headings and small values are mostly malformed |
+
+Across all eight cases, the 500K decoder removes the community decoder's
+dominant grid, echo, and temporal-smear artifacts. The remaining failures are
+mostly inherited semantic-generation limits or known decoder weaknesses in
+exact text and phase-sensitive detail.
+
+## Image-edit observations
+
+The edit workflow is H3 first-frame-conditioned video used as a static-edit
+experiment. The source and fixed latent slices `0`, `9`, `18`, `27`, and `36`
+are shown in order. These are not best-frame selections.
+
+![Two fixed image-edit transitions](edit-overview.png)
+
+| Case | What works | What remains difficult |
+|---|---|---|
+| [Audio-interface enclosure](edit/audio-interface-color/transition-grid.png) | The requested blue-to-orange enclosure change is clear by latent 18 and later slices retain the main product form | Fine labels change, controls drift, exposure rises, and exact preservation weakens as the edit strengthens |
+| [Reading-chair material](edit/reading-chair-material/transition-grid.png) | The chair becomes green with a visible material pattern while the broad room layout remains recognizable | Poster text, books, window detail, furniture, lighting, and local geometry drift across the transition |
+
+These edits demonstrate why this checkpoint is presented as a single-frame
+decoder rather than a native image-edit model. The decoder can render an H3
+edit slice sharply, but it cannot correct semantic drift already present in
+the generated latent sequence.
+
+## Review boundary
+
+This is an AI visual review of ten fixed cases, not a human preference study.
+The complete prompt and seed registration is in
+[`hard-example-prompts-v1.json`](../../hard-example-prompts-v1.json), and
+checksums are in [`public-manifest.json`](public-manifest.json).

hard-example-prompts-v1.json

@@ -0,0 +1,76 @@
+{
+  "schema_version": 1,
+  "selection_policy": {
+    "prompts_fixed_before_generation": true,
+    "seed_search": false,
+    "t2i_latent_index": 0,
+    "edit_latent_indices": [0, 9, 18, 27, 36],
+    "show_failures": true
+  },
+  "text_to_image": [
+    {
+      "id": "scene-text-package",
+      "seed": 2026081921,
+      "steps": 30,
+      "prompt": "Premium studio product photograph of a small matte cream coffee package standing upright on a dark stone pedestal, centered, full package visible, crisp folded edges and subtle paper fibers. The package must display the exact large title 'NORTH / 27', the exact subtitle 'HIGH ALTITUDE COFFEE', and three small lines 'ORIGIN: KENYA', 'ROAST: LIGHT', '250 g'. Soft directional light, realistic shadows, no people, no extra text, no logos, no watermark."
+    },
+    {
+      "id": "wet-kingfisher",
+      "seed": 2026081922,
+      "steps": 30,
+      "prompt": "Ultra-detailed natural wildlife photograph of a wet kingfisher perched on a mossy branch immediately after rain, full bird visible, individual blue and orange feather barbs, tiny water droplets, delicate claws, realistic eye reflection, soft forest bokeh, natural overcast light, accurate anatomy, no text, no watermark."
+    },
+    {
+      "id": "exploded-camera-diagram",
+      "seed": 2026081923,
+      "steps": 30,
+      "prompt": "Clean industrial-design exploded technical diagram of a compact mirrorless camera on white paper, precise black line work with pale gray shading, twelve aligned mechanical components, thin leader lines, dimension arrows, consistent perspective, and exact labels 'LENS', 'SENSOR', 'SHUTTER', 'BODY', 'BATTERY'. Entire diagram visible with generous margins, no hands, no watermark."
+    },
+    {
+      "id": "rainy-neon-street",
+      "seed": 2026081924,
+      "steps": 30,
+      "prompt": "Photorealistic rainy night street in a dense modern city, wide-angle eye-level view, wet pavement with crisp neon reflections, bicycles, umbrellas, layered shop windows, fine electrical cables, realistic people at a distance. One dominant red sign must read exactly 'NIGHT MARKET' and a smaller blue sign must read exactly 'OPEN 24 HOURS'. No other legible text, no watermark."
+    },
+    {
+      "id": "woven-textile-macro",
+      "seed": 2026081925,
+      "steps": 30,
+      "prompt": "High-resolution macro product photograph of a folded navy and ivory technical textile with a tiny repeating herringbone weave, visible individual fibers, precise alternating thread pattern, one stitched orange seam, a brushed aluminum ruler for scale, sharp focus across the fabric plane, neutral studio light, no moire, no text, no watermark."
+    },
+    {
+      "id": "skeleton-watch",
+      "seed": 2026081926,
+      "steps": 30,
+      "prompt": "Extreme-detail studio photograph of a mechanical skeleton wristwatch, straight-on view, entire round watch visible, transparent dial showing coherent gears, jewels and springs, sixty evenly spaced minute tick marks, two slender hands set to 10:10, polished steel case, black leather strap, controlled reflections, dark gray background, no brand, no text, no watermark."
+    },
+    {
+      "id": "library-interior",
+      "seed": 2026081927,
+      "steps": 30,
+      "prompt": "Photorealistic contemporary library interior with strict two-point architectural perspective, full-height walnut shelves containing hundreds of distinct books, perforated brass stair rail, slatted ceiling, patterned terrazzo floor, reading tables and precise chair legs, warm indirect lighting, deep focus, straight verticals, no people, no text, no watermark."
+    },
+    {
+      "id": "dense-ui-dashboard",
+      "seed": 2026081928,
+      "steps": 30,
+      "prompt": "High-fidelity dark desktop analytics dashboard shown perfectly front-on, 16:9 interface, crisp twelve-column grid, four line charts, one candlestick chart, a map, compact tables and aligned controls. Exact visible headings: 'SYSTEM HEALTH', 'LATENCY', 'THROUGHPUT', 'ERROR RATE'. Small values must remain sharp and logically aligned. No device frame, no perspective, no logo, no watermark."
+    }
+  ],
+  "image_edits": [
+    {
+      "id": "audio-interface-color",
+      "source": "source-fixtures-hard-v1/audio-interface.png",
+      "seed": 2026081931,
+      "steps": 25,
+      "instruction": "Change only the cobalt-blue metal enclosure to warm matte orange anodized aluminum. Preserve the black front panel, exact text including AERON 07, every knob and port, LED colors and states, ventilation holes, camera angle, crop, lighting, shadows, and background."
+    },
+    {
+      "id": "reading-chair-material",
+      "source": "source-fixtures-hard-v1/reading-room.png",
+      "seed": 2026081932,
+      "steps": 25,
+      "instruction": "Change only the mustard-yellow chair upholstery to deep emerald-green herringbone velvet. Preserve the chair shape, room geometry, poster text, lamp, plant, books, window, rain, rug grid, table, glass, lighting, camera angle, and crop."
+    }
+  ]
+}

source-fixtures-hard-v1/README.md

@@ -0,0 +1,26 @@
+# Hard edit-source provenance
+
+These two source images were generated specifically for this fixed public
+example suite using the built-in OpenAI image-generation workflow. They are
+not training images and were not selected from multiple generated variants.
+
+## `audio-interface.png`
+
+SHA-256:
+`108a9d4b1975d29a80d85f7c6097495677732aeae65ceb16a0e66342d5bccc3f`
+
+Prompt summary: a photorealistic cobalt-blue desktop audio interface with an
+exact `AERON 07` label, dense controls, LEDs, perforations, fine tick marks,
+and legible functional labels on a neutral studio background.
+
+## `reading-room.png`
+
+SHA-256:
+`eee847c1cc5e311d3c62405820b65a0d17ae599dc440eadc8d47eba236b4ab18`
+
+Prompt summary: a photorealistic reading room with a mustard boucle chair,
+rainy window, lamp, plant, books, fine rug grid, multiple preservation-sensitive
+objects, and a poster reading `READ / REST / REPEAT`.
+
+The subsequent H3 edit instructions and fixed seeds are recorded in
+[`hard-example-prompts-v1.json`](../hard-example-prompts-v1.json).