Publish validation-selected SD1.5 to SANA text bridge and evaluated alternatives

Three exact inference checkpoints, pinned configurations, validation selection, held-out quality measurements, reviewed comparison panels, and SHA256 checksums. No test-driven reselection.

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+figures/bench-comparison.jpg filter=lfs diff=lfs merge=lfs -text
+figures/first-four-test-prompts.jpg filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 255BITS
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -0,0 +1,147 @@
+---
+license: mit
+language:
+- en
+library_name: pytorch
+pipeline_tag: text-to-image
+tags:
+- model-glue
+- sana
+- stable-diffusion
+- model-stitching
+- safetensors
+base_model:
+- stable-diffusion-v1-5/stable-diffusion-v1-5
+- Efficient-Large-Model/Sana_600M_512px_diffusers
+inference: false
+---
+
+# SD1.5 CLIP → SANA text bridges
+
+Three trained bridges translate **77 × 768 SD1.5 CLIP text states** into
+**77 × 2304 SANA conditioning states and a mask**. The bridge runs as an ordinary
+deterministic network forward, followed by normal frozen SANA sampling. There is
+no fitting, retrieval, particle sampling, or candidate selection during inference.
+
+These are bridge weights for the [model-glue implementation](https://github.com/255BITS/model-glue),
+not a complete Diffusers pipeline. The donor text encoder, SANA transformer, and
+DC-AE decoder are loaded separately at pinned revisions. SANA's native empty
+conditioning is stored in each checkpoint, so deployment needs no Gemma encoder.
+
+## Choose a checkpoint
+
+| File | Role | Selected training update |
+|---|---|---:|
+| [model.safetensors](https://huggingface.co/ntc-ai/model-glue-sd15-sana-text/resolve/main/model.safetensors) | Default: validation-selected linear continuation with fixed mask calibration | 1250 |
+| [alternatives/linear-uncalibrated.safetensors](https://huggingface.co/ntc-ai/model-glue-sd15-sana-text/resolve/main/alternatives/linear-uncalibrated.safetensors) | Same trained head before calibration; retained for its mask tradeoff | 1250 |
+| [alternatives/expanded-source.safetensors](https://huggingface.co/ntc-ai/model-glue-sd15-sana-text/resolve/main/alternatives/expanded-source.safetensors) | Expanded output head with full-source access and fixed mask calibration | 1000 |
+
+The expanded-source architecture repairs a verified blind space in the old source
+projection, but did not win the combined validation ranking. The default therefore
+retains the linear head. All three checkpoints include the complete bridge, cloud,
+calibration buffers, architecture metadata, and pinned source/recipient identities.
+They load directly without the parent checkpoint or training caches.
+
+## Run
+
+Use the implementation revision that was tested for this release:
+
+```bash
+git clone https://github.com/255BITS/model-glue.git
+cd model-glue
+git checkout f67054d329fe5246510a094dd791d485526e3dd6
+python3 -m venv .venv
+.venv/bin/python -m pip install -e .
+.venv/bin/python - <<'PY'
+from huggingface_hub import snapshot_download
+snapshot_download(
+    repo_id="ntc-ai/model-glue-sd15-sana-text",
+    local_dir="artifacts/huggingface/sd15-sana-text",
+    allow_patterns=["model.safetensors", "manifest.json", "SHA256SUMS"],
+)
+PY
+CUDA_VISIBLE_DEVICES=1 .venv/bin/python -m model_glue.sana_infer \
+  --checkpoint artifacts/huggingface/sd15-sana-text/model.safetensors \
+  --prompt 'A park bench on a beach of white sand looking out to the ocean.' \
+  --seed 87654 --steps 20 --height 512 --width 512 --guidance 4.5 \
+  --out artifacts/sana-text-example
+```
+
+Choose an available CUDA device for your machine. The output directory must be new.
+The command saves `image.png` and `render.json`, including the checkpoint hash and
+deterministic-forward verification. To use an alternative, download its path in
+`allow_patterns` and pass that checkpoint to the same CLI. For a reproducible
+download, set `revision` to the desired Hugging Face commit hash.
+
+The tested deployment uses CLIP FP16 with a fixed batch padded to 16, a FP32
+bridge, SANA BF16, and the pinned 512-pixel sampler recipe. Input is limited to 77
+tokens. The source skip uses matching sequence lengths; it does not establish
+word-level alignment between tokenizers. `config.json` describes the default
+bridge, while the loader treats checkpoint metadata as authoritative.
+
+## Measured quality
+
+Selection used **58 validation prompts**, averaging seeds within each prompt.
+The final test was held out until selection was frozen: **16 prompts × 3 seeds**,
+20 sampling steps, 512 × 512, guidance 4.5. Every declared output was scored.
+
+| Method | Native-image LPIPS ↓ | Pixel MAE ↓ | CLIP cosine ↑ |
+|---|---:|---:|---:|
+| Original parent, comparison baseline | 0.38347 | 0.14984 | 0.30620 |
+| Uncalibrated linear continuation | 0.37711 | 0.14652 | 0.30744 |
+| Validation-selected default | 0.37802 | 0.14609 | 0.30757 |
+| Expanded-source alternative | 0.37778 | 0.14716 | 0.30815 |
+| Native SANA reference | 0.00000 | 0.00000 | 0.30820 |
+
+The default improves LPIPS by **1.42%** and pixel MAE by **2.50%** against the
+original parent. LPIPS improves on 11/16 prompt averages, but the paired difference
+of −0.00545 has a 95% prompt-bootstrap interval of **[−0.01769, +0.00817]**.
+The interval includes no improvement. These are modest observed fidelity gains,
+not established gains in general prompt accuracy or aesthetics. The test ranking
+did not replace the frozen validation choice.
+
+Mask calibration reduces validation errors from 6 to 4, but increases missed test
+positions from **1 to 3**; exact test masks decline from 15/16 to 14/16 prompts.
+The uncalibrated control has slightly better test LPIPS, while the calibrated
+default has better pixel MAE and CLIP. Composition and attribute errors remain.
+
+In this illustrative beach-bench case, the default's orientation is closer to
+native, while the expanded-source alternative changes the bench color. Columns
+are native SANA, original parent, selected default, and expanded-source alternative.
+
+![Bench comparison](https://huggingface.co/ntc-ai/model-glue-sd15-sana-text/resolve/main/figures/bench-comparison.jpg)
+
+The first four test prompts include failures: the small car between trucks is
+not recovered, and the swan remains goose-like. These panels show seed 87654;
+all three seeds contribute to the table. Both displayed panels were visually
+reviewed; neither contains people.
+
+![First four test prompts and failures](https://huggingface.co/ntc-ai/model-glue-sd15-sana-text/resolve/main/figures/first-four-test-prompts.jpg)
+
+## Training and provenance
+
+Seven continuations completed **8,500 updates on GPU 1**, starting from the same
+trained parent. The original stem, cloud, and masks stayed fixed during training;
+linear runs trained the output projection, and source-head runs trained the added
+branch. Mask calibration later folded one validation-fitted scalar into the mask
+bias. All new cloud controls are fixed; this study establishes no particle-movement
+advantage. SANA and its decoder were frozen throughout.
+
+Training used 8,192 paired captions, paired embedding supervision, and frozen-SANA
+functional distillation. Full-trajectory and repaired b_cap arms were also tested
+but did not win selection. These released winners were trained with the supervised
+distillation recipe. There is no new Anima result in this release.
+
+- [Complete study and reproduction details](https://github.com/255BITS/model-glue/blob/f67054d329fe5246510a094dd791d485526e3dd6/docs/results/formulation-retraining-20260905.md)
+- [Release manifest and per-file hashes](https://huggingface.co/ntc-ai/model-glue-sd15-sana-text/blob/main/manifest.json)
+- [Checksums](https://huggingface.co/ntc-ai/model-glue-sd15-sana-text/blob/main/SHA256SUMS)
+- [Training configurations](https://huggingface.co/ntc-ai/model-glue-sd15-sana-text/tree/main/configs)
+- [Selection and measured results](https://huggingface.co/ntc-ai/model-glue-sd15-sana-text/tree/main/reports)
+
+The implementation passed 82 tests, including checkpoint migration, resume,
+source reachability, mask geometry, and differentiable rollout checks. The
+Gaussian reference audit has zero cap-parameter-gradient discrepancy. Exact
+upstream revisions and checkpoint SHA256 values are in the manifest. The release
+contains learned bridge weights under the repository's MIT license; the separate
+upstream models retain their own licenses. No donor or recipient model weights,
+training datasets, optimizer states, or credentials are included.

SHA256SUMS

@@ -0,0 +1,18 @@
+f50d7a3340d8e33d65e6905d6471ee64c2217a0a20f54a67922f96e02ec6a42b  LICENSE
+8786e5eaf37799205bbc1f25cdd4cf4d965a9ad208e544e3a53d2656f7ae62a6  README.md
+1a2db74c6c2d940ee15a845a7e60404ffbe32573757672586a18c578db36a10a  alternatives/expanded-source.safetensors
+fc7f44369cf21791c4f9ad6b74978dd420be41a76330a1fe32b11f440f3bf925  alternatives/linear-uncalibrated.safetensors
+aea09893f34769c9b4eef9c30ab9711c935854ec3cf29571aee81cc2367109d7  config.json
+7d352ba100c91c6167a3926275d6cbb9ea784de7912d83919202d1de55e621f4  configs/linear-distill-fixed.json
+aea09893f34769c9b4eef9c30ab9711c935854ec3cf29571aee81cc2367109d7  configs/linear_distill.architecture.json
+ceca1ec0fcac308fc6a565452f3dc67ce90dcd21e323acefd6df21cc9c1b78b4  configs/source-residual-fullscale-fixed.json
+726250e6d98abe97c35e61b066ac9211c8ef1759504020c0dd11205105d5880c  configs/source_cal.architecture.json
+4dc2536374ff7b33c85e7bf21e55977737d4e3cd7d0a6db9189812972dbdf469  configs/validation-confirmation.json
+12de8ae1a69eff31879911c50068af1a2fe77bd5f56dbbf5cf1e6743464eca06  figures/bench-comparison.jpg
+d3f12c915d977e11bbffc83b6e1ba94fabe5feafe2a3a35b0ee6c84e34e77654  figures/first-four-test-prompts.jpg
+635da76e075a43329ac4700a2f3a0e5cb71434be04f0828de29fc0a5b69acbcb  manifest.json
+d3be2492e49f6eb1bc28a406d35108f1055e3dd19e412792442b4c5d9274bb4d  model.safetensors
+5416eac9605b2e0af6b5325be168dbd86bdb8a7d5f241f5481dee417e439c349  reports/mask-metrics.json
+fabd67b38e5a5e878f63bbb0f6de0278c4bcc459d0603957ce3603417f1fddf2  reports/retraining.json
+e57fe58d25d63912d2bbf3bbc814301f760b841c1f49babc648a9d05396910c3  reports/selection.json
+4014fe544e8087ad654a596d7010212b3be99880d609ff22c6ad0fd8d637b36f  reports/test-results.json

alternatives/expanded-source.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a2db74c6c2d940ee15a845a7e60404ffbe32573757672586a18c578db36a10a
+size 62037769

alternatives/linear-uncalibrated.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc7f44369cf21791c4f9ad6b74978dd420be41a76330a1fe32b11f440f3bf925
+size 30167265

config.json

@@ -0,0 +1,17 @@
+{
+  "source_dim": 768,
+  "target_dim": 2304,
+  "tokens": 77,
+  "width": 384,
+  "blocks": 4,
+  "particle_dim": 4,
+  "num_particles": 256,
+  "learnable_particles": false,
+  "source_residual": true,
+  "native_unconditional": true,
+  "output_head": "linear",
+  "head_width": 0,
+  "use_particles": true,
+  "output_source_residual": false,
+  "output_source_init_scale": 0.01
+}

configs/linear-distill-fixed.json

@@ -0,0 +1,75 @@
+{
+  "seed": 90526,
+  "source": {
+    "repo_id": "stable-diffusion-v1-5/stable-diffusion-v1-5",
+    "revision": "451f4fe16113bff5a5d2269ed5ad43b0592e9a14"
+  },
+  "dataset": "artifacts/data/sana-transfer-sd15",
+  "state_cache": "artifacts/data/sana-transfer-sd15-states",
+  "steps": 1500,
+  "batch_size": 32,
+  "lr": 1e-05,
+  "adv_weight": 0.0,
+  "feature_weight": 0.0,
+  "embedding_weight": 0.1,
+  "mask_weight": 0.0,
+  "particle_weight": 0.0,
+  "functional_weight": 1.0,
+  "functional_every": 1,
+  "functional_start": 0,
+  "functional_batch": 4,
+  "eval_every": 250,
+  "log_every": 50,
+  "functional_normalize": true,
+  "native_state_probability": 0.5,
+  "ema_decay": 0.98,
+  "branch_weight": 1.0,
+  "guided_weight": 0.05,
+  "guidance_scale": 4.5,
+  "matmul_tf32": false,
+  "max_grad_norm": 1.0,
+  "selection_metric": "native_lpips",
+  "bridge": {
+    "source_dim": 768,
+    "width": 384,
+    "blocks": 4,
+    "source_residual": true,
+    "native_unconditional": true,
+    "learnable_particles": false,
+    "output_head": "linear"
+  },
+  "functional_validation": {
+    "batch_size": 4
+  },
+  "render_validation": {
+    "seeds": [
+      87654
+    ],
+    "steps": 20,
+    "height": 512,
+    "width": 512,
+    "guidance_scale": 4.5,
+    "scorer_device": "cpu",
+    "include_native_empty": false,
+    "measure_native_fidelity": true
+  },
+  "scope": "Matched old-linear-head distillation control for residual-distill-fixed. Same parent, frozen resampler/cloud, batches, functional states, losses, learning rate, and update budget. Mask weight zero and deployed hard masks yield zero mask-head gradient. Validation selection only.",
+  "patience": 0,
+  "critic": {
+    "architecture": "coordinate",
+    "width": 128
+  },
+  "training_paths": "routed",
+  "regularizer": "b_cap",
+  "cap_coefficient": 1.0,
+  "cap_kappa": 1.0,
+  "cap_damping": 0.01,
+  "cap_interpolation_coefficient": 0.1,
+  "feature_normalization": "centered",
+  "padding_weight": 0.0,
+  "particle_center_weight": 0.0,
+  "routing_diagnostics": true,
+  "trainable_parts": "heads",
+  "supervised_only": true,
+  "soft_mask_weight": 0.0
+}

configs/linear_distill.architecture.json

@@ -0,0 +1,17 @@
+{
+  "source_dim": 768,
+  "target_dim": 2304,
+  "tokens": 77,
+  "width": 384,
+  "blocks": 4,
+  "particle_dim": 4,
+  "num_particles": 256,
+  "learnable_particles": false,
+  "source_residual": true,
+  "native_unconditional": true,
+  "output_head": "linear",
+  "head_width": 0,
+  "use_particles": true,
+  "output_source_residual": false,
+  "output_source_init_scale": 0.01
+}

configs/source-residual-fullscale-fixed.json

@@ -0,0 +1,77 @@
+{
+  "seed": 90526,
+  "source": {
+    "repo_id": "stable-diffusion-v1-5/stable-diffusion-v1-5",
+    "revision": "451f4fe16113bff5a5d2269ed5ad43b0592e9a14"
+  },
+  "dataset": "artifacts/data/sana-transfer-sd15",
+  "state_cache": "artifacts/data/sana-transfer-sd15-states",
+  "steps": 1500,
+  "batch_size": 32,
+  "lr": 1e-05,
+  "adv_weight": 0.0,
+  "feature_weight": 0.0,
+  "embedding_weight": 0.1,
+  "mask_weight": 0.0,
+  "particle_weight": 0.0,
+  "functional_weight": 1.0,
+  "functional_every": 1,
+  "functional_start": 0,
+  "functional_batch": 4,
+  "eval_every": 250,
+  "log_every": 50,
+  "functional_normalize": true,
+  "native_state_probability": 0.5,
+  "ema_decay": 0.98,
+  "branch_weight": 1.0,
+  "guided_weight": 0.05,
+  "guidance_scale": 4.5,
+  "matmul_tf32": false,
+  "max_grad_norm": 1.0,
+  "selection_metric": "native_lpips",
+  "bridge": {
+    "source_dim": 768,
+    "width": 384,
+    "blocks": 4,
+    "source_residual": true,
+    "native_unconditional": true,
+    "learnable_particles": false,
+    "output_head": "expanded_residual",
+    "output_source_residual": true,
+    "output_source_init_scale": 1.0
+  },
+  "functional_validation": {
+    "batch_size": 4
+  },
+  "render_validation": {
+    "seeds": [
+      87654
+    ],
+    "steps": 20,
+    "height": 512,
+    "width": 512,
+    "guidance_scale": 4.5,
+    "scorer_device": "cpu",
+    "include_native_empty": false,
+    "measure_native_fidelity": true
+  },
+  "scope": "Matched source-access initialization control: identical to source-residual-distill-fixed except source projection uses its full fan-in initialization rather than 0.01 scaling. The final output residual still starts exactly at zero, preserving parent predictions. Same training/validation schedule and fixed cloud.",
+  "patience": 0,
+  "critic": {
+    "architecture": "coordinate",
+    "width": 128
+  },
+  "training_paths": "routed",
+  "regularizer": "b_cap",
+  "cap_coefficient": 1.0,
+  "cap_kappa": 1.0,
+  "cap_damping": 0.01,
+  "cap_interpolation_coefficient": 0.1,
+  "feature_normalization": "centered",
+  "padding_weight": 0.0,
+  "particle_center_weight": 0.0,
+  "routing_diagnostics": true,
+  "trainable_parts": "residual",
+  "supervised_only": true,
+  "soft_mask_weight": 0.0
+}

configs/source_cal.architecture.json

@@ -0,0 +1,17 @@
+{
+  "source_dim": 768,
+  "target_dim": 2304,
+  "tokens": 77,
+  "width": 384,
+  "blocks": 4,
+  "particle_dim": 4,
+  "num_particles": 256,
+  "learnable_particles": false,
+  "source_residual": true,
+  "native_unconditional": true,
+  "output_head": "expanded_residual",
+  "head_width": 0,
+  "use_particles": true,
+  "output_source_residual": true,
+  "output_source_init_scale": 1.0
+}

configs/validation-confirmation.json

@@ -0,0 +1,54 @@
+{
+  "dataset_parent": "artifacts/data/text-integrated",
+  "source_split": "val",
+  "val_indices": [
+    4,
+    5,
+    8,
+    9,
+    16,
+    18,
+    20,
+    23,
+    27,
+    28,
+    29,
+    30,
+    31,
+    32,
+    36,
+    42,
+    43,
+    50,
+    51,
+    52,
+    57,
+    62,
+    63,
+    67,
+    69,
+    70,
+    76,
+    78,
+    81,
+    82,
+    86,
+    87,
+    88,
+    90,
+    91,
+    100,
+    101,
+    103,
+    106,
+    107,
+    109,
+    116,
+    121,
+    123,
+    124,
+    125
+  ],
+  "purpose": "Additional validation confirmation outside the twelve checkpoint-selection prompts; predeclared objects, animals, buildings and vehicles. No test tensors and no training.",
+  "source_manifest_sha256": "2bf44e5af1d94ccb054b242669a89f4ce56b36973dbd86c9d6dd8706d3f0ae2e"
+}

manifest.json

@@ -0,0 +1,222 @@
+{
+  "repo_id": "ntc-ai/model-glue-sd15-sana-text",
+  "code_repository": "https://github.com/255BITS/model-glue",
+  "code_revision": "f67054d329fe5246510a094dd791d485526e3dd6",
+  "default_checkpoint": "model.safetensors",
+  "source": {
+    "repo_id": "stable-diffusion-v1-5/stable-diffusion-v1-5",
+    "revision": "451f4fe16113bff5a5d2269ed5ad43b0592e9a14"
+  },
+  "recipient": {
+    "repo_id": "Efficient-Large-Model/Sana_600M_512px_diffusers",
+    "revision": "83d7a190bfd1fd070570a793d2dab5c7a3231b9d"
+  },
+  "decoder": {
+    "repo_id": "mit-han-lab/dc-ae-f32c32-sana-1.0-diffusers",
+    "revision": "ca69e17e97609e64ce055115a6515215109b1f50"
+  },
+  "selection_frozen_before_test": true,
+  "checkpoints": {
+    "model.safetensors": {
+      "role": "validation-selected default",
+      "selection_key": "linear_cal",
+      "sha256": "d3be2492e49f6eb1bc28a406d35108f1055e3dd19e412792442b4c5d9274bb4d",
+      "size_bytes": 30168265,
+      "training_step": 1250,
+      "architecture": {
+        "source_dim": 768,
+        "target_dim": 2304,
+        "tokens": 77,
+        "width": 384,
+        "blocks": 4,
+        "particle_dim": 4,
+        "num_particles": 256,
+        "learnable_particles": false,
+        "source_residual": true,
+        "native_unconditional": true,
+        "output_head": "linear",
+        "head_width": 0,
+        "use_particles": true,
+        "output_source_residual": false,
+        "output_source_init_scale": 0.01
+      },
+      "mask_calibration": {
+        "parent": "artifacts/runs/formulation-linear-distill-fixed/best.safetensors",
+        "parent_sha256": "fc7f44369cf21791c4f9ad6b74978dd420be41a76330a1fe32b11f440f3bf925",
+        "datasets": [
+          {
+            "path": "artifacts/data/sana-transfer-sd15",
+            "split": "val",
+            "rows": 12,
+            "manifest_sha256": "208e35938dc8a58cae55143b826f6227b7d1e3e1f2d5a76f1945b680afdc4789"
+          },
+          {
+            "path": "artifacts/data/formulation-validation-confirmation",
+            "split": "val",
+            "rows": 46,
+            "manifest_sha256": "a9923ec7b7b8bc9efcfd73d277a437b4f70ba3b67f7036532088c2938352f0bd"
+          }
+        ],
+        "threshold_offset": 2.0417721271514893,
+        "errors_before": 6,
+        "errors_after": 4,
+        "actual_network_errors_verified": true,
+        "positions": 4408,
+        "selection": "Minimum validation mask Hamming error; zero wins ties, otherwise closest-to-zero optimal interval midpoint. BOS excluded and always valid.",
+        "inference": "Fixed exported mask-head bias. No fitting, selection, target tokenizer or retrieval during inference."
+      }
+    },
+    "alternatives/linear-uncalibrated.safetensors": {
+      "role": "uncalibrated control",
+      "selection_key": "linear_distill",
+      "sha256": "fc7f44369cf21791c4f9ad6b74978dd420be41a76330a1fe32b11f440f3bf925",
+      "size_bytes": 30167265,
+      "training_step": 1250,
+      "architecture": {
+        "source_dim": 768,
+        "target_dim": 2304,
+        "tokens": 77,
+        "width": 384,
+        "blocks": 4,
+        "particle_dim": 4,
+        "num_particles": 256,
+        "learnable_particles": false,
+        "source_residual": true,
+        "native_unconditional": true,
+        "output_head": "linear",
+        "head_width": 0,
+        "use_particles": true,
+        "output_source_residual": false,
+        "output_source_init_scale": 0.01
+      },
+      "mask_calibration": null
+    },
+    "alternatives/expanded-source.safetensors": {
+      "role": "expanded-source alternative",
+      "selection_key": "source_cal",
+      "sha256": "1a2db74c6c2d940ee15a845a7e60404ffbe32573757672586a18c578db36a10a",
+      "size_bytes": 62037769,
+      "training_step": 1000,
+      "architecture": {
+        "source_dim": 768,
+        "target_dim": 2304,
+        "tokens": 77,
+        "width": 384,
+        "blocks": 4,
+        "particle_dim": 4,
+        "num_particles": 256,
+        "learnable_particles": false,
+        "source_residual": true,
+        "native_unconditional": true,
+        "output_head": "expanded_residual",
+        "head_width": 0,
+        "use_particles": true,
+        "output_source_residual": true,
+        "output_source_init_scale": 1.0
+      },
+      "mask_calibration": {
+        "parent": "artifacts/runs/formulation-source-residual-fullscale-fixed/best.safetensors",
+        "parent_sha256": "ceeb067e4c96b545f5bc094d660ce1e0e1a1953dd14edbf3270a32017f1acdd9",
+        "datasets": [
+          {
+            "path": "artifacts/data/sana-transfer-sd15",
+            "split": "val",
+            "rows": 12,
+            "manifest_sha256": "208e35938dc8a58cae55143b826f6227b7d1e3e1f2d5a76f1945b680afdc4789"
+          },
+          {
+            "path": "artifacts/data/formulation-validation-confirmation",
+            "split": "val",
+            "rows": 46,
+            "manifest_sha256": "a9923ec7b7b8bc9efcfd73d277a437b4f70ba3b67f7036532088c2938352f0bd"
+          }
+        ],
+        "threshold_offset": 2.0417721271514893,
+        "errors_before": 6,
+        "errors_after": 4,
+        "actual_network_errors_verified": true,
+        "positions": 4408,
+        "selection": "Minimum validation mask Hamming error; zero wins ties, otherwise closest-to-zero optimal interval midpoint. BOS excluded and always valid.",
+        "inference": "Fixed exported mask-head bias. No fitting, selection, target tokenizer or retrieval during inference."
+      }
+    }
+  },
+  "files": {
+    "LICENSE": {
+      "sha256": "f50d7a3340d8e33d65e6905d6471ee64c2217a0a20f54a67922f96e02ec6a42b",
+      "size_bytes": 1064
+    },
+    "README.md": {
+      "sha256": "8786e5eaf37799205bbc1f25cdd4cf4d965a9ad208e544e3a53d2656f7ae62a6",
+      "size_bytes": 7968
+    },
+    "alternatives/expanded-source.safetensors": {
+      "sha256": "1a2db74c6c2d940ee15a845a7e60404ffbe32573757672586a18c578db36a10a",
+      "size_bytes": 62037769
+    },
+    "alternatives/linear-uncalibrated.safetensors": {
+      "sha256": "fc7f44369cf21791c4f9ad6b74978dd420be41a76330a1fe32b11f440f3bf925",
+      "size_bytes": 30167265
+    },
+    "config.json": {
+      "sha256": "aea09893f34769c9b4eef9c30ab9711c935854ec3cf29571aee81cc2367109d7",
+      "size_bytes": 371
+    },
+    "configs/linear-distill-fixed.json": {
+      "sha256": "7d352ba100c91c6167a3926275d6cbb9ea784de7912d83919202d1de55e621f4",
+      "size_bytes": 2092
+    },
+    "configs/linear_distill.architecture.json": {
+      "sha256": "aea09893f34769c9b4eef9c30ab9711c935854ec3cf29571aee81cc2367109d7",
+      "size_bytes": 371
+    },
+    "configs/source-residual-fullscale-fixed.json": {
+      "sha256": "ceca1ec0fcac308fc6a565452f3dc67ce90dcd21e323acefd6df21cc9c1b78b4",
+      "size_bytes": 2214
+    },
+    "configs/source_cal.architecture.json": {
+      "sha256": "726250e6d98abe97c35e61b066ac9211c8ef1759504020c0dd11205105d5880c",
+      "size_bytes": 380
+    },
+    "configs/validation-confirmation.json": {
+      "sha256": "4dc2536374ff7b33c85e7bf21e55977737d4e3cd7d0a6db9189812972dbdf469",
+      "size_bytes": 763
+    },
+    "figures/bench-comparison.jpg": {
+      "sha256": "12de8ae1a69eff31879911c50068af1a2fe77bd5f56dbbf5cf1e6743464eca06",
+      "size_bytes": 120617
+    },
+    "figures/first-four-test-prompts.jpg": {
+      "sha256": "d3f12c915d977e11bbffc83b6e1ba94fabe5feafe2a3a35b0ee6c84e34e77654",
+      "size_bytes": 513851
+    },
+    "model.safetensors": {
+      "sha256": "d3be2492e49f6eb1bc28a406d35108f1055e3dd19e412792442b4c5d9274bb4d",
+      "size_bytes": 30168265
+    },
+    "reports/mask-metrics.json": {
+      "sha256": "5416eac9605b2e0af6b5325be168dbd86bdb8a7d5f241f5481dee417e439c349",
+      "size_bytes": 717
+    },
+    "reports/retraining.json": {
+      "sha256": "fabd67b38e5a5e878f63bbb0f6de0278c4bcc459d0603957ce3603417f1fddf2",
+      "size_bytes": 175016
+    },
+    "reports/selection.json": {
+      "sha256": "e57fe58d25d63912d2bbf3bbc814301f760b841c1f49babc648a9d05396910c3",
+      "size_bytes": 5669
+    },
+    "reports/test-results.json": {
+      "sha256": "4014fe544e8087ad654a596d7010212b3be99880d609ff22c6ad0fd8d637b36f",
+      "size_bytes": 6071
+    }
+  },
+  "visual_review": {
+    "date": "2026-09-05",
+    "finding": "Both published panels visually reviewed; neither contains people.",
+    "sha256": {
+      "figures/bench-comparison.jpg": "12de8ae1a69eff31879911c50068af1a2fe77bd5f56dbbf5cf1e6743464eca06",
+      "figures/first-four-test-prompts.jpg": "d3f12c915d977e11bbffc83b6e1ba94fabe5feafe2a3a35b0ee6c84e34e77654"
+    }
+  }
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3be2492e49f6eb1bc28a406d35108f1055e3dd19e412792442b4c5d9274bb4d
+size 30168265

reports/mask-metrics.json

@@ -0,0 +1,37 @@
+{
+  "native": {
+    "wrong_positions": 0,
+    "false_positive": 0,
+    "false_negative": 0,
+    "exact_prompts": 16,
+    "prompts": 16
+  },
+  "parent": {
+    "wrong_positions": 1,
+    "false_positive": 0,
+    "false_negative": 1,
+    "exact_prompts": 15,
+    "prompts": 16
+  },
+  "linear_uncalibrated": {
+    "wrong_positions": 1,
+    "false_positive": 0,
+    "false_negative": 1,
+    "exact_prompts": 15,
+    "prompts": 16
+  },
+  "selected": {
+    "wrong_positions": 3,
+    "false_positive": 0,
+    "false_negative": 3,
+    "exact_prompts": 14,
+    "prompts": 16
+  },
+  "repaired_source": {
+    "wrong_positions": 3,
+    "false_positive": 0,
+    "false_negative": 3,
+    "exact_prompts": 14,
+    "prompts": 16
+  }
+}

reports/selection.json

@@ -0,0 +1,150 @@
+{
+  "selected": "linear_cal",
+  "selected_validation": {
+    "native_lpips": 0.37923671806166914,
+    "native_pixel_mae": 0.14399123039824524,
+    "clip_text_image_cosine": 0.30746305802430224
+  },
+  "parent_validation": {
+    "native_lpips": 0.3845586983979434,
+    "native_pixel_mae": 0.14540668043853908,
+    "clip_text_image_cosine": 0.30728059424751103
+  },
+  "validation_prompts": 58,
+  "validation_prompt_seed_pairs": 82,
+  "ranking": [
+    "linear_cal",
+    "linear_distill",
+    "source_cal",
+    "parent_cal",
+    "source_full",
+    "parent",
+    "combined_025",
+    "linear_rollout",
+    "orthogonal_100",
+    "orthogonal_025"
+  ],
+  "scores": {
+    "native": {
+      "native_lpips": 0.0,
+      "native_pixel_mae": 0.0,
+      "clip_text_image_cosine": 0.31729713510507823
+    },
+    "parent": {
+      "native_lpips": 0.3845586983979434,
+      "native_pixel_mae": 0.14540668043853908,
+      "clip_text_image_cosine": 0.30728059424751103
+    },
+    "linear_distill": {
+      "native_lpips": 0.3811477638952348,
+      "native_pixel_mae": 0.14415318966340063,
+      "clip_text_image_cosine": 0.3076438892675542
+    },
+    "linear_rollout": {
+      "native_lpips": 0.38574381789256784,
+      "native_pixel_mae": 0.14572909688470007,
+      "clip_text_image_cosine": 0.3078894349692881
+    },
+    "source_full": {
+      "native_lpips": 0.3832593902774241,
+      "native_pixel_mae": 0.14520142021878013,
+      "clip_text_image_cosine": 0.30850052174137926
+    },
+    "orthogonal_025": {
+      "native_lpips": 0.3886523847949916,
+      "native_pixel_mae": 0.1462558062597253,
+      "clip_text_image_cosine": 0.3085590165922011
+    },
+    "orthogonal_100": {
+      "native_lpips": 0.38771238516288237,
+      "native_pixel_mae": 0.1463224577407042,
+      "clip_text_image_cosine": 0.30837413865601876
+    },
+    "combined_025": {
+      "native_lpips": 0.3855607927530661,
+      "native_pixel_mae": 0.14555561092906985,
+      "clip_text_image_cosine": 0.30844964797811947
+    },
+    "parent_cal": {
+      "native_lpips": 0.38301196291871453,
+      "native_pixel_mae": 0.14532601702744247,
+      "clip_text_image_cosine": 0.30715905900659235
+    },
+    "linear_cal": {
+      "native_lpips": 0.37923671806166914,
+      "native_pixel_mae": 0.14399123039824524,
+      "clip_text_image_cosine": 0.30746305802430224
+    },
+    "source_cal": {
+      "native_lpips": 0.3814662068054594,
+      "native_pixel_mae": 0.14504165844670658,
+      "clip_text_image_cosine": 0.3084425563956129
+    }
+  },
+  "checkpoints": {
+    "parent": {
+      "path": "/ml2/model-glue/artifacts/runs/sana-transfer-sd15-movable/best.safetensors",
+      "sha256": "73fbd8ff783c8b6570f01d69551193c814c6baaa7233c7b058bf47fb4782cad9"
+    },
+    "linear_distill": {
+      "path": "/ml2/model-glue/artifacts/runs/formulation-linear-distill-fixed/best.safetensors",
+      "sha256": "fc7f44369cf21791c4f9ad6b74978dd420be41a76330a1fe32b11f440f3bf925"
+    },
+    "linear_rollout": {
+      "path": "/ml2/model-glue/artifacts/runs/formulation-linear-rollout-fixed/best.safetensors",
+      "sha256": "065d88db8ad73f904b3681f2a45bcacafd6b86af7b32f69aabab08a2585ef2f5"
+    },
+    "source_full": {
+      "path": "/ml2/model-glue/artifacts/runs/formulation-source-residual-fullscale-fixed/best.safetensors",
+      "sha256": "ceeb067e4c96b545f5bc094d660ce1e0e1a1953dd14edbf3270a32017f1acdd9"
+    },
+    "orthogonal_025": {
+      "path": "/ml2/model-glue/artifacts/formulation-repair/candidates/orthogonal-025.safetensors",
+      "sha256": "34d9ad66dff950cca7afec85bbac1046e05ff759e4eee9057d41d0416415fc7a"
+    },
+    "orthogonal_100": {
+      "path": "/ml2/model-glue/artifacts/formulation-repair/candidates/orthogonal-100.safetensors",
+      "sha256": "85986a987e6d5a6b5ad52f651b2267c319107927431afd685437dc829f11ba6a"
+    },
+    "combined_025": {
+      "path": "/ml2/model-glue/artifacts/formulation-repair/candidates/combined-025.safetensors",
+      "sha256": "73f11ee497640c4ce4cb09b115e940e84233e4a36130e73bb30cfa15c61a3124"
+    },
+    "parent_cal": {
+      "path": "/ml2/model-glue/artifacts/formulation-repair/parent-calibrated.safetensors",
+      "sha256": "ee3fe905e68d9fa9859ef5cd303ad86868b3ae680bd81c4e8f4954f2ba6d13ad"
+    },
+    "linear_cal": {
+      "path": "/ml2/model-glue/artifacts/formulation-repair/candidates/linear-distill-calibrated.safetensors",
+      "sha256": "d3be2492e49f6eb1bc28a406d35108f1055e3dd19e412792442b4c5d9274bb4d"
+    },
+    "source_cal": {
+      "path": "/ml2/model-glue/artifacts/formulation-repair/candidates/source-full-calibrated.safetensors",
+      "sha256": "1a2db74c6c2d940ee15a845a7e60404ffbe32573757672586a18c578db36a10a"
+    }
+  },
+  "selected_checkpoint": "/ml2/model-glue/artifacts/runs/formulation-selected/best.safetensors",
+  "selected_sha256": "d3be2492e49f6eb1bc28a406d35108f1055e3dd19e412792442b4c5d9274bb4d",
+  "selection": "Minimum native-image LPIPS averaged equally across 58 validation prompts; three seeds averaged within each original prompt and one seed for each additional prompt. Mask calibration used validation labels only.",
+  "test_used_for_selection": false,
+  "test_recipe": {
+    "split": "test",
+    "prompts": 16,
+    "seeds": [
+      87654,
+      12001,
+      12002
+    ],
+    "steps": 20,
+    "height": 512,
+    "width": 512,
+    "guidance_scale": 4.5
+  },
+  "final_test_comparators": [
+    "original parent",
+    "uncalibrated linear continuation",
+    "validation-selected calibrated linear continuation",
+    "calibrated expanded source-head alternative"
+  ],
+  "calibration_additional_panel_reuse": "Verified bitwise equality of every deployed conditioning tensor, positive mask and native empty branch for all 46 additional prompts."
+}

reports/test-results.json

@@ -0,0 +1,154 @@
+{
+  "metrics": {
+    "native": {
+      "clip_text_image_cosine": 0.3082017917186022,
+      "sampling_seconds": 0.6179222891029591,
+      "bridge_seconds": 0.0,
+      "image_rgb_std": 0.2777115295951565,
+      "image_count": 48,
+      "quality_pass": true,
+      "bridge_repeat_bitwise_equal": null,
+      "native_lpips": 0.0,
+      "native_pixel_mae": 0.0,
+      "native_detail_mae": 0.0,
+      "detail_energy_ratio": 1.0
+    },
+    "parent": {
+      "clip_text_image_cosine": 0.30619997655351955,
+      "sampling_seconds": 0.5849423649196979,
+      "bridge_seconds": 0.0062157050997484475,
+      "image_rgb_std": 0.26411043759435415,
+      "image_count": 48,
+      "quality_pass": true,
+      "bridge_repeat_bitwise_equal": true,
+      "native_lpips": 0.3834717807670434,
+      "native_pixel_mae": 0.14984353189356625,
+      "native_detail_mae": 0.03409249125979841,
+      "detail_energy_ratio": 1.062919204433759
+    },
+    "linear_uncalibrated": {
+      "clip_text_image_cosine": 0.30744258500635624,
+      "sampling_seconds": 0.5833559354844814,
+      "bridge_seconds": 0.0016241110861301422,
+      "image_rgb_std": 0.26165843568742275,
+      "image_count": 48,
+      "quality_pass": true,
+      "bridge_repeat_bitwise_equal": true,
+      "native_lpips": 0.3771136791134874,
+      "native_pixel_mae": 0.1465170100952188,
+      "native_detail_mae": 0.03367988064807529,
+      "detail_energy_ratio": 1.0336223902801673
+    },
+    "selected": {
+      "clip_text_image_cosine": 0.3075694603224595,
+      "sampling_seconds": 0.5856305301519266,
+      "bridge_seconds": 0.0017263079935219139,
+      "image_rgb_std": 0.26142118747035664,
+      "image_count": 48,
+      "quality_pass": true,
+      "bridge_repeat_bitwise_equal": true,
+      "native_lpips": 0.3780206674709916,
+      "native_pixel_mae": 0.14609215129166842,
+      "native_detail_mae": 0.03372150473296642,
+      "detail_energy_ratio": 1.036277251938979
+    },
+    "repaired_source": {
+      "clip_text_image_cosine": 0.3081482208023469,
+      "sampling_seconds": 0.5845581179504128,
+      "bridge_seconds": 0.0018588434904813766,
+      "image_rgb_std": 0.2612251853570342,
+      "image_count": 48,
+      "quality_pass": true,
+      "bridge_repeat_bitwise_equal": true,
+      "native_lpips": 0.37778033611054224,
+      "native_pixel_mae": 0.14715851237997413,
+      "native_detail_mae": 0.033508921546551086,
+      "detail_energy_ratio": 1.0336602094272773
+    }
+  },
+  "selected_paired_metrics": {
+    "native_lpips": {
+      "candidate_minus_baseline": -0.00545111329605182,
+      "prompt_bootstrap_95_percent_interval": [
+        -0.017689170230490468,
+        0.008170036304121215
+      ],
+      "prompt_count": 16,
+      "prompts_improved": 11,
+      "per_prompt_deltas": {
+        "69c6b91f7189f44577d2": 0.019887487093607586,
+        "978f7cb5fb01bf22523c": 0.007638911406199138,
+        "9f0be48a1ef2dc4ce885": -0.013423313697179159,
+        "06dcbd59e5bc658ec580": 0.01774933934211731,
+        "f6ea3fb9a241d37c4fb6": -0.021033987402915955,
+        "07075c6a148e625a997a": -0.015316779414812723,
+        "593e8a026f6a96531c26": -0.00507734219233195,
+        "f2e1e3d4c4157387c817": -0.004540830850601196,
+        "18949233e5ae12021263": -0.011010746161142984,
+        "60db927e2b28e456a102": 0.06870140631993611,
+        "b71eae96f6d63766bc0f": -0.006747742493947347,
+        "d85d9a84c873983c21e1": 0.007935047149658203,
+        "c96cd1fce8e47a2aac9a": -0.00905078649520874,
+        "740624682fb48ef9de2f": -0.03408389290173849,
+        "0564e20e17f92949faea": -0.05274855593840281,
+        "ba346dce185dbc104295": -0.036096026500066124
+      }
+    },
+    "native_pixel_mae": {
+      "candidate_minus_baseline": -0.003751380601897836,
+      "prompt_bootstrap_95_percent_interval": [
+        -0.008547981436519572,
+        0.0006537250126712024
+      ],
+      "prompt_count": 16,
+      "prompts_improved": 10,
+      "per_prompt_deltas": {
+        "69c6b91f7189f44577d2": 0.010721762975056967,
+        "978f7cb5fb01bf22523c": 0.0007752527793248495,
+        "9f0be48a1ef2dc4ce885": -0.009089951713879904,
+        "06dcbd59e5bc658ec580": 0.008430302143096924,
+        "f6ea3fb9a241d37c4fb6": -0.0071954553325970965,
+        "07075c6a148e625a997a": -0.006219935913880666,
+        "593e8a026f6a96531c26": -0.002699047327041626,
+        "f2e1e3d4c4157387c817": 0.001969166100025177,
+        "18949233e5ae12021263": 0.0006917491555213928,
+        "60db927e2b28e456a102": 0.009469276914993921,
+        "b71eae96f6d63766bc0f": -0.010789374510447184,
+        "d85d9a84c873983c21e1": -0.0003184378147125244,
+        "c96cd1fce8e47a2aac9a": -0.004497090975443522,
+        "740624682fb48ef9de2f": -0.01392831156651179,
+        "0564e20e17f92949faea": -0.027514191965262096,
+        "ba346dce185dbc104295": -0.009827802578608194
+      }
+    },
+    "clip_text_image_cosine": {
+      "candidate_minus_baseline": 0.001369483768939972,
+      "prompt_bootstrap_95_percent_interval": [
+        -0.0007196499345203237,
+        0.0035478313298275064
+      ],
+      "prompt_count": 16,
+      "prompts_improved": 9,
+      "per_prompt_deltas": {
+        "69c6b91f7189f44577d2": -0.008427510658899942,
+        "978f7cb5fb01bf22523c": -0.001944790283838908,
+        "9f0be48a1ef2dc4ce885": 0.0018561085065205891,
+        "06dcbd59e5bc658ec580": 0.0008262892564137777,
+        "f6ea3fb9a241d37c4fb6": -0.0013627906640370686,
+        "07075c6a148e625a997a": 0.005472660064697266,
+        "593e8a026f6a96531c26": 0.0020618836085001626,
+        "f2e1e3d4c4157387c817": -0.0011492172876993816,
+        "18949233e5ae12021263": 0.0030751427014668784,
+        "60db927e2b28e456a102": -0.0015757282574971516,
+        "b71eae96f6d63766bc0f": -0.0002011259396870931,
+        "d85d9a84c873983c21e1": 0.012138674656550089,
+        "c96cd1fce8e47a2aac9a": 0.003370086352030436,
+        "740624682fb48ef9de2f": 0.0026403963565826416,
+        "0564e20e17f92949faea": -0.0008615354696909586,
+        "ba346dce185dbc104295": 0.005993197361628215
+      }
+    }
+  },
+  "selection_frozen_before_test": true,
+  "remaining_composition_errors": true
+}