hammamiomar/SAM_CAM_BAM-models

SAM_CAM_BAM Dream model pack

This is the immutable browser artifact pack for the experimental Dream mode in SAM_CAM_BAM. It is not a general-purpose pipeline. The app combines these files with a small handwritten WebGPU block-transfer kernel.

Dream encodes a live source frame and an uploaded donor image with DINOv2-B/14-with-registers, selects eight learned rank-3 block subspaces from a clicked donor patch, transfers those projections to every source patch, and decodes the edited patch field with an RAE ViT-B decoder. Inference stays in the browser; this repository only serves static model bytes.

Files

File	Bytes	SHA-256	Purpose
dinov2-vitb14-reg-q4f16.onnx	51,374,908	ef0feac82fd3fc1310aba8c38390fee86367b8c9e553a9846ce3488c29494ee5	DINO encoder when shader-f16 is available
dinov2-vitb14-reg-q4.onnx	59,978,957	02996745913d0654d8571cef0de689a7d3a4870b3deacf9b41d6c13b6dca7fd1	Portable DINO encoder fallback
dinov2-vitb14-reg-bsf-g8192-r3-f16.bin	38,547,520	54f6243b339af889ac826faf9fb5b535e3135e41068117ba6c78f95ad7ecbf61	8,192 orthonormal rank-3 frames, position mean, and encoder affine data
rae-vitb-decoder-int8-qdq.onnx	87,946,239	cb31eec571791e3f3ab9cc03abd63841603ffbce85853423f47a75e53a907691	Static 256-patch to 256×256 RGB decoder

manifest.json records the complete source identities, tensor contracts, and verification results. The pack is consumed by commit-pinned URLs; main is not a runtime contract.

Verification

The build script refuses source artifacts whose hashes differ from its pinned inputs. The published pack passed:

• maximum BSF frame Gram error: 5.96e-7;
• decoder int8-QDQ versus FP32 PSNR: 39.45 dB;
• float16 versus float32 selector mean top-8 overlap: 1.0; and
• exact ordered top-8 agreement on 254 of 256 fixture patches, including the default strongest patch.

The reproducible build and checks live in tools/dream_models.py.

Provenance and terms

• The two encoder graphs are byte-identical copies from onnx-community/dinov2-with-registers-base, revision fcd050203baf48e192055356f379b5949a111255. DINOv2 code and model weights are Apache-2.0.
• The decoder was exported from RAE source bytetriper/RAE at a4d18c4db766419cbe7cb8c02cd9f7ceb0ec9041, then its 74 constant matrix weights were quantized to int8 QDQ. The RAE source repository is MIT.
• The ViT-B decoder checkpoint is decoders/dinov2/wReg_base/decB_ganv3/dinov2_decoder.pt from nyu-visionx/RAE-collections, revision 1be4f03273523431f099a934da4cf1940dc6039f. That repository declares the MIT license.
• The BSF field is an Omar Hammami-trained artifact over calibrated DINOv2 patch states. Its exact checkpoint and calibration hashes are in the manifest.

Limits

Dream is a creative feature-visualization experiment, not segmentation, depth, restoration, or identity-preserving image editing. Inputs are resampled to 224×224 and output is 256×256. Results can distort content, faces, text, and fine structure and inherit limitations from the source models and their training data.