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.
