| # Verified independent inference examples |
|
|
| These examples were generated on 2026-10-05 UTC from adapter revision `76d175cfa584e93b61f7c49ced6a03e1538ee176` with the repository's unmodified `infer.py`. The base model is `stabilityai/stable-diffusion-3.5-medium` at revision `b940f670f0eda2d07fbb75229e779da1ad11eb80`. The repository remains private. |
|
|
| ## Reproduce the default entry |
|
|
| With access to the gated base model and this private adapter repository, install the public packages named by the model card: `torch==2.7.0`, `diffusers==0.35.1`, `transformers==4.57.1`, `peft==0.17.1`, `accelerate==1.8.1`, `safetensors==0.5.2`, `sentencepiece==0.2.1`, and `protobuf==5.29.5`. Then run: |
|
|
| ```bash |
| python infer.py --prompt "A horse running through a sunlit meadow." --output image.png |
| ``` |
|
|
| The verified defaults are 512 × 512 pixels, 40 sampling steps, guidance scale 1.0, and seed 42. The output matched `p01_s42_optimized.png` byte for byte in this acceptance run. |
|
|
| ## Paired examples and checks |
|
|
| Six prompts were each generated at seeds 42 and 123 with the same base pipeline, dimensions, steps, guidance scale, CPU generator, and prompt for each pair. The base control temporarily disabled the loaded PEFT adapter; the optimized image used the adapter. All 24 PNGs decoded as 512 × 512 RGB images, all 12 pairs differed, and the adapter file SHA256 was unchanged throughout. `manifest.json` records the prompts, parameters, dependency versions, exact revisions, GPU, file SHA256 hashes, and per-pair pixel difference RMS. The source adapter SHA256 is `caa8c9408dc5f2270f89e83009557af9e6de78ebec94c56220fd3b3cf0e058c4`. |
|
|
| This verifies independent inference and adapter effect. It does not establish a quantitative reward improvement or exact reproduction of the historical training run: that run did not record an immutable base-model revision. Images should be reviewed for the intended downstream use. |
|
|
