Reproducibility¶
Same seed, same image¶
Pass seed and the initial noise is drawn on the CPU with torch.Generator("cpu").manual_seed(seed),
the way the reference does it (shown for Stable Diffusion; each model page
notes where an architecture differs), whatever device runs the model.
a = pipe("a cat", seed=1).images[0]
b = pipe("a cat", seed=1).images[0]
assert a.tensor.equal(b.tensor)
Leave seed out and a random one is chosen and recorded in result.metadata, so an image can be
repeated afterwards. Pass either seed or latents (your own noise tensor), not both.
Same image as the reference¶
For the recorded checkpoints, on CPU in float32, imagehub produces the same pixels as each model's recorded reference implementation. The validation record identifies the exact diffusers, transformers or official-code revision used. The claim comes with its evidence:
- Results: the recorded differences, one row per model.
- Gallery: the reference and imagehub images side by side.
- Validation record: the method, the tolerances and the bugs found on the way.
What can still differ¶
- A different device or precision: see Devices and precision.
- A different library version of the reference: the record names the versions it was measured with.
- Different weights: the catalog pins a revision; another revision is not covered.
Alignment matters
Safetensors files are memory-mapped so weights are read at their published alignment. CPU BLAS can pick a different summation order for differently aligned data, which was one of the real sources of a mismatch found during validation.