Parity¶
imagehub adds a model only after it reproduces the original implementation. This section shows what was measured, how, and where the claim stops.
123 / 123checkpoints with recorded end-to-end evidence
0largest recorded pixel difference
CPU · fp32where parity is claimed
80models in the side-by-side gallery
The method¶
For each checkpoint, in this order:
- Strict load. Every published tensor is loaded; any missing, extra or mismatched name or shape fails the model.
- Components. The tokenizer, text encoders, transformer or U-Net, VAE and scheduler are each compared with the reference on the same inputs.
- End to end. A full generation from the same initial noise is compared with the reference: final latents and final pixels.
- Tolerance. The accepted difference is written down before the run; for most models it is exactly zero.
The recorded reference is the model's implementation in diffusers 0.40.0 and transformers 5.18.0, or its official implementation where those libraries do not cover it. Measurements use PyTorch 2.14.1.
Reading the results¶
| Page | Contents |
|---|---|
| Gallery | The reference image and the imagehub image side by side, with a slider |
| Results | One row per checkpoint with steps, size and the largest recorded difference |
| Full validation record | Method, tolerances, device tests, limits and the bugs found |
Where the claim stops¶
Parity has a scope
- It holds on CPU in float32 against the stated library versions.
- MPS and CUDA runs are smoke tests, not parity.
- There is no float16 or bfloat16 parity claim.
- It is not a quality benchmark: no FID or CLIP score was run.
- There is no safety checker or invisible watermark; the reference was run without one.
The complete list is in the limits.