| { |
| "strict_load": true, |
| "finite": true, |
| "diffusers_version": "0.36.0", |
| "source_declared_diffusers_version": "0.36.0.dev0", |
| "torch_version": "2.14.1", |
| "device": "cpu", |
| "dtype": [ |
| "torch.bfloat16" |
| ], |
| "parameters": 715746304, |
| "tensor_count": 56, |
| "noise_refiner_validation": { |
| "input": "three real COCO128 images encoded by the authentic Z-Image VAE, patchified and embedded with the authentic all_x_embedder, noised at flow-matching t=0.5, real t_embedder adaLN conditioning and upstream create_coordinate_grid RoPE", |
| "vae_sha256": "f5b59a26851551b67ae1fe58d32e76486e1e812def4696a4bea97f16604d40a3", |
| "vae_identical_to_z_image_turbo": true, |
| "assumed_caption_length": 64, |
| "assumption_note": "image position ids are offset by the padded caption length upstream; the official Qwen3 text encoder is out of scope, so a fixed 64-token caption length is assumed", |
| "per_image": [ |
| { |
| "image": "000000000009.png", |
| "source_size": [ |
| 256, |
| 256 |
| ], |
| "latent_shape": [ |
| 1, |
| 16, |
| 32, |
| 32 |
| ], |
| "patch_shape": [ |
| 256, |
| 64 |
| ], |
| "noise_refiner_output_shape": [ |
| 1, |
| 256, |
| 3840 |
| ], |
| "timestep": 0.5, |
| "finite": true, |
| "repeat_max_abs_error": 0.0, |
| "output_abs_mean": 0.1828508973121643 |
| }, |
| { |
| "image": "000000000030.png", |
| "source_size": [ |
| 256, |
| 256 |
| ], |
| "latent_shape": [ |
| 1, |
| 16, |
| 32, |
| 32 |
| ], |
| "patch_shape": [ |
| 256, |
| 64 |
| ], |
| "noise_refiner_output_shape": [ |
| 1, |
| 256, |
| 3840 |
| ], |
| "timestep": 0.5, |
| "finite": true, |
| "repeat_max_abs_error": 0.0, |
| "output_abs_mean": 0.17951683700084686 |
| }, |
| { |
| "image": "000000000034.png", |
| "source_size": [ |
| 256, |
| 256 |
| ], |
| "latent_shape": [ |
| 1, |
| 16, |
| 32, |
| 32 |
| ], |
| "patch_shape": [ |
| 256, |
| 64 |
| ], |
| "noise_refiner_output_shape": [ |
| 1, |
| 256, |
| 3840 |
| ], |
| "timestep": 0.5, |
| "finite": true, |
| "repeat_max_abs_error": 0.0, |
| "output_abs_mean": 0.16232071816921234 |
| } |
| ] |
| }, |
| "context_refiner_validation": { |
| "input": "synthetic Qwen3-shaped caption features (seeded), passed through the authentic cap_embedder", |
| "cap_len": 64, |
| "output_shape": [ |
| 1, |
| 64, |
| 3840 |
| ], |
| "finite": true, |
| "repeat_max_abs_error": 0.0, |
| "output_abs_mean": 0.5141564011573792 |
| }, |
| "elapsed_seconds": 20.54, |
| "limitations": "Component-level CPU test only. The 30-layer DiT trunk, final layer, VAE decoder and text encoder are excluded, so this does not evaluate image generation quality or end-to-end parity with the full NitroFusion2 pipeline. context_refiner inputs are synthetic because the official Qwen3 text encoder was not downloaded. bfloat16 CPU execution is not claimed to match GPU inference. COCO128 may overlap training data." |
| } |
|
|
