Nacre v1 — MLX layout
The Nacre v1 weights converted for the Swift/MLX port
(mlx-nacre-swift): conv weights OHWI, per-tensor exact from the release (oracle/convert_weights.py). Two
tensors of the release are not parameters of the MLX module tree and are omitted: the Swin
relative_position_index buffers (rebuilt from the window size) and the VQ codebook (Nacre decodes continuous
latents).
| file | lane |
|---|---|
nacre_v1_mlx_fp16.safetensors + vq_f4_mlx_fp16.safetensors |
shipping (every reduction is fp32 inside the port) |
nacre_v1_mlx_fp32.safetensors + vq_f4_mlx_fp32.safetensors |
parity |
Licences, provenance, metrics and limitations: see the main model card. The VQ-f4 autoencoder is CompVis (MIT); see NOTICE.
- Downloads last month
- 15
Hardware compatibility
Log In to add your hardware
Quantized
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support