DiT-XL-2-512

Backup of the facebook/DiT-XL-2-512 Diffusers weights used for the Caaret image-mining run on 2026-10-26.

These are the original weights, not a re-quantized Caaret checkpoint. The noisy GEMM is applied at runtime to transformer_blocks.*.ff.net.2 (4608 → 1152), which is the width the node accepts (K >= 2048). Attention and the MLP up-projections stay in bf16 because their K is 1152 and the node rejects those certificates.

The run wrote a 512×512 ImageNet class 207 image and the node accepted block 179 (e63c6da479afe06d2cb8b6ec65c52ba2f24f8cbd8e1d6efc38623bf7b787a365).

License follows the upstream model: CC-BY-NC-4.0. Original weights: https://huggingface.co/facebook/DiT-XL-2-512

Downloads last month
13
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for CaaretAI/DiT-XL-2-512

Finetuned
(2)
this model