Bernini-MNIST: Latent Semantic Planning for Continuous Flow Diffusion

Pretrained model checkpoints for Bernini-MNIST, a toy reproduction of ByteDance's Bernini architecture (Latent Semantic Planning via Multimodal LLMs).

GitHub Repository: https://github.com/ruwwww/bernini-mnist

Visual Results

Checkpoint Manifest

File	Size	Description
checkpoints/vit_mnist.pt	~13 MB	16-patch Vision Transformer Oracle (98.48% classification accuracy)
checkpoints/planner_qwen.pt	~29 MB	Stage 1 Semantic Planner (Qwen3-0.6B + MaskGIT + AdaLN Flow Matching Head)
checkpoints/renderer_semantic.pt	~6.5 MB	Stage 2 2D Spatial ConvFlow Renderer (Smooth Bilinear Continuous Upsampling)
checkpoints/joint_pipeline.pt	~41 MB	Stage 3 Jointly fine-tuned end-to-end weights

Quantitative Benchmarks

• Classification Accuracy: 100.00% on 300 test digits under held-out ViT Oracle.
• Reconstruction Fidelity: 100.00% validation accuracy from real continuous tokens.