Z-Image-Turbo ONNX (Browser-Oriented Sharded Packaging)

What Changed

• The WebNN-style transformer packaging was first reordered into a cleaner, more contiguous execution layout. In practice this acts like a "defragmented" transformer graph.
• That reordered transformer was then split into multiple ONNX shards.
• The goal was to reduce peak per-session model size for browser runtimes, especially on tighter mobile GPU memory budgets.

Repository Layout

• tokenizer/: tokenizer files used to turn prompt text into model inputs
• onnx/text_encoder_model_q4f16.onnx + .onnx_data: text encoder that converts tokenized prompts into conditioning embeddings
• onnx/vae_decoder_model_f16.onnx: VAE decoder that turns final latents into pixels
• onnx/vae_pre_process_model_f16.onnx: VAE helper used in the decode path
• onnx/scheduler_step_model_f16.onnx: scheduler update model used between denoising steps to produce the next latent sample
• onnx/transformer_model_q4f16.onnx + .onnx_data: reordered / defragmented monolithic transformer kept as the base form used to build the shards
• onnx/transformer_model_q4f16_shard*.onnx + matching .onnx_data: the main transformer split from the reordered monolithic form into multiple shards to reduce peak session size
• onnx/transformer_shards.json: manifest describing the shard set

Notes

• Base model license and usage terms come from the upstream Tongyi-MAI/Z-Image-Turbo release.
• This repo focuses on browser-friendly packaging rather than training or architecture changes.