SceneWorks/Mage-Flow-Components-mlx

Mage-Flow — shared components (MLX, per-tier)

The Qwen3-VL-4B text encoder and Mage-VAE shared by every Mage-Flow variant, re-hosted for SceneWorks as physically distinct MLX quantization tiers.

Why this repo exists

Microsoft publishes six Mage-Flow checkpoints (Base, RL, Turbo, and the three Edit twins). The text encoder (8.875 GB) and VAE (0.345 GB) are bit-identical across all six — only the 8.232 GB DiT differs. Mirroring each variant as a complete snapshot would cost 105.04 GB for a full install. Hosting the shared components once here, and only the per-tier DiT in each variant mirror, brings that to 58.65 GB, and makes a second variant an ~8.2 GB delta rather than another 17.5 GB.

Layout

Each tier directory is self-contained:

<tier>/text_encoder/   Qwen3-VL-4B — config, tokenizer/processor assets, model.safetensors
<tier>/vae/            Mage-VAE

The VAE is dense in every tier, deliberately. 59 of its 60 quantizable 2-D weights are adaLN_modulation projections that the decoder folds away at load (adaLN depends only on t, and the one-step decode always runs t = 0). Packing them on disk would fold quantized codes and silently corrupt the decode. The disk cost of that correctness is ~0 — the VAE is conv-dominated, and the comparable Z-Image VAE shrinks only 1.8% under q4.

Quantization is MLX q4/q8 at group size 64, packed with the same mlx_rs::ops::quantize(bf16, 64) call the load-time path uses — a pre-quantized tier is bit-identical to quantizing the dense weights at load.

Provenance

Converted from microsoft/Mage-Flow-Base at revision 59a9cfd58cf6ecef28245852c6bdace3f12428a2 with `mlx-gen-mage`'s examples/mage_prequant.rs. The text encoder and VAE tensors are identical in all six upstream repos, so the choice of source variant is immaterial.

Upstream: microsoft/Mage · arXiv:2607.19064 · MIT.