SeedVR2 1.4B MLX Q4
Unofficial MLX affine 4-bit / group-32 conversion of lvladikov/SeedVR2-1.4B, a community six-block distillation of ByteDance SeedVR2 7B. This is a complete still-image upscaling package: Q4 transformer, unmodified FP16 VAE, and fixed positive text embeddings. No text encoder is required. This checkpoint requires the six-block, 3072-wide architecture with normal GELU MLPs; a standard SeedVR2 3B loader cannot load it. Quantization changes the model's predictions; sample comparisons and measured results are in validation.json and the examples directory.
Native Swift/MLX inference and the macOS model selector are available in zimageapp/zimage, on the seedvr2 branch:
swift build -c release --product ZImageCLI
.build/release/ZImageCLI seedvr2 --model-path /path/to/package \
--input input.png --output output.png --resolution 2x --steps 1 \
--tile-size 512 --tile-overlap 64 --seed 42
Use one step, without approximate prediction reuse. The VAE uses overlapping tiles for larger images; the transformer attends over the whole latent image. Peak memory therefore grows with output resolution. Supported app scales are 2x and 4x. The 3B model remains available separately.
The original Apache-2.0 license applies. Credit to ByteDance Seed and lvladikov; this derivative is not an official release from either author. conversion.json records the pinned upstream revision, source hashes, and quantization details; package-manifest.json records the distributed file hashes.
4-bit