alastorid/SeedVR2-3B-Swift-MLX-4bit

SeedVR2 — fused Swift/MLX video bundles

Ready-to-use Apple Silicon bundles for WaifuFinder. The root contains 3B (2.4 GB); 7b/ contains 7B (5.14 GB). Both include the positive text embedding. Setup downloads prepared weights directly, without Python, conversion, training or further quantization.

Optimizations

The native exporter joins 3B's packed SwiGLU gate/value projections and both models' VAE query/key/value projections, including biases. Packed integers, scales and offsets are reordered without dequantization or rounding. 198 byte-identical shared attention arrays are represented by aliases, removing 325 MB of duplicate 3B weights. Shared video/text attention and MLP projection calls are batched after verifying weight identity. Transformer attention already has packed QKV. One-block transformer shards support bounded loading. The 7B architecture uses GELU, so it receives VAE attention fusion and block sharding, not SwiGLU fusion.

The Swift runtime uses matching F16 VAE activations/weights and F16 3B transformer activations. The 7B transformer retains F32 to avoid F16 overflow. Normalization and rotary calculations remain F32. The runtime compiles the small SwiGLU elementwise chain and retains a capped 64 MB reusable buffer cache. Streamed inference loads only the active encoder or decoder and one transformer block. Temporal context, causal state and overlapping tiles remain. F16 computation introduces small output differences; this is not bit-exact F32 inference.

Use

./tools/setup-native-uncover.sh       # 3B default
./tools/setup-native-uncover.sh --7b  # optional larger model
./build.sh

Setup verifies pinned hashes and resumes interrupted downloads. The worker uses a physical-memory budget of the smaller of 6 GB and half installed RAM. Streaming is selected automatically for 7B, 8 GB Macs and single-pass 3B jobs. Repeated 3B passes on larger Macs keep weights resident to avoid reloading. --stream-weights also enables it manually. The release targets WaifuFinder's native loader, not Diffusers or an unchanged image loader. See validation.json for measured runtime, memory, pixel differences and test scope. Measurements on a 16 GB M3 with a bounded process budget do not establish zero swap on an actual 8 GB machine. Generated details and spatial tiling can differ from the source.

Measured results

Three paired runs on a 16 GB Apple M3, five consecutive real-video frames, 128×128 input restored to a 256×256 ROI at 2×, seed 42. Time is the median; memory is the highest sampled physical footprint across those runs.

Variant	Seconds	Peak GB
Original 3B, resident	11.58	5.19
Prepared 3B, resident	9.85	4.03
Prepared 3B, streamed single pass	9.03	2.27
Original 7B, streamed	13.49	2.89
Prepared 7B, streamed	12.60	2.24

Each prepared path was faster in all three corresponding runs. The F16 runtime changes produced a mean RGB pixel error around 0.13 on a 0–255 scale, with maximum error 4. Alias-only and shared projection batching preserved output PNG bytes. These measurements describe this fixture, not a speed guarantee for every clip.

Authors and license

Original models: ByteDance Seed, 3B and 7B, Apache-2.0. MLX q4/group64 conversion: lpalbou / AbstractFramework, 3B revision 9b2f6c3c1b1b4548c66f302ec5a4cd3cd9937eb5, 7B revision 22f491ba03647864491520cb803916c21eaad303. Embedding: mlx-community/SeedVR2-3B-mlx revision 46f851fe23b2a9fbefee6419b94e4fdf07b92965. Implementation credit: mflux, MLX-Gen and the Swift reference port.

Model weights and embedding retain Apache-2.0; LICENSE, NOTICE and upstream model cards are included. WaifuFinder adds packed projection fusion, sharding, checksums and native video runtime optimizations. These are derivatives of the named authors' models, not independently trained models. Runtime source retains its separate MIT/upstream notices.