SeedVR2-3B-MLX-Serve

mlx-serve-native repack of ByteDance-Seed/SeedVR2-3B for one-step diffusion image/video restoration and upscaling.

Weights are a straight fp16 passthrough from the Comfy-Org/SeedVR2 fp16 mirror (diffusion_models/seedvr2_3b_fp16.safetensors, vae/ema_vae_fp16.safetensors) plus the original repo's pos_emb.pt text-conditioning tensor — no requantization. Converted with mlx-serve's tests/convert_seedvr2_weights.py.

Files

config.json — model_type: seedvr2
dit.safetensors — the 635 NaDiT tensors (~6.8 GB fp16)
vae.safetensors — encoder + decoder (~0.5 GB fp16)
pos_emb.safetensors — the fixed text-conditioning tensor (~1 MB)

Usage

mlx-serve --model-dir ~/.mlx-serve/models/ByteDance-Seed

Serves POST /v1/images/upscales (images) and POST /v1/video/upscales (video frames).

License

Apache-2.0, inherited from the original ByteDance-Seed/SeedVR2-3B release.

Downloads last month: -

MLX

Hardware compatibility

Quantized

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for justintime47/SeedVR2-3B-MLX-Serve

Base model

ByteDance-Seed/SeedVR2-3B

Quantized

(3)

this model