themindstudio/SeedVR2-3B-FP8-e4m3fn

SeedVR2 Optimized Weights (FP8 + FP16 VAE)

This repository provides a curated set of weights for the SeedVR2 upscaler, optimized for high-quality inference with a reduced memory footprint on modern hardware.

Original Model & Research

• Title: SEED-VR: Image-to-Video Upscaling with Specialized Research
• Original Repo: ByteDance-Seed/SeedVR2-3B
• Organization: ByteDance (Seed Story Team)

Included Weights

The files in this repository were adapted from the ComfyUI distribution by numz/SeedVR2_comfyUI to provide a streamlined experience for MindCraft Studio.

• seedvr2_ema_3b_fp8_e4m3fn.safetensors: The main 3B parameter transformer model quantized to FP8 (e4m3fn). This allows the model to run on consumer hardware with significantly less VRAM while maintaining state-of-the-art upscaling quality.
• ema_vae_fp16.safetensors: The Exponential Moving Average (EMA) VAE kept in FP16. This ensures that the final image generation avoids the artifacts often introduced by low-precision VAEs.

Usage

These files are intended to be used as a pair within the MindCraft Studio generation engine or any compatible MLX/Python environment.

License

As with the original upstream model from ByteDance, these weights are distributed under the Apache License 2.0.