SCAIL-2 14B MLX for mere.run
This repository contains an immutable MLX-native packaging of the official
zai-org/SCAIL-2 model for native Swift/MLX inference in mere.run.
It supports reference-subject animation, performer replacement, and up to six palette-masked subjects. The package contains model artifacts only. It does not contain Python, a checkpoint conversion utility, the upstream SCAIL inference repository, Maestro, WanGP, ComfyUI, or translated source code.
Provenance
- Source model:
zai-org/SCAIL-2 - Source model revision:
150cc0ca4e98e50e60b9295dacde39442fdccab2 - Packaging: lossless tensor layout/name conversion to safetensors for the independently maintained native Swift/MLX runtime
- Precision: BF16 transformer and UMT5, FP16 OpenCLIP and Wan VAE, with explicitly recorded FP32 islands
- License: the official SCAIL-2 model card declares the model MIT
MERERUN_CONVERSION.json records the immutable source hashes, output artifact
hashes, byte counts, component mapping, and packaging contract.
provenance.json retains the exact source component paths and source hashes.
The separately published SCAIL reference-code repository is Apache-2.0. No files from that repository are included in this model package.
Install and use
mere.run model pull video-scail2-14b-mlx
mere.run video animate "a dancer in a red silk dress" \
--reference reference.png \
--reference-mask reference-mask.png \
--driving-video driving.mp4 \
--driving-mask driving-mask.mov \
--tail-policy pad-trim \
--audio-source driving \
--output result.mp4
Use mere.run video prepare-masks with the separately licensed native
vision-segment-sam31 model to prepare palette-safe reference and driving
masks. SAM weights are not included here.
Required layout
config.jsonmodel.safetensors.index.jsonand eight transformer shardsmodel-fp32-islands.safetensorst5_encoder.safetensors.index.jsonand three UMT5 shardsclip.safetensorsvae.safetensorstokenizer.jsonmererun_model.jsonprovenance.jsonMERERUN_CONVERSION.jsonLICENSE