A small, staged image-to-video workflow for an Apple Silicon Mac with 16 GB unified memory. Neural Zero contributes local integration, bounded weight loading, compatibility patches, process separation and reproducible documentation. The foundation model is Lightricks LTX-Video, using baisampayans/ltx-mlx.
This Static Space hosts documentation, source and an example. It does not generate videos online. Inference runs on your Mac.
These results come from a test on a Mac mini M4 with 16 GB unified memory, using an existing environment. A completely fresh installation on another Mac has not been verified.
requirements.txt.0d884df6c5f086915e4ffdab950d914757c306f2.ltxv-2b-0.9.8-distilled, T5v1.1-XXL Q8 affine / group size 64.Read benchmarks, third-party notices and review results before use.
Use an Apple Silicon Python 3.11 environment with Metal support and a working FFmpeg/ffprobe installation. The verified FFmpeg build was 9.0.2 with libx264. Do not install PyTorch, Diffusers or Transformers for this workflow.
From the directory containing this README:
python3.11 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
git clone https://github.com/baisampayans/ltx-mlx.git ../ltx-mlx-upstream
git -C ../ltx-mlx-upstream checkout 0d884df6c5f086915e4ffdab950d914757c306f2
ffmpeg -version
ffprobe -version
The backend checkout must be clean. run.py copies its Python package into a
new work directory and applies backend.patch there. The patch preserves
checkpoint-compatible parameter-free block RMSNorm and additive attention-mask
dtype parity. The diffusion worker also casts the FP32 RoPE result back to the
query/key dtype. The original checkout is preserved.
Read the Lightricks model terms and the T5 information linked in the notices. Model terms are separate from the Apache-2.0 license on this integration.
.venv/bin/python download_weights.py ../ltx-models --accept-model-license
This uses only Python's standard library. It downloads eight required files from
the official Lightricks/LTX-Video repository at the revision in
weights.lock.json, then verifies SHA-256. No authentication is passed by the
script. No weights or derived weight caches are included in this release.
An interrupted .partial download requires manual inspection/removal before retry;
the script never silently replaces it.
The official FP32 T5 shards are about 19 GB, in addition to the LTX checkpoint. Keep at least 35 GB free after downloading for the BF16 and Q8 conversions, temporary staging and outputs. The first run intentionally regenerates caches.
Close other memory-intensive applications. Run one job at a time:
.venv/bin/python run.py \
--backend ../ltx-mlx-upstream \
--models ../ltx-models \
--image examples/input.png \
--prompt "A golden sphere on a blue floor. Very subtle slow camera push in. Preserve the sphere and lighting. No text, no logos." \
--work ../local-lab-run-001
The output is ../local-lab-run-001/result.mp4. The directory must not already
exist and must be outside the release directory. To use your own image, replace
--image and describe it in an English prompt. Use only images you have rights to.
This minimal version fixes geometry, seed and step count; it does not expose
unverified resolution presets.
The work directory contains logs, checksummed conditioning handoffs, Q8 weights and temporary runtime metadata. Keep it private: it records your prompt and local paths. Do not upload it to a Space. The packaged benchmark logs have their local paths replaced with placeholders.
The clip lasts 2.04 seconds, has no audio and is low resolution. Technical validation does not establish cinematic quality, motion quality, semantic fidelity or performance at other resolutions. The included geometric image is a simple reproducibility fixture, not a production-quality showcase.
The decoder previously reached about 12.57 GB of MLX allocations. This is not total system memory. A 16 GB Mac can still experience compression, swap or memory pressure. The lightweight runner checks pressure, stops on critical pressure, limits each child to 900 seconds and waits for normal pressure between stages. It is not a full system memory safety guarantee.
No Neural Zero Studio engine, commercial QA system, FastMetal implementation, paid service, hosted inference, Blender project or third-party media is included. This release does not claim to train a new foundation model.
Los resultados corresponden a una prueba en Mac mini M4 de 16 GB con un entorno existente. No se ha comprobado una instalación completamente nueva en otro Mac.
Edición mínima para generar un clip local de 384×224, 49 fotogramas y cuatro pasos en Apple Silicon. El ejemplo se ejecutó desde los pesos originales en 83,28 s, sin contar descarga ni instalación. La medición histórica de 25,76 s reutilizaba cachés y no equivale a una ejecución desde cero. Hugging Face aloja la documentación y el código; la generación ocurre en tu Mac. Revisa las licencias de los pesos y las limitaciones de memoria. El producto comercial Neural Zero Studio queda separado.