Access Sulphur 2 Distilled MLX 8-bit
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This is a derivative of LTX-2. Read and accept the complete LTX-2 Community License Agreement and Attachment A included in this repository before access.
Log in or Sign Up to review the conditions and access this model content.
Sulphur 2 Distilled MLX 8-bit
PocketAI's native Apple Silicon conversion of SulphurAI's Sulphur 2 distilled BF16 checkpoint, based on LTX-2.3. The transformer block linear layers use MLX affine 8-bit quantization with group size 64. Other tensors retain their converted source precision. This was converted directly from BF16; the existing 4-bit release was not used as a source.
Powered by LTX-2 from Lightricks. SulphurAI supplied the fine-tuned distilled checkpoint; PocketAI performed the MLX conversion and validation using Damien Grauet's MLX tools. This community derivative is not an official Lightricks or SulphurAI release.
Package and requirements
- Model package: approximately 29.75 GB; Q8 transformer: 20.59 GB.
- Includes the connector, video VAE encoder/decoder, audio VAE, vocoder, and spatial ×2 upscaler for the distilled 8+3-step pipeline.
- Uses the separate Gemma 3 12B 4-bit text encoder, approximately 8.07 GB. Gemma is not redistributed here and retains its own license. The optional Sulphur prompt enhancer is not required.
- Apple Silicon Mac; 64 GB unified memory or more is recommended and roughly 38 GB disk space including the encoder. Memory requirements increase with frame count and resolution. The tested machine had 128 GB.
Use in PocketAI
Choose Sulphur 2 Distilled MLX 8-bit from PocketAI's gated video catalog,
accept the Hugging Face access terms, and install it. The model ID is
sulphur-2-distilled-mlx-q8, matching the locally validated installation.
PocketAI manages the native MLX runtime and separate text encoder.
The validated preset is 512×320, 25 frames, 24 FPS, 8 initial denoising steps and 3 refinement steps. CFG is 1 and negative prompts are unused.
Python inference
Use the ltx-2-mlx runtime
and its setup instructions. ffmpeg must be available. After downloading
this repository and the pinned Gemma encoder to local folders:
from ltx_pipelines_mlx import DistilledPipeline
pipeline = DistilledPipeline(
model_dir="./Sulphur-2-Distilled-MLX-Q8",
gemma_model_id="./gemma-3-12b-it-4bit",
low_memory=True,
)
pipeline.generate_and_save(
prompt="A small bright red ball rolls across a wooden table in daylight.",
output_path="sulphur.mp4",
width=512,
height=320,
num_frames=25,
frame_rate=24.0,
seed=42,
stage1_steps=8,
stage2_steps=3,
)
Validation
Validated on an Apple M5 Max with 128 GB unified memory:
- All 4,186 transformer tensors were checked against the source.
- 1,632 matrices were quantized to Q8; 2,554 other transformer tensors remained numerically identical. All shared components matched the BF16 conversion.
- Energy-weighted relative weight RMSE after quantization: 0.7946%. This is a numerical weight metric, not a video-quality score.
- A 512×320, 25-frame test with the PocketAI pipeline adapter completed in 22.21 seconds. All frames decoded and the sampled frames were visually inspected. This is a short smoke test, not a comprehensive benchmark.
See the test clip and the provenance and validation manifest.
Sources and license
- SulphurAI/Sulphur-2-base,
sulphur_distil_bf16.safetensors. - Lightricks/LTX-2.3, spatial upscaler ×2 1.1.
- mlx-forge, conversion and affine quantization.
The full LTX-2 Community License, including Attachment A, governs this derivative. The source repository's license is preserved verbatim as well. Read the complete terms before use or redistribution. See MODIFICATIONS.md for conversion details.
- Downloads last month
- -
Quantized