ORIGINAL OPEN MODEL / MIT SOURCE / LOCAL INFERENCE

ENGINEER
THE MOTION.

A small, trainable text-conditioned spatiotemporal diffusion model built for GPT-DOUG. Own the architecture, training loop, checkpoints and export process. The original model is research-scale—not a copy of Higgsfield's proprietary video engine.

TEXT CONDITIONING3D TEMPORAL CONVOLUTIONDDPM SAMPLINGAPPLE SILICON MPSNO PAID API

SYNTHETIC MOTION TOPOLOGY
Concept illustration; not a generated model output

Checking whether the local Motion-1 engine is connected…

CODE LICENSEMITOriginal model implementation

NETWORK INFERENCE0Paid APIs or remote inference calls

BASE CLIP8×32²8 frames at 32 × 32 pixels

TRAININGLOCALRequires checkpoint to synthesize neural video

01 / LOCAL AI VIDEO STUDIO

Generate a motion sequence

The hosted XUNIA page is a public demonstration of the source. Actual model execution happens only in the local studio on your Mac.

OUTPUT / AWAITING ENGINE

⌁

No generated frames yet

02 / ORIGINAL MODEL DESIGN

Neural architecture

TEXT PROMPT STABLE HASHED TOKENS

↓

TRAINABLE EMBEDDINGS TEXT + TIMESTEP

↓

3 × TEMPORAL RESIDUAL BLOCKS 3D SPATIAL + TEMPORAL CONV · FiLM

↓

DDPM DENOISING REVERSE DIFFUSION

↓

FRAMES → GIF / MP4 LOCAL EXPORT + RECEIPT

03 / INSTALL & TRAIN ON MAC

Launch with free tools

Use Python 3.12; it avoids the Python 3.15 package conflicts encountered with older ML dependencies.

cd "$HOME"
if [ -d "$HOME/gpt-doug-llm/.git" ]; then
  cd "$HOME/gpt-doug-llm"
elif [ -d "$HOME/GPT-DOUG-Motion1/.git" ]; then
  cd "$HOME/GPT-DOUG-Motion1"
else
  git clone https://github.com/sonoxo/gpt-doug-llm.git "$HOME/GPT-DOUG-Motion1"
  cd "$HOME/GPT-DOUG-Motion1"
fi
git pull --ff-only
bash open_video_model/run_mac.sh

This detects an existing Git checkout. If ~/gpt-doug-llm is a non-Git folder, it preserves your files and clones into ~/GPT-DOUG-Motion1. The studio starts before supervised starter training finishes; the neural button unlocks automatically. Synthetic training works without downloading media. You can supply authorized local frame folders for custom training.

04 / MODEL HONESTY

What Motion-1 is—and is not

✓ Genuine open training code

Independent small diffusion architecture, prompt conditioning, local optimizer, and safe state-dict checkpoints. MIT source.

✓ Real output files

Generated GIFs and optional FFmpeg MP4s, with local metadata receipts. Procedural demo mode is labeled separately.

! No photorealistic weights bundled

Trained from scratch, initially on moving shapes. High-quality cinematic synthesis requires large training corpora and compute; the existing free Wan studio remains available separately.

Compare the existing Wan2.2 free cinematic studio ↗

ENGINEERTHE MOTION.

Generate a motion sequence

What Motion-1 is—and is not

ENGINEER
THE MOTION.