wabibito/Onyx-FastWan22-TI2V-5B-mlx-4bit

Onyx-FastWan22-TI2V-5B-mlx-4bit

On-device video generation model for Onyx (iPhone + Mac): a 4-bit MLX conversion of FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers, FastVideo's 3-step DMD distillation of Wan-AI/Wan2.2-TI2V-5B-Diffusers.

• —Transformer: 5B dense DiT, 4-bit affine (group 64), norms/embeddings high-precision — 2.6 GB
• —Text encoder: umT5-XXL, 4-bit affine (group 64) — 3.0 GB
• —VAE: Wan 2.2 3D causal VAE, float32 (decode precision is load-bearing) — 2.6 GB
• —Inference: 3 steps, no CFG, DMD timesteps [1000, 757, 522] (σ = t/1000 — the list is already in post-shift space; do not re-apply flow_shift)
• —Modes: text-to-video and image-to-video (TI2V: conditioning image pinned over latent frame 0, its tokens at timestep 0)
• —Native grid: 121 frames × 704 × 1280 @ 24 fps (other resolutions run, quality degrades away from the trained grid)

Converted with Onyx's own streaming per-tensor exporter (bounded memory; only matmul weights quantize, norms/embeddings/adaln stay high-precision). Loads in OnyxDiffusion's Wan 2.2 engine.

Attribution & license

Apache-2.0, inherited from both upstreams. Distillation by the FastVideo team (hao-ai-lab); base model by the Wan team at Alibaba. This repo repackages their work for Apple-silicon MLX inference; all credit for the model itself belongs to them.