AnimeGen-T2V-A14B-Lightning-int4

AnimeGen-T2V-A14B-Lightning-int4 (MLX)

An MLX (Apple Silicon) conversion of aidealab/AnimeGen-T2V — an anime-specialized full fine-tune of Wan2.2-T2V-A14B (dual-expert MoE) — quantized to int4 and shipped as a pre-merged Lightning checkpoint for few-step, CFG-free inference.

This variant is built for the Swift-MLX wan-core / Bernini-R runtime: it uses the exact flat, 1095-key A14B checkpoint layout that runtime loads, so the DiT experts drop in unchanged.

What's in here

Flat checkpoint (stock-Wan VAE + umT5 reused from Wan2.2-T2V-A14B, byte-identical):

How it was made

1. The two fine-tuned experts from aidealab/AnimeGen-T2V (original-Wan key naming, bf16) were converted 1:1 into the MLX A14B layout (sanitize_wan_transformer_weights; the key set is asserted equal to the base A14B contract, 1095 keys, 0 added / 0 dropped).
2. AIdeaLab's acceleration recipe — the lightx2v Wan2.2-Lightning Wan2.2-T2V-A14B-4steps-lora-250928 LoRA — was merged offline per-expert with AIdeaLab's weighting (set_adapters([high, low], [2.0, 1.0]); kohya W += strength·(alpha/rank)·(up @ down)).
3. The attention + FFN linears of both experts were quantized to int4 (group size 64); VAE + umT5 kept full precision.

Recommended inference recipe (from the AnimeGen model card)

• Sampler: euler, CFG-free (guidance 1.0)
• 8 steps, shift 3.0 (FlowMatchEuler)
• 832×480 or 1280×720, frames 16·seconds + 1 (must be 4n+1), 16 fps

Footprint

int4 measured ~42 GB peak at 832×480 / 17 frames / 8 steps on an M-series GPU (fits a 64 GB Mac).

License & attribution

Apache-2.0. Derived from aidealab/AnimeGen-T2V (AIdeaLab, developed with GENIAC/METI support) and the lightx2v Lightning LoRA. This is an unofficial community MLX conversion; not affiliated with AIdeaLab.