AnimeGen-I2V-A14B-Lightning-int4

AnimeGen-I2V-A14B-Lightning-int4 (MLX)

An MLX (Apple Silicon) conversion of aidealab/AnimeGen-I2V — an anime-specialized image-to-video fine-tune of Wan2.2-I2V-A14B (dual-expert MoE) — quantized to int4 and shipped as a pre-merged Lightning checkpoint for few-step, CFG-free inference.

Built for the Swift-MLX wan-core / Bernini-R runtime. The i2v sibling of AnimeGen-T2V-A14B-Lightning-int4.

What's in here

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

Image conditioning (channel-concat, no CLIP)

The DiT input is [noisy_latent(16), mask(4), image_latent(16)] = 36 channels. Encode a video whose first frame is the conditioning image and the rest zeros through the 16-ch WanVAE → z_video(16); build a first-frame temporal mask(4); y = [mask, z_video](20) is constant across steps and concatenated onto the 16-ch noisy latent each denoise step. No separate image encoder (image_dim is null).

Recommended inference recipe (from the AnimeGen-I2V model card)

• Sampler: euler, CFG-free (guidance 1.0)
• 4 steps, shift 3.0 (FlowMatchEuler)
• 832×480, frames 16·seconds + 1 (must be 4n+1), 16 fps
• Prompt prefix: "Japanese anime style, "

How it was made

Experts converted 1:1 from aidealab/AnimeGen-I2V (original-Wan keys → sanitize; key set == the A14B 1095-key contract). The lightx2v Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1 LoRA merged offline per-expert (set_adapters([high,low],[1.0,1.0]), kohya). Attention + FFN linears quantized int4/g64.

Footprint

int4 measured ~43 GB peak at 832×480 / int4 / 4 steps on an M-series GPU (fits a 64 GB Mac).

License & attribution

Apache-2.0. Derived from aidealab/AnimeGen-I2V (AIdeaLab / GENIAC) + the lightx2v Lightning LoRA. Unofficial community MLX conversion; not affiliated with AIdeaLab.