Wan 2.2 Animate 14B, lightx2v 4-step LoRA merged, Q4_0 GGUF
Wan 2.2 Animate 14B with lightx2v's I2V 480p step and CFG distill LoRA (rank 64, Kijai's
lightx2v_I2V_14B_480p_cfg_step_distill_rank64_bf16) merged in at strength 1, quantized to the
same per-tensor types as QuantStack's Q4_0.
Runs in 4 steps at cfg 1 with no LoRA loaded at run time.
How it was made. The LoRA was merged into Comfy-Org's bf16 weights
(wan2.2_animate_14B_bf16.safetensors), then each tensor was quantized to the type QuantStack used for it,
under QuantStack's own GGUF header. Merging into an already quantized Q4_0 loses most of the LoRA, so the
merge happens before quantizing. Check of the method: quantizing the UNMERGED bf16 the same way reproduces
QuantStack's file byte for byte (q4_0, q4_1, f32, bf16 tensors). 1,261 of 1,441 tensors change (lora_up/lora_down
pairs, diff added to weights and diff_b to biases, scale 1, no alphas); the other 180 and the header are identical.
Measured on a Galaxy S25 Ultra (Snapdragon 8 Elite, Hexagon v79) with
Nightmare's stable-diffusion.cpp fork (backend-patches/dit/054):
the same picture as the runtime LoRA (33 dB between the two), and only about 3 to 7% faster at 832x480.
Status: archived. Nightmare does not ship Wan Animate (for motion transfer it uses Wan 2.2 Fun 5B Control, about 18x faster on the phone). Kept here as the record of the conversion. Apache 2.0, as Wan 2.2 and the LoRA.