Giniiki/FastWan2.2-TI2V-5B-mlx-q8

FastWan 2.2 TI2V-5B — MLX q8 (Pithos)

The image-to-video checkpoint Pithos downloads. DMD-distilled Wan 2.2 TI2V-5B (3 denoise steps, sigmas [1.0, 0.757, 0.522, 0.0], renoise, guide 1), quantized for Apple Silicon.

file	size	what
model.safetensors	5.4 GB	DiT, q8 (group 64) — attention + FFN Linears quantized, per mlx-video's predicate
t5_encoder_q8.safetensors	6.4 GB	umT5-XXL, q8 (group 64) everywhere, scales kept float32
vae.safetensors	2.8 GB	Wan 2.2 VAE (encoder + decoder), unquantized
tokenizer.json + configs	17 MB	umT5 tokenizer (T5Tokenizer / Unigram)
config.json	—	model config incl. the fastwan_dmd recipe and quantization block

Provenance: DiT quantized 2026-08-30 from lBroth/FastWan2.2-TI2V-5B-MLX (bf16, rev de2998ac446581895089ae363f4dcc0c02ca9ea2), itself converted from FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers. The T5 and VAE come from that same conversion (T5 quantized here); the tokenizer files are copied from Wan-AI/Wan2.2-TI2V-5B google/umt5-xxl/. Quantization used mlx.nn.quantize via mlx-video's converter (mlx 0.32.2).

Two facts written into Pithos's tests, kept here so a re-conversion does not lose them: the T5's quantization scales must stay float32 (bf16 scales measurably degrade the encoder, 0.023 → 0.042 relative), and the DMD distill must be sampled at its trained 121-frame profile with renoise — plain euler over its sigmas, or off-profile frame counts, dissolve the clip's tail.

License: Apache-2.0, inherited from the base model.