Wan 2.2 TI2V 5B — 4-bit weights at 3.39x, and the second model family where 4-bit activations fail

asym_w4a8_int8 build of wan2.2_ti2v_5B, measured against the FP16 original and against a pure W4A4 build of the same source.

9.31 GiB → 2.75 GiB, 3.39x lighter and 1.59x faster per step, with output this bench cannot tell apart from the original at a glance. The W4A4 build is published as a measured negative, not as a file to use.

Method, tools and the full measurement log: https://github.com/JoaoZaokk/comfy-quant-bench

---

The measurement

6 runs (6 prompts × 1 seed), 33 frames at 480 px, 20 steps, euler/simple, one RTX 3090. All three arms are resident on the card — none is offloaded, none is block-split, so the s/step column is a real comparison and not a placement artifact.

Four prompts, one seed. Left FP16, middle this build, right W4A4. The middle column tracks the left one. The right column smears every face, every brick, every hand.

The two findings that outlive this file

1. Four-bit activations fail here too, and this is the second family. W4A4 and W4A8 carry the same 4-bit weights; only the activation path differs. On Qwen-Image-Edit 2511 the W4A4 build rendered pure static; here it renders blur. Different failure mode, same axis. Weights survive 4 bits in both families; activations do not.

2. Latent divergence has a direction bias, and this is the counterexample that shows it. 0.3847 sits inside the band that has always worked on this bench — below Z-Image v2's 0.7854 and Krea 2 Turbo's 0.5843, both of which render fine. And the picture is degraded.

Until now this bench only had the opposite failure: high divergence with a good picture, because a tiny perturbation reroutes the sampler somewhere else that is also good. This is the mirror, and the mechanism is plain once seen: blur is a small latent change. Smoothing moves you less far from the reference than going somewhere sharp and different does. So the metric rewards degradation, as long as the degradation is smooth.

That is stronger than "there is no threshold". It means divergence is not merely noisy — it is biased toward soft failure, and a build that blurs will always flatter itself on this number.

The file

Source wan2.2_ti2v_5B_fp16.safetensors, 9,999,658,848 B, sha256 0a09fde7bc7f6a0d456f87acc914d3380bef60bc0f47f8ff0a94e2ae6a95ef76. The W4A4 build carries the same source hash, so the two are comparable by construction. Backend recorded in the sidecar: comfy_kitchen.backends.cuda.

Calibrated on real activations — 6 sampling runs with forward hooks on all 300 Linear layers, reservoir-sampled across steps, not just step 0.

The profile, checked against someone else's work

The layer set comes from a wan_2_2 profile that was validated against Comfy-Org's own wan2.2_animate_14B_int8_convrot: 480 layers matched, 480 they quantize, zero either way. The 2.1 profile this bench already had missed 80 of them — cross_attn.k_img and .v_img, the image cross-attention projections, which do not exist in a pure T2V model. On this TI2V 5B the same pattern degrades to 300 by itself, because this model has no k_img/v_img either.

Running it

Needs a 48-channel VAE: wan2.2_vae.safetensors. The wan_2.1_vae that ships beside it in many installs has 16 channels and will decode this model's latents into a plausible, silently wrong picture. That is not a hypothetical — it was caught here by counting channels before decoding, not after.

What is NOT covered

• Six runs, one seed, 33 frames, 480 px, one sampler, one card. The ladder's own output says six runs is a small sample for a quantity this noisy, and the spread proves it: W4A8's divergence spans 0.1185–0.3275 around a mean of 0.2115.
• "Indistinguishable at a glance" is a human verdict on a contact sheet, not a perceptual metric. No LPIPS, no FVD, no user study.
• 33 frames is 1.3 seconds. Nothing here says how either build behaves over a long generation.
• The audio path does not exist in this model, so nothing about audio is claimed.
• convrot_groupsize 256 only, and no mixed build was tried — the choice here was the clean single axis (all-W4A4 against all-W4A8), not a search.
• The W4A4 weights are published as evidence, not as a usable model. Use the W4A8.

Credits

• Wan-AI — Wan 2.2 TI2V 5B itself.
• Comfy-Org — the repackaged single-file FP16 weights this was converted from, the 48-channel VAE, and the animate_14B_int8_convrot build whose layer list was used to validate the profile. That validation is why 80 layers are not silently missing from this file.
• ComfyUI / comfy-kitchen — the asym_w4a8_int8 format and the CUDA kernels that run it.