Qwen-Image 2512 — 4-bit weights at 3.53x, and the third family where 4-bit activations fail

asym_w4a8_int8 build of qwen_image_2512, measured against the BF16 original and against a pure W4A4 build of the same source.

38.05 GiB → 10.79 GiB, 3.53x lighter and 3.65x faster per step, with output this bench cannot separate from the original on a contact sheet. The W4A4 build is published as a measured negative, not as a file to use.

This is the non-edit half of the Qwen-Image family. The editor lives in its own repo: Qwen-Image-Edit-2511-W4A8-ConvRot.

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

---

The measurement

12 runs (6 prompts × 2 seeds), 1024 px, 20 steps, cfg 2.5, euler/simple, one RTX 3090. The transformer is the only thing that changes — same encoder (qwen_2.5_vl_7b), same VAE, same prompts, same seeds.

Four prompts, one seed. Left BF16, middle this build, right W4A4. The middle column tracks the left one — apple, face, OPEN sign, night market. The right column still shows the subject and is unusable.

The BF16 arm's 5.722 s/step is not a fair kernel comparison: 38.05 GiB does not fit a 24 GB card, so it streams weights every step. The two quantized arms are resident. What the column does say honestly is the thing a user feels — the smaller file finishes 3.65x sooner on this hardware.

Third family, third failure mode, one axis

W4A4 and W4A8 carry the same 4-bit weights. Only the activation path differs. Across three families on this bench:

Three different-looking failures, one cause. Weights survive four bits in every family measured here; activations do not survive in any of them. That is the transferable result, and it is worth more than any single file in this repo.

Note also what the divergence numbers do not do: Wan's blur scored 0.3847, lower than this build's perfectly good 0.3875. The metric is biased toward soft failure — blur is a small latent change — so it cannot be used alone to accept a build. Only a render decides.

The files

Source qwen_image_2512_bf16.safetensors, 40,861,031,488 B, sha256 beginning d8c57b76f262e1ef27f5eca1. The W4A4 build carries the same source hash, so the two arms are comparable by construction.

Calibrated on this checkpoint, not on the Edit's analysis. The two models differ by 72 bytes in file size and share the MMDiT architecture, so reusing would have saved an hour of GPU — and quant_mixed refuses foreign analyses by default, correctly: different weights give different per-layer errors, and reusing would pick formats from measurements of another model.

Backend recorded in the sidecar: comfy_kitchen.backends.cuda on all four ops.

What is NOT covered

• 12 runs, 6 prompts, 2 seeds, one resolution, one sampler, one card.
• "Indistinguishable on a contact sheet" is a human verdict, not a perceptual metric. No LPIPS, no FID, no user study. Judge the sheet yourself.
• The text encoder is not in this repo and is shared with the Edit model; a convrot_w4a4 build of qwen_2.5_vl_7b exists separately on this account.
• The VAE is not quantized and will not be — 254 MiB against a 38 GiB transformer.
• convrot_groupsize 256 only; no mixed build was tried. The choice here was the clean single axis, all-W4A4 against all-W4A8, not a search for an optimum between them.
• The W4A4 weights are not published. Its contact sheet is, so the claim is checkable.

Credits

• Qwen (Alibaba) — Qwen-Image itself. Everything here is a re-encoding of their weights.
• Comfy-Org — the repackaged BF16 single-file weights this was converted from, and the qwen_image layer profile this bench uses, which was derived from their own int8_convrot build of the Edit model and matches it 840/840 with no slack either way.
• ComfyUI / comfy-kitchen — the asym_w4a8_int8 format and the CUDA kernels that run it.