JoaoZaokk/Z-Image-Turbo-W4A4-ConvRot

Z-Image Turbo — ConvRot 4-bit for ComfyUI, 3.75x lighter and 2.65x faster per step

Native convrot_w4a4 and mixed 4/8-bit builds of the official Tongyi-MAI/Z-Image-Turbo, in ComfyUI's own quantization format. No custom node: the stock UNETLoader reads the per-layer metadata and dispatches on it.

11.46 GiB → 3.06 GiB, and 0.903 → 0.341 s/step on an RTX 3090 at 1024×1024. Thirty-six renders across six prompts and two seeds, none broken.

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

---

The files

file	format	GiB	median per-layer error	latent divergence	s/step
zimage_turbo_w4a4.safetensors	170 × convrot_w4a4	3.06	0.1228	0.6589	0.341
zimage_turbo_mixed.safetensors	113 × 4-bit + 57 × 8-bit	3.18	—	0.5500	0.391

Source: 11.46 GiB BF16, 6.15 B parameters. Eight steps, 1024×1024, euler/simple, cfg 1.0, two seeds, six prompts.

"Per-layer error" is, for each layer, the measured relative error of the format that layer actually received, on the activations it saw during real sampling — captured with forward hooks and reservoir-sampled across the steps, not on random input. The floor for comparison is err_bf16 = 0.0019: that much error is not quantization at all.

err_bf16   median 0.0019
err_w4a4   median 0.1228
err_w4a8   median 0.0389      <- 3.16x cheaper, at 8 bits

Dispatch was counted, not assumed, with the weights on the card: 170 quantized modules, 8/8 quantized forwards, 0 dequantize, convrot_linear_dtype=int4 on comfy_kitchen.backends.cuda — the native INT4 MMA path, not the INT8 fallback branch.

---

What they look like

Six prompts, seed 1. Reference on top.

The hardest cell is the enamel sign, because text is where quantization damage shows first:

BF16 reference	W4A4	mixed

---

How to read the divergence column, and why it is not "damage"

0.6589 is the mean distance between this build's latent and the BF16 latent on the same prompt and seed. It sounds alarming and it is not: at 8 steps a small perturbation reroutes the sampler and it arrives at a different good image. All 36 renders are usable.

The between-run spread inside a single arm is 0.54, which is close to the gap between the two arms' means — so the mean comparison is weak evidence here and the paired comparison is the real one:

paired against zimage_turbo_w4a4, run by run
zimage_turbo_mixed    mean delta -0.1089    wins 11/12

The mixed build is measurably closer to the BF16 latent — it wins 11 of 12 paired runs — for 0.12 GiB and 15% more time per step. If you want the smallest and fastest, take W4A4; if you want the one that tracks the original trajectory, take mixed.

That result does not carry across the family. The same converter at the same threshold, run on the de-distilled sibling (ostris/Z-Image-De-Turbo), gives the mixed build only 7 of 12 — a split decision. Whether promoting 57 layers to 8 bits buys anything depends on the checkpoint, not on the recipe.

---

Where Z-Image sits

The per-layer error a model survives before its output breaks is not a property of the format. Measured across four architecture families on this bench:

model	parameters	tolerated	not tolerated
Wan 2.1 VACE	1.3 B	0.0546	0.0793
Z-Image	~6 B	0.1421	0.1848
Krea 2 Turbo	12.82 B	0.1377	not reached on the only axis available
HunyuanVideo 1.5	~13 B	0.1837	0.2147

Z-Image's upper bound was filled by pushing convrot_groupsize down: cg 256 (0.1312) good, cg 64 (0.1421) good, cg 16 (0.1848) destroyed in 3 of 3 renders. At 0.1228 these builds sit well inside the working band.

The mechanism written down before that run was backwards, and the pre-written control caught it: a Hadamard rotation of size N spreads each outlier across N channels, so a larger group mixes more and flattens outliers harder. "Coarser rotation, more error" was the intuition of a per-group quantizer, and the rotation is not that.

---

Using them

Drop into ComfyUI/models/diffusion_models/ and load with UNETLoader.

text encoder   qwen_3_4b.safetensors
VAE            ae.safetensors
sampler        euler / simple, 8 steps, cfg 1.0

A quantized qwen_3_4b text encoder is published separately at JoaoZaokk/Qwen3-4B-W4A4-ConvRot — 7.49 GiB → 2.4 GiB. Note that stock ComfyUI runs every text encoder with dequantized math behind two independent locks, so a quantized encoder saves VRAM and no time unless those are released.

---

What this does not cover

• Six prompts, two seeds, one card, one resolution, no perceptual metric. "36 of 36 usable" is the judgement of someone who looked at the contact sheets.
• No image-space metric. Latent divergence measures trajectory, not fidelity — a free-running generated image cannot separate two quantizations of the same model, only a matched-input probe can, and that probe was not run on these builds.
• The calibration used one concatenated prompt block, not six separate ones: the tool that captures activations read a whole prompt file as a single prompt, a defect found and fixed the same day these were built. The published beyond-reality-zimage-v2 numbers this family is compared against were produced the same way, so the comparison is internally consistent, and a fresh calibration would shift the medians by a few percent — the calibration seed alone already moves them 2-6%.
• Nothing here was tested with LoRAs, ControlNet, or img2img.