Z-Image De-Turbo — ConvRot 4-bit for ComfyUI, 3.75x lighter and 2.62x faster per step

Native convrot_w4a4 and mixed 4/8-bit builds of ostris/Z-Image-De-Turbo, the de-distilled fine-tune of Z-Image Turbo that people train LoRAs on. ComfyUI's own quantization format — stock UNETLoader, no custom node.

11.46 GiB → 3.06 GiB, and 0.911 → 0.348 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

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

err_bf16   median 0.0019     <- the floor. This much is not quantization.
err_w4a4   median 0.1211
err_w4a8   median 0.0384

Dispatch counted with the weights on the card: 170 quantized modules, 8/8 quantized forwards, 0 dequantize, convrot_linear_dtype=int4 on comfy_kitchen.backends.cuda.

---

Take W4A4 here. The mixed build does not earn its extra bits on this checkpoint.

paired against zimage_deturbo_w4a4, run by run
zimage_deturbo_mixed    mean delta +0.0199    wins 7/12

Seven of twelve is a coin flip, and the tool that produced it prints Split decisions, i.e. not separated at 12 run(s) rather than a winner. So the mixed build costs 0.11 GiB and 14% more time per step for no measurable gain here.

That is the interesting part, because the same recipe does earn its keep on the parent model. Run on the official Tongyi-MAI/Z-Image-Turbo at the identical threshold, the mixed build wins 11 of 12 paired runs. Same converter, same --promote-error 0.15, same six prompts, same two seeds, same card — a model and its own de-distilled fine-tune disagree about whether promoting ~55 layers to 8 bits buys anything.

Nothing about the per-layer error predicted that: the two checkpoints measure 0.1228 and 0.1211, within 1.4% of each other. Whatever separates them is not in that number.

One more reason not to read too much into the means: the between-run spread inside a single arm is 0.37–0.42 on this checkpoint, which is larger than the 0.02 gap between the arm means. Always read the paired row.

---

Using them

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

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

De-Turbo is not distilled — it exists precisely to undo the turbo distillation so LoRA training behaves. The 8-step / cfg 1.0 settings above are what these measurements used, for comparability with the Turbo builds; for actual generation this model wants more steps and real cfg.

A quantized qwen_3_4b text encoder is published separately at JoaoZaokk/Qwen3-4B-W4A4-ConvRot.

---

What this does not cover

• Six prompts, two seeds, one card, one resolution, no perceptual metric. "36 of 36 usable" is a human judgement over contact sheets.
• LoRA training is this model's whole reason to exist and was not tested. Nothing here says whether a LoRA trains, or trains as well, on 4-bit weights.
• Sampled at 8 steps / cfg 1.0, which is the Turbo regime and not this model's own.
• No image-space metric. Latent divergence measures trajectory, not fidelity.
• The calibration used one concatenated prompt block, not six separate ones — a defect in the activation-capture tool, found and fixed the same day these were built. It is the same method the Turbo builds and the published beyond-reality-zimage-v2 used, so the family stays internally comparable; a fresh calibration would move the medians a few percent.