Z-Image Turbo — NVFP4 pack for ComfyUI
An NVFP4-quantized build of Z-Image Turbo plus everything ComfyUI needs to run it, packaged for Blackwell GPUs (RTX 5090, DGX Spark / GB10) where NVFP4 runs natively.
The full writeup — how it was quantized and how it actually performs — is at ai-muninn.com.
What's in the pack
| File | Role | Size |
|---|---|---|
z_image_turbo_nvfp4.safetensors |
Z-Image Turbo diffusion model — NVFP4 | 4.5 GB |
ae.safetensors |
VAE (autoencoder) | 335 MB |
qwen_3_4b.safetensors |
Qwen3-4B text encoder (bf16) | 8.0 GB |
qwen_3_4b_fp8_mixed.safetensors |
Qwen3-4B text encoder (FP8 mixed) | 5.6 GB |
gemma_3_12B_it_fp4_mixed.safetensors |
Gemma-3-12B-IT text encoder (FP4 mixed) | 9.4 GB |
Z-Image-Turbo-Fun-Controlnet-Union.safetensors |
ControlNet Union | 3.1 GB |
Why NVFP4
NVFP4 is Blackwell's native 4-bit format. On a 5090 / GB10 it roughly halves the diffusion model's VRAM versus fp8/bf16 while keeping the tensor cores fed — so Z-Image Turbo's few-step generation stays fast without spilling to system RAM. FP8 / FP4 text-encoder variants are included so you can trade quality vs. footprint on the encoder side too.
Use in ComfyUI
Drop the files into your ComfyUI model folders:
z_image_turbo_nvfp4.safetensors→models/diffusion_models/ae.safetensors→models/vae/qwen_3_4b*.safetensors/gemma_3_12B_it_fp4_mixed.safetensors→models/text_encoders/Z-Image-Turbo-Fun-Controlnet-Union.safetensors→models/controlnet/
Credits & licensing
- Base model —
Tongyi-MAI/Z-Image-Turbo· Apache-2.0 - ControlNet —
alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union· Apache-2.0 - Text encoders — Qwen3-4B (Apache-2.0). The bundled
gemma_3_12B_it_fp4_mixedfile is Google Gemma-3, covered by the Gemma Terms of Use, not Apache-2.0 — use that file under those terms.
NVFP4 quantization & packaging by ai-muninn. This repackaging is Apache-2.0; the Gemma component keeps its own license as noted above.
Made by Thor Lin. Full benchmarks, setup notes and the runs that failed are on ai-muninn.com. More quantized models: huggingface.co/coolthor.
Provenance
This section documents release 85d8426976b4cf8e92c2e037bd153e9ccfc6a8fc, inspected on 2026-11-14. Where the exact upstream commit used at build time is unknown, it is listed as not recorded rather than backfilled with today's upstream HEAD.
Source lineage
| Source repository | Revision used for this build | Role |
|---|---|---|
| Tongyi-MAI/Z-Image-Turbo | not recorded | Documented diffusion-model lineage |
| alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union | not recorded | Bundled ControlNet lineage |
Build details
Tools and versions: NVFP4 conversion and mixed FP8/FP4 encoder packaging; build-time converter identity/version and commands not recorded.
This repository is a multi-component Z-Image pack despite its name: diffusion model, VAE, Qwen3-4B BF16 and FP8 mixed encoders, Gemma-3-12B FP4 mixed encoder and ControlNet Union. Exact immediate source repositories for the VAE and encoder files were not recovered; architecture names alone do not establish an immutable input lineage.
The recovered records describe quantization and repackaging, with no additional fine-tuning or merge by coolthor recorded. ControlNet is an upstream trained component; the pack is not solely the original Z-Image weights. Provenance of each bundled encoder/VAE is incomplete.
Exact build command: not recorded in the recovered material. The method description above is not a complete executable reproduction recipe.
Weight file SHA-256 checksums
The values below are Hugging Face LFS object SHA-256 values for the pinned release above. File sizes and digests agree with pinned tree metadata. They identify full file contents, not LFS pointer text or Git blob IDs. Calibration imatrices are included when present.
| File | Bytes | SHA-256 |
|---|---|---|
Z-Image-Turbo-Fun-Controlnet-Union.safetensors |
3101572408 | 86c085c0d7853f12ce5183499934b54d08371c60f549c5a6b20615cd23989388 |
ae.safetensors |
335304388 | afc8e28272cd15db3919bacdb6918ce9c1ed22e96cb12c4d5ed0fba823529e38 |
gemma_3_12B_it_fp4_mixed.safetensors |
9447702218 | aaca463d11e6d8d2a4bdb0d6299214c15ef78a3f73e0ef8113d5a9d0219b3f6d |
qwen_3_4b.safetensors |
8044982048 | 6c671498573ac2f7a5501502ccce8d2b08ea6ca2f661c458e708f36b36edfc5a |
qwen_3_4b_fp8_mixed.safetensors |
5631994051 | 72450b19758172c5a7273cf7de729d1c17e7f434a104a00167624cba94f68f15 |
z_image_turbo_nvfp4.safetensors |
4509509600 | a553c889dbcb910de4c98293237573219a37007c1074a3f04576646a088bd5c8 |
Verifying the files
Download the desired files from revision 85d8426976b4cf8e92c2e037bd153e9ccfc6a8fc and compare sha256sum <file> (Linux) or shasum -a 256 <file> (macOS) with the table. If cloning with Git LFS, hash the downloaded weight files, not the LFS pointer files. A matching hash confirms your files match this release; it does not prove the weights came only from the listed base model with no other changes.
Checking how the weights were derived requires the exact upstream commit, tool versions, recipe and any calibration or training data, then re-running the build and diffing the weights with a format-aware tool (or comparing file hashes where serialization is deterministic). Because of the gaps listed below, this build can't be reproduced exactly today. Quantized files are expected to differ numerically from BF16 sources, and a raw file hash cannot establish a base-to-quant weight diff. No quantization rerun or weight-diff audit was performed for this provenance review.
Known gaps: Build-time upstream commits for every component; Exact VAE/encoder source repositories and per-file conversion/copy mapping; Tool identities/versions and commands.
Sources: existing model card and HF metadata.