Z-Image Turbo ControlNet Union 2.1

Base model

• Quantized from: alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.1

• ControlNet revision: 5155fc56d17821007d6f62ac192c09e0f0e72016

• Turbo revision: Tongyi-MAI/Z-Image-Turbo @ f332072aa78be7aecdf3ee76d5c247082da564a6

• Licence of the weights: Apache-2.0

• Card date: 2026-10-11

Method

Dtype conversion only. Linear weights are stored as float8_e4m3fn with a per-tensor BF16 scale (amax/448) under the safetensors metadata format voxprint-fp8-v1. Embeddings, norms, and the layers named in the job stay BF16. This is not an ablation and it is not a fine-tune. Quality checks dequantize FP8 back to BF16, because the A100 has no FP8 tensor cores.

Calibration data

None. No calibration set and no training.

Quantization

• Union 2.1 full: BF16 (24 GB tier)

• Union 2.1 lite: publish fp8 (16 GB tier when FP8 passes)

• Norms, embeddings, and zero-convs stay BF16

Evaluation

• lite FP8 vs BF16 mean LPIPS 0.07501369922969249 (gate 0.10), decision fp8

• 16 prompts x canny/depth/pose x 2 seeds, plus 8 inpaint cases

• Control maps are generated in the job from synthetic scenes

Known limits

The FP8 files are a storage format. On an A100 they are dequantized to BF16 before the matmul. A component that misses its LPIPS or CLIP gate is published as BF16. This card is not a safety filter. TensorRT engines are not part of this job.

Acceptable use

Use of this model is subject to the VOXPrint Acceptable Use Policy, in addition to the upstream licence.

Credits

• alibaba-pai Z-Image-Turbo-Fun-Controlnet-Union-2.1

• Tongyi-MAI / Alibaba Z-Image-Turbo

• diffusers ZImageControlNetPipeline

Modification notice

Modified by Voxprint: dtype conversion.

License

Apache-2.0, same as the base model. See https://www.apache.org/licenses/LICENSE-2.0

Made for Voxprint AI Media Suite

Voxprint AI Canvas loads these weights for text-to-image and image-to-image. Tiers are selected with the VRAM hints in MODELS.md.

Author's original files (unmodified)

These files are byte-identical copies of alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.1 @ 5155fc56d17821007d6f62ac192c09e0f0e72016 (Apache-2.0), kept at the upstream file names. They are what VOXPrint AI Canvas loads by default.

zimage-turbo-controlnet-union-2.1-bf16.safetensors is the same bytes as Z-Image-Turbo-Fun-Controlnet-Union-2.1-2602-8steps.safetensors under an older name. zimage-turbo-controlnet-union-2.1-lite-fp8.safetensors is a VOXPrint dtype conversion (see Method above).