Tongyi-MAI

/

Z-Image

like 1.11k

Follow

Tongyi-MAI 5.26k

I'm getting rainbow/garbage output when running the official Z-Image example with diffusers on an RTX 5090 (Blackwell), Windows. Same code from the model card works in terms of running, but the decoded image is wrong.

Setup

• GPU: NVIDIA GeForce RTX 5090
• OS: Windows
• Python: 3.12, in a venv (not global)
• PyTorch: 2.7.1+cu128
• diffusers: 0.37.0.dev0

What I'm running

The exact example from the model card (prompt, ZImagePipeline.from_pretrained("Tongyi-MAI/Z-Image", torch_dtype=torch.bfloat16), then pipe(...) with num_inference_steps=30, guidance_scale=4, cfg_normalization=False, height=512, width=512). Image saves without error but looks like rainbow/noise.

What I've already tried

• Triton / Sage Attention: Not installed in this env. importlib.util.find_spec("triton") and "sageattention" are both None. So the pipeline is using default PyTorch attention.
• TORCHAO_DISABLE_INTMM=1: Set before running; no change. Still get torchao.kernel.intmm warning and rainbow output.
• fp32: With torch_dtype=torch.float32 the output is the same rainbow junk.
• With a step callback: NaNs first appear at denoising step 0 (first transformer forward). So the issue is the Z-Image transformer on RTX 5090, not the VAE or quantization.
• Z-Image Turbo (local, same 5090 setup) runs fine; the issue appears specific to the base Z-Image model.

Question

Is Z-Image (base) with the official diffusers pipeline known to work on RTX 5090 / Blackwell? If yes, what environment (PyTorch/diffusers versions, dtype, any env vars or flags) are you using? If there’s a known workaround (e.g. disable a backend, use a specific diffusers version), I’d appreciate a pointer.

Thanks.