Z-Image
Qwen3 text encoder with a flow-matching transformer using 3D RoPE positional encoding. Excellent quality at just 9 steps.
- Developer: Z-Potentials
- License: Apache 2.0
- HuggingFace: Z-Potentials/Z-Image-v1-Turbo
Variants
| Model | Steps | Size | Notes |
|---|---|---|---|
z-image-turbo:q8 | 9 | 6.6 GB | Fast, great |
z-image-turbo:q6 | 9 | 5.3 GB | Best quality/size |
z-image-turbo:q4 | 9 | 3.8 GB | Lighter |
z-image-turbo:bf16 | 9 | 12.2 GB | Full precision |
Defaults
- Resolution: 1024x1024
- Guidance: 0.0
- Steps: 9
Recommended Dimensions
| Width | Height | Aspect Ratio |
|---|---|---|
| 1024 | 1024 | 1:1 (native) |
| 1024 | 768 | 4:3 |
| 768 | 1024 | 3:4 |
Using non-recommended dimensions will trigger a warning. All values must be multiples of 16.
Example
Z-Image Turbo — 9 steps, seed 777:
mold run z-image-turbo:q8 \
"An astronaut floating through a bioluminescent underwater cave, \
reflections on the helmet visor, science fiction art" \
--seed 777
Notes
Z-Image uses a Qwen3 text encoder (BF16 or GGUF with auto-fallback). The quantized transformer is implemented directly in mold (not upstream candle) due to GGUF tensor naming differences.
On CUDA, the quantized transformer's linears dequantize each weight per forward rather than feeding candle's quantized fast-matmul kernels, which return non-finite values for Z-Image's layers and produced solid-black renders. Metal keeps the fast quantized path, which is validated against stable-diffusion.cpp. MOLD_ZIMAGE_QMATMUL=1 re-enables the CUDA fast path for kernel debugging only.
GGUF transformers always stay quantized at rest. Mold does not expose the old dense-map diagnostic route: expanding a Q4 checkpoint from roughly 3.4 GB to about 12 GB on CUDA or 24 GB on Metal was never memory-planned, its speed advantage was unmeasured, and the Metal route produced corrupted renders under that pressure (#1109).
Catalog cv:* Z-Image checkpoints use the hidden z-image-te companion for the same Qwen3 text encoder shards, tokenizer, and VAE. When the Civitai version publishes its own text-encoder file, that per-version file is downloaded and used instead of the shared encoder shards. An existing z-image-turbo install satisfies the shared files.