Hi Z-Image team,
I built crispz-studio, an open-source desktop UI running entirely on Z-Image Turbo
through diffusers (ZImagePipeline). Sharing it here for users who want a local,
single-purpose tool without a node graph.
Repo: https://github.com/mikecastrodemaria/crispz-studio
What it does with Z-Image:
- txt2img with the Turbo profile (8 steps, guidance 0), plus profiles for Base
checkpoints (CFG ~4, more steps). Euler and UniPC samplers, sgm_uniform / beta /
karras / exponential sigma schedules.
- Upscale + detail: Real-ESRGAN followed by a Z-Image img2img refine pass, with
tiling for 4K+ outputs on a single GPU.
- Inpaint / outpaint / reframe bounded to the model's ~1 MP sweet spot, with
feathered seams and an optional harmonize pass.
- Community checkpoints: single-file
.safetensors from Civitai load as a
transformer override while the base repo keeps the VAE and Qwen3 text encoder.
Multi-LoRA (up to 10 slots), X/Y/Z grids, a job queue for unattended batches.
- 100% local, Gradio UI + CLI. Runs well on a single consumer GPU (developed on an
RTX 5090, BF16).
Z-Image Turbo is the default model and the reason this tool exists: 8-step
generations at this quality level made a lightweight local studio practical.
If this kind of community project is worth a mention in the README, I would be glad
to see it listed. Feedback on how I use the pipeline is welcome too, and I will close
the issue if showcase posts do not belong in the tracker.
Hi Z-Image team,
I built crispz-studio, an open-source desktop UI running entirely on Z-Image Turbo
through diffusers (
ZImagePipeline). Sharing it here for users who want a local,single-purpose tool without a node graph.
Repo: https://github.com/mikecastrodemaria/crispz-studio
What it does with Z-Image:
checkpoints (CFG ~4, more steps). Euler and UniPC samplers, sgm_uniform / beta /
karras / exponential sigma schedules.
tiling for 4K+ outputs on a single GPU.
feathered seams and an optional harmonize pass.
.safetensorsfrom Civitai load as atransformer override while the base repo keeps the VAE and Qwen3 text encoder.
Multi-LoRA (up to 10 slots), X/Y/Z grids, a job queue for unattended batches.
RTX 5090, BF16).
Z-Image Turbo is the default model and the reason this tool exists: 8-step
generations at this quality level made a lightweight local studio practical.
If this kind of community project is worth a mention in the README, I would be glad
to see it listed. Feedback on how I use the pipeline is welcome too, and I will close
the issue if showcase posts do not belong in the tracker.