If you haven't tried Z-Image Turbo yet, you're missing one of the most exciting local models of the year — and this is the tool that makes it feel like Fooocus.
Why Z-Image Turbo deserves your attention
Z-Image Turbo (Tongyi-MAI / Alibaba) is a distilled text-to-image model that produces sharp, coherent, beautifully lit images in 8 steps at guidance 0 — a full render in a couple of seconds on a modern GPU. Skin texture, hands, typography-free scenes, natural lighting: for its ~1 MP sweet spot it trades punches with models many times slower.
The base model isn't listed on Civitai yet (it lives on Hugging Face), but community fine-tunes and single-file checkpoints are already appearing right here — Juggernaut-Z and friends — and they're worth the detour. You just need a proper studio to run them in.
The approach: what we learned from Fooocus
I run a maintained fork of Fooocus (Fooocus2026), and its philosophy shaped this tool from day one: opinionated defaults, one big Generate button, and power hidden behind an "Advanced" checkbox — not a node graph. No ComfyUI, no SwarmUI, no wiring. You type a prompt, you get an image; every advanced feature is one click away when you want it, invisible when you don't.
crispz-studio applies that recipe to Z-Image, then goes further: several features we shipped on Fooocus2026 (job queue, X/Y/Z grids, tag autocomplete, atomic downloads) were ported and adapted the same week. Everything is config-gated, enabled by default, and removable in one line — zero cost when off.
What's inside (v1.5)
Create
Text → Image with an optional "Upscale after generate" toggle — each image chains through the upscale pipeline automatically, no manual step.
Tag autocomplete in the prompt fields: danbooru-style suggestions under the caret (↑/↓, Tab/Enter), merged with your local
__wildcards__, popularity-ranked, sub-millisecond per keystroke. Drop any CSV intags/to add a source.277 Fooocus/SDXL styles with search and hover previews.
Enhance
Upscaler/detailer: Real-ESRGAN + Z-Image refine with 4K tiling.
Inpaint / Outpaint in one tab, three modes: brush inpaint, directional expand (left/right/top/bottom/center, Fooocus-style), and reframe to a new aspect ratio — with blurred-edge fill, feathered seams, an optional Harmonize pass, and Auto-describe (local captioner, no server needed) so empty-prompt outpaints stay coherent.
Remove background (rembg) and face swap with GFPGAN restore.
Compare & batch
Job queue: snapshot ALL current settings (including model, LoRAs, sampler) into labeled jobs, reorder them, run them overnight. Stop pauses; nothing is lost. VRAM is purged only when the model actually changes between jobs.
X/Y/Z grid: vary up to 3 parameters (checkpoint, sampler, steps, guidance, denoise, LoRA weight, Prompt S/R…) and get an annotated contact sheet per Z value — from the UI or the CLI (
--xyz "Steps=4,8,12"), Ctrl+C assembles a partial sheet.
Civitai users, this part is for you
One checkpoint dropdown merging the official base repos with your single-file
.safetensorsfrom Civitai (two folders supported). BF16/FP16 checkpoints load as transformer overrides — VAE and text encoder stay from the base, so incomplete community repos still work. Multi-LoRA (3 slots, trigger-word helper). FP8 files are detected and skipped with a clear message.
Quality of life
Asset Browser: instant-open output gallery with full metadata (prompt/seed/params), search, day filter, NSFW blur.
Full metadata embedded in every image (PNG text + EXIF + JSON sidecar).
Scriptable CLI and a persistent server mode — every feature works headless.
Optional Ollama integration (describe an image into a prompt, improve prompts, blend reference images) with offline fallbacks for everything.
100% local, nothing phones home
Python + Gradio + diffusers. Your images, your models, your machine. Developed and tested on an RTX 5090 (CUDA 12.8); the model itself is comfortable on much less, and CPU-offload modes are built in for smaller cards.
Get it
👉 GitHub: github.com/mikecastrodemaria/crispz-studio
Clone, run install.bat / install.sh, then run.bat — the README covers everything (there's a CLI cheat sheet too). Issues and feedback welcome on GitHub.
Happy rendering!