Z-Image Base & Z-Image-Edit: What Actually Exists Beyond Turbo

Short answer: the "base" model is real and downloadable — it shipped January 27, 2026 under the plain name Tongyi-MAI/Z-Image (there is no repo called "Z-Image-Base"): 6B parameters, Apache 2.0, non-distilled, 28-50 steps at CFG 3.0-5.0, fine-tunable, with a 12.31GB BF16 checkpoint that runs in ComfyUI on a 16GB card and GGUF quants down to 4.01GB. Z-Image-Edit and Z-Image-Omni-Base are still marked "to be released" on the official repo as of August 2026 — no public Edit weights exist yet.

If you found us through the Z-Image Turbo ComfyUI guide, this page covers everything Turbo deliberately left out: real classifier-free guidance, negative prompts that actually work, seed diversity, LoRA training — and an honest, repo-verified status check on the Edit and Omni variants the family microsites keep hyping. Every number below was checked against the official Tongyi-MAI repositories in August 2026, because third-party Z-Image blogs contradict each other constantly.

---

The Z-Image Family Map {#family-map}

Four variants are announced; exactly two are downloadable today. Here is the family status straight from the official Tongyi-MAI/Z-Image GitHub repository's news log, checked August 2026:

Source: Tongyi-MAI/Z-Image GitHub README, news section and model zoo.

Two naming traps worth clearing up before you go downloading things:

• "Z-Image-Base" is a community nickname. The official Hugging Face repo is Tongyi-MAI/Z-Image — searching for a literal "Z-Image-Base" repo gets you nothing. Everything in this guide about "Base" refers to that checkpoint.
• The family is bigger on marketing sites than on Hugging Face. A whole ecosystem of lookalike Z-Image domains exists. The authoritative sources are exactly two: the Tongyi-MAI GitHub repo and the Tongyi-MAI org on Hugging Face/ModelScope. If a variant is not there, it is not out.

For pedigree: this is Alibaba's Tongyi Lab (the Qwen lineage). The repo notes Z-Image-Turbo ranked 8th overall — and #1 among open-source models — on the Artificial Analysis text-to-image leaderboard as of December 8, 2025, and the team published a technical report on arXiv on December 1, 2025. This is a serious model family, which is exactly why the unreleased variants attract so much premature coverage.

---

Base vs Turbo: When 28-50 Steps Beats 8 {#base-vs-turbo}

Download Turbo for iteration speed, Base for control and training. The official repo publishes a feature table that settles most arguments — here it is, verbatim in substance:

Source: Tongyi-MAI/Z-Image GitHub README and Hugging Face model card.

The speed cost is real, and you can do the arithmetic yourself. Turbo runs ~8 network evaluations per image with guidance off. Base runs 28-50 steps, and because CFG above 1.0 requires two forward passes per step (conditional + unconditional), that is 56-100 evaluations — roughly 7-12x the compute per image. In our Turbo testing an RTX 3090 produced a warm 1024px image in ~3-4 seconds; scale that and you should expect very roughly 25-50 seconds per Base image on the same card. That is an estimate from verified step counts, not a benchmark we have run — but it sets honest expectations: Base is an SDXL-cadence model, not an interactive one.

What you buy with that time:

• Prompt adherence you can tune. CFG 3.0-5.0 is a real dial. On Turbo, raising CFG just burns the image.
• Negative prompts that work. The repo strongly recommends them on Base; on Turbo they are decorative.
• Diversity. Turbo's distillation collapses variety — its repo rating is "Low", and different seeds look samey. Base's "Medium" makes batch-of-8 seed exploration worthwhile again. (The promised "High"-diversity member of the family is the still-unreleased Omni-Base.)
• One rating Turbo keeps: polish. The repo's own table scores Turbo's visual quality "Very High" against Base's "High" — Turbo went through an RL post-training stage that Base skipped. Base's case is control, diversity, and trainability, not extra prettiness per image; the repo is candid about that and so are we.
• The 2048px ceiling. The repo specifies 512×512 up to 2048×2048 total pixel area at any aspect ratio for Base.
• Trainability. The big one — Turbo is flagged not fine-tunable; Base is the family's training target (section below).

A practical pattern we like: keep both checkpoints, prototype prompts on Turbo at 8 steps, then re-render keepers on Base with a negative prompt and CFG ~4. Same text encoder, same VAE, one swapped file. Where does the result sit against FLUX and friends? Our best local image models compared roundup covers quality/speed/license across the field, and the FLUX local guide is the deep dive on Base's closest 12B rival.

---

VRAM and Checkpoints: What to Download {#vram}

The BF16 diffusion checkpoint is a 12.31GB file — plan for a 16GB card. An official-repack int8 build (6.20GB) and community GGUFs (4.01-7.22GB) cover 8-12GB cards. File sizes below were read directly from the Hugging Face repos in August 2026:

Plus the shared support files every build needs:

An attribution note, because we are careful with this: Tongyi's official "fits comfortably within 16G VRAM" statement was written about Turbo. Base is the same 6B S3-DiT architecture — the BF16 weights are the same size class — so the working-memory picture is equivalent; what Base costs you is time (more steps, doubled CFG passes), not VRAM. ComfyUI additionally offloads the text encoder after the prompt is encoded, which is how a 12.31GB DiT plus an 8GB encoder coexist on a 16GB card.

Practical tiers: a 24GB card like the RTX 4090 runs BF16 with room to spare — see what else a 4090 unlocks. A 16GB card runs BF16 comfortably with the FP8 text encoder. On 8-12GB, take the int8 build or a Q5/Q6 GGUF (the GGUF path needs the ComfyUI-GGUF custom node). Choosing a card for image work generally? Our best GPU for image generation pick ranks by text-to-image throughput rather than LLM specs.

---

ComfyUI Setup for Z-Image (Base) {#comfyui-setup}

Three files, three folders, one official template — image_z_image.json. ComfyUI documents Z-Image (Base) natively at docs.comfy.org, and the layout mirrors Turbo's with one swapped file. Update ComfyUI first; Z-Image support is recent enough that stale builds miss the template.

1. Download the three files (BF16 route, from the Comfy-Org/z_image repackage):

ComfyUI/models/
├── diffusion_models/
│   └── z_image_bf16.safetensors      ← 12.31 GB
├── text_encoders/
│   └── qwen_3_4b.safetensors         ← 8.04 GB
└── vae/
    └── ae.safetensors                ← 0.34 GB

If you already run Z-Image Turbo, you only need the first file — the text encoder and VAE are shared across the family, so adding Base costs one 12.31GB download. Direct downloads, using the verified Comfy-Org paths:

cd ComfyUI/models

curl -L -o diffusion_models/z_image_bf16.safetensors \
  https://huggingface.co/Comfy-Org/z_image/resolve/main/split_files/diffusion_models/z_image_bf16.safetensors

curl -L -o text_encoders/qwen_3_4b.safetensors \
  https://huggingface.co/Comfy-Org/z_image/resolve/main/split_files/text_encoders/qwen_3_4b.safetensors

curl -L -o vae/ae.safetensors \
  https://huggingface.co/Comfy-Org/z_image/resolve/main/split_files/vae/ae.safetensors

Prefer Python over nodes? The model card's route is diffusers from source — pip install git+https://github.com/huggingface/diffusers, then ZImagePipeline.from_pretrained("Tongyi-MAI/Z-Image", torch_dtype=torch.bfloat16).

2. Load the official template. Workflow → Browse Templates → Image → Z-Image (the template file is image_z_image.json). Point the three loader nodes at your files.

3. Set the sampling recipe. These ranges come from the official repo — treat them as the operating envelope, not suggestions:

4. Queue. Expect tens of seconds per image, not Turbo's 2-3 — that is the trade you chose. Note that ComfyUI's shipped template loads with its own conservative defaults (a step count in the mid-20s at CFG ~4) rather than the repo's recommended envelope — the official guidance above (28-50 steps, CFG 3.0-5.0, per the Tongyi-MAI model card and README) is what to use for final quality, so check the sampler node before you queue. Otherwise keep the sampler/scheduler as shipped; the distillation-era advice ("never touch steps/CFG") does not apply here. Base responds to these dials the way a normal diffusion model does.

New to the node graph entirely? Start with our complete ComfyUI guide — loaders, Manager, and how the sampler chain fits together — then come back.

---

Fine-Tuning: Why Base Is the One You Train {#fine-tuning}

Base is the first Z-Image you can meaningfully fine-tune: the official repo's model zoo rates its fine-tunability "Easy" (Turbo's entry is "N/A"), the model card cites LoRA training and ControlNet-style structural conditioning, and ostris AI Toolkit already lists Tongyi-MAI/Z-Image as a supported training target.

The reason is structural. Turbo is a distillation — its 8-step behavior is a carefully compressed trajectory, and naive training on top of it degrades the very thing that makes it fast. Base is a full-capacity, undistilled transformer (the model card's phrase), which is why Tongyi's own positioning for it is "community-driven fine-tuning and custom development."

What is verified about the tooling today:

• AI Toolkit (ostris) lists Z-Image, Z-Image-Turbo, and a community "Z-Image-De-Turbo" among its supported image models. It is the same trainer many people already use for FLUX and Qwen-Image LoRAs.
• A 6B DiT is a friendly training target. It is half of FLUX's 12B, which has made FLUX LoRA training on 24GB cards routine — but note that official VRAM requirements for Z-Image training have not been published, so any specific number you see is a community estimate, including that inference.
• The Apache 2.0 license carries through — you can train and ship commercial LoRAs without FLUX-dev-style license anxiety.

The craft of it — dataset curation, captioning, regularization, when a LoRA beats a full fine-tune — is model-agnostic, and our local image LoRA training guide covers that end of the pipeline.

---

Where Is Z-Image-Edit? {#z-image-edit}

Not out. As of August 2026 the official repo marks Z-Image-Edit "to be released" on both Hugging Face and ModelScope, and a Hugging Face search confirms no official Edit repo exists. If a website is offering "Z-Image Edit" today, whatever it is running, it is not the official open checkpoint — none has been published.

What the repo actually says about it: Z-Image-Edit is "fine-tuned on Z-Image specifically for image editing tasks" and will support creative image-to-image generation with instruction-following — precise edits from natural-language prompts. That places it in the same product category as Qwen-Image-Edit and FLUX Kontext-style editors: type "make the jacket red, keep everything else," get a targeted edit. No release date has been announced, and we would rather tell you that than invent one. Same status for Z-Image-Omni-Base: listed, undated, undetailed.

What you can do locally today instead:

• Classic img2img with Base. ComfyUI's standard image-to-image workflow (load an image, set denoise below 1.0) works with Z-Image like any diffusion model. It is restyling, not instruction editing — but it is available now and free.
• Qwen-Image-Edit, from the same Alibaba stable. For genuine instruction-following edits on local hardware, the Qwen team's image family is the released option — our Qwen-Image local guide covers it.
• Watch exactly one URL. The news section of the Tongyi-MAI/Z-Image GitHub repo is where both Turbo and Base were announced first. When Edit lands, it lands there — and given this family's cadence (Turbo in November, Base in January), "watch the repo" is not empty advice.

We will update this page when Edit ships with real weights, real settings, and real VRAM numbers — the same treatment Base gets above.

---

Limitations and Honest Caveats {#limitations}

• Base is slow by design. 28-50 steps with CFG means roughly 7-12x Turbo's compute per image. If you never use negative prompts, never train, and mostly iterate — Turbo remains the better daily driver.
• Our Base speed figure is an estimate. The 25-50s/image expectation is arithmetic from official step counts scaled off our own Turbo measurement, not a benchmark of Base itself. Treat it as a planning number.
• No official VRAM statement exists for Base. The 16GB guidance extrapolates from Turbo's official claim plus identical architecture and file size. It is a well-grounded extrapolation — but it is one, and we have labeled it.
• The ecosystem is young. Base is a January 2026 release; the LoRA/checkpoint ecosystem around SDXL and FLUX has a multi-year head start. Expect that gap to close (Apache 2.0 helps), but today it is real.
• The text encoder is a chunky dependency. Qwen 3 4B at 8.04GB full precision — budget disk for it, or take the FP8-mixed (5.63GB) variant.
• Naming confusion is the tax on this family. "Z-Image-Base" is not a repo, Edit is not out, and lookalike domains outnumber official sources. When in doubt: Tongyi-MAI on GitHub and Hugging Face, nothing else.

---

Verdict {#verdict}

1. Base is the real story of the Z-Image family right now. A 6B, Apache-2.0, non-distilled model with proper CFG, working negative prompts, better seed diversity than Turbo, and official fine-tuning support — running on the same 16GB-class hardware as Turbo, just slower.
2. Keep both checkpoints. They share the text encoder and VAE; the only cost of having both is 12.31GB of disk. Prototype on Turbo, finish on Base.
3. If you came to train, Base is your target — repo-flagged finetunable, AI Toolkit support already in place, commercial-friendly license.
4. Do not chase Z-Image-Edit yet. It is officially unreleased as of August 2026. For instruction editing today, use the released alternatives; for restyling, Base img2img already works.

The pattern with this family has been consistent: announced, then shipped within weeks, then repackaged for ComfyUI within days. Bookmark this page for the Edit update — and in the meantime, the model you can actually download is worth your 28 steps.

---

Sources {#sources}

• Tongyi-MAI/Z-Image (GitHub) — release news log, official sampling settings (steps/CFG/resolution), family feature table, Edit/Omni "to be released" status
• Tongyi-MAI/Z-Image (Hugging Face) — model card: non-distilled architecture, LoRA/structural-conditioning notes, download command
• Comfy-Org/z_image (Hugging Face) — repackaged checkpoint file names and sizes (BF16 12.31GB, int8 6.20GB, encoders, VAE)
• unsloth/Z-Image-GGUF (Hugging Face) — GGUF quant sizes (Q2_K 4.01GB through Q8_0 7.22GB)
• ComfyUI documentation — official native workflow, file locations, image_z_image.json template
• ostris/ai-toolkit (GitHub) — Z-Image listed among supported training targets

All repo facts and file sizes checked August 2026. Our RTX 3090 Turbo timing is our own measurement from the Turbo guide; the Base per-image time is an estimate derived from it.

---

FAQ {#faq}