Z-Image Edit: Why the Weights Still Aren't Published, and How to Edit Images With Z-Image Today (2026)

I lost two hours last week to a .safetensors file that called itself Z-Image Edit. It downloaded fine, dropped into models/diffusion_models/, and then ComfyUI either ignored it or threw a key-mismatch error depending on which loader I used. No crash, no useful message. Just a file that did not belong to the architecture it claimed.

That is the state of "Z-Image Edit" in November 2026. The model exists as an announced variant in Alibaba Tongyi's own model zoo, the editing quality is described there in detail, and the weights are still listed as to be released. Everything else — the download pages, the mirrors, the ComfyUI tutorials, the "how to download z-image-edit" videos — is either anticipation of that release or something else wearing its name.

This guide is what I worked out the hard way: what Z-Image Edit actually is per the official model card, how to check its status in under a minute, what file you are really getting when a site offers it today, and the editing workflows that produce usable results right now with the Z-Image models that are published.

TL;DR

• Z-Image Edit is not downloadable yet. The official Tongyi-MAI model zoo lists it with pre-training and SFT complete, RL not applied, 50 sampling steps, CFG enabled, task = editing — and both Hugging Face and ModelScope columns read To be released.
• The official Hugging Face org has exactly two public repos: Z-Image-Turbo and Z-Image. There is no Z-Image-Edit repository to clone.
• Files circulating as "z-image-edit" are one of three things: a community upload of unclear provenance, another architecture's model under a renamed file, or a LoRA/merge trained on the base. None of them load into a Z-Image pipeline as an editing model.
• The failure signature is quiet. A wrong-architecture checkpoint usually produces a key-mismatch error or a black frame, not a clear "this is not an edit model" message.
• Editing today runs through four routes: Qwen-Image-Edit for instruction-based edits, Z-Image Turbo plus a community IPAdapter for style transfer, Z-Image Turbo reference workflows for consistency, and hosted generators when you do not want a graph at all.
• Z-Image Turbo is the 6B, 8-step, no-CFG model — fast, Apache 2.0, and the correct base for anything you build before the edit weights ship.
• If you want to skip the graph while testing ideas, the free Z-Image generator runs the model in the browser.

Quick Verdict: What You Can Do Today vs What Waits on the Weights

Job	Route that works now	What changes when Z-Image Edit ships
Change a background, keep the product	Qwen-Image-Edit (dedicated editing model)	One more option at Z-Image's speed
Replace a material or colour on a product	Qwen-Image-Edit, or masked edit in any editor	Faster iteration at 50 steps
Apply a reference photo's style to new content	Z-Image Turbo + community IPAdapter	Possibly a first-party path
Keep a character consistent across images	Z-Image Turbo reference workflow	Same job, fewer workarounds
Batch 200 clean generations	Z-Image Turbo at 8 steps	No change — generation is already solved
Precise pixel retouching	A normal image editor	No change

The honest summary: Z-Image Edit is a convenience release for people already inside the Z-Image ecosystem. If your bottleneck is editing existing images, the tool that solves it today is a dedicated editing model, not the promise of a variant.

What Z-Image Edit Is Supposed to Be

The model zoo on the official Z-Image-Turbo card is unusually informative for an unreleased model:

Property	Value
Pre-training	Complete
SFT	Complete
RL	Not applied
Sampling steps	50
CFG	Enabled
Task	Editing
Visual quality	High
Diversity	Medium
Fine-tunability	Easy
Weights (HF / ModelScope)	To be released / To be released

Read those rows as an architecture description rather than a release note. The model is Z-Image's 6B single-stream diffusion transformer, fine-tuned specifically on editing instructions, at the base model's 50-step / CFG-on regime — not Turbo's 8-step, no-CFG regime. That difference matters practically: if you are used to Turbo's near-instant results, the edit model will feel slower per image, in exchange for instruction following that Turbo was never trained for.

The card also notes bilingual editing instructions as a highlight, which is consistent with everything else in the Z-Image line: the text encoder is a Qwen 3 4B model, the same family that handles Chinese and English prompts without a separate translation step.

Verify the Status Yourself in 60 Seconds

Do not take a blog post's word for it — including this one. Four checks, all free:

#	Check	What a real release looks like
1	Hugging Face org page for Tongyi-MAI	A repository named Z-Image-Edit (currently only Z-Image-Turbo and Z-Image exist)
2	The model zoo table on the Z-Image-Turbo card	The HF and ModelScope columns change from To be released to links
3	ModelScope mirror	Same repository, same day — the two usually publish together
4	ComfyUI's supported-model list / template workflows	A Z-Image Edit template appears

Check 1 is the fastest filter. The official organisation's repo list is short, public, and cannot be faked: a release exists there or it does not. Community uploads with "z-image-edit" in the name live under individual user accounts, not under the vendor's org — which is exactly the tell.

If you want output today, start here: Launch Z-Image Now →

What You're Actually Downloading When a Page Offers "Z-Image Edit"

What the file is	How to recognise it	What happens when you load it
A community upload of unclear provenance	Individual user account, few downloads, vague card, no eval	Loader errors, or output that looks like a different model
Another architecture's model, renamed	Card mentions a different family, or the file size is wrong for a 6B DiT	Key mismatch, black frames, or noise
A LoRA or merge built on Z-Image / Turbo	Base model field names Z-Image or Z-Image-Turbo	Loads — but as an adapter, not an edit model
A quantised or FP8 build of the base	Filename contains fp8, gguf, q4	Loads — but it is generation, not editing

The size sanity check is worth internalising. Z-Image Turbo runs as a 6B model in three pieces, and the ComfyUI layout that works is:

ComfyUI/models/
├── diffusion_models/z_image_turbo_bf16.safetensors
├── text_encoders/qwen_3_4b.safetensors
└── vae/ae.safetensors

A "Z-Image Edit checkpoint" of a few hundred megabytes is not a 6B diffusion transformer. It is a LoRA, a pruned merge, or nothing at all. And a file that lands in the right folder but throws missing-key errors on load is telling you something specific: the weights were trained against a different architecture than the one you are loading them into.

Where Editing Actually Happens Today

Four routes, in order of how well they match an "edit this image" brief.

Route A — Qwen-Image-Edit (the dedicated editor)

If the job is modify this existing image and keep everything else, a model trained for editing is the correct tool, and Qwen-Image-Edit is the most direct one available. It handles background replacement, material swaps, face fixes and character consistency without asking you to describe the whole scene again. Apache 2.0, and it runs in the same ComfyUI you already have.

The trade is speed: it is not an 8-step model, and its outputs are less stylised than Z-Image's. For product and retouching work that is a feature.

Route B — Z-Image Turbo + community IPAdapter (style transfer only)

This is the one Z-Image-specific editing path that works today, and it does one thing: apply the look of a reference image to new content. There is no first-party adapter; the working node is community-built (Boyo Apply Z-Image IP-Adapter), and it patches Z-Image's transformer blocks with IP projection layers fed by a converted SD3 adapter.

Settings that behave, from the write-up I did on this route: adapter weight 0.7, start_percent 0.0, end_percent 0.6–1.0, fp16 on CUDA. Below the weight floor you get almost no effect; above it the reference overwhelms your prompt.

Two caveats worth stating plainly. First, the loader needs a converted SD3→Z-Image adapter file that the node's own README does not link, which is where most people stop. Second, it is style conditioning, not identity: it will not keep a face or a product exactly, and for those jobs you want Route C or A.

Route C — Z-Image Turbo reference workflow (consistency)

Generate a reference frame, then reuse it as the visual anchor for every subsequent generation, with the prompt carrying the invariants and the image carrying the details words cannot pin down. It is the same discipline that makes reference systems work in video models, applied to stills — one property per reference, stated explicitly.

This is also the workaround for the missing edit model: instead of editing an existing image, you regenerate it with a locked reference and a tighter prompt. It is slower than an edit, and it avoids the failure mode where an editor reinterprets the subject entirely.

Route D — Hosted generation, no graph at all

When you are testing whether a prompt is worth committing to, a local install is the wrong first step. The free Z-Image generator runs the same model family in the browser, and the free image-to-video generator takes the result straight into motion work. Iterate on the look there, then move to the graph when you need LoRAs or exact settings.

Instruction Prompts That Survive Any Editor

Whether you end up in Qwen-Image-Edit, a hosted editor, or Z-Image Edit the day it ships, the prompt structure that works is the same: operation + scope + preservation clause. Edits fail when the model is told what to change and left to guess what to keep.

Brief	Weak prompt	Prompt that holds
Background swap	"Put the mug on a beach"	"Replace only the background with a pale sand beach at golden hour. Keep the mug, its glaze, its shadow direction and the rim highlight unchanged."
Material change	"Make the watch gold"	"Change the strap material to brushed gold. Keep the dial, hands, indices and lighting identical."
Season change	"Make it winter"	"Add snow cover to the ground and frost on the window glass. Keep the building, camera position and light direction unchanged."
Style restyle	"Make it a watercolour"	"Restyle as loose watercolour with visible paper texture. Preserve subject pose, framing and composition exactly."

The preservation clause is not politeness. Editing models regenerate regions; naming what must not move is the only mechanism you have for constraining that regeneration. On 50-step models the clause is also where most of your quality comes from, because a vague scope invites the model to redraw more than you asked for.

The Limits You'll Hit on a 6B Model at 8 Steps (and at 50)

Limit	What it looks like	Workaround
Fine text in edited regions	Lettering warps or re-renders as plausible nonsense	Composite type in a normal editor after generation
Hands and small props	Fingers merge; thin objects gain or lose parts	Close crops, one subject per generation
Style range	Prompt-driven styles converge toward a photoreal default	Reference images, or a LoRA on an open base
Repeated identities	Faces drift between generations	Reference workflow, not prompt description
Extreme aspect ratios	Composition collapses at very wide or tall sizes	Generate at 1024-square and crop deliberately
Step-count wishful thinking	More steps, no more detail	Stay in the regime the model was trained for — 8 for Turbo, 50 for base

That last row is the most expensive habit in the Z-Image community. Turbo was trained for the low-step regime; raising the step count buys time and nothing else. Base and the eventual edit model sit at the other end, at 50 steps with CFG on — the regimes are not interchangeable, and a workflow copied from Turbo to base without changing both settings will look broken. For a closer look at how it stacks up against other models, see Krea 2 vs Qwen Image Edit vs Z.

Z-Image Turbo vs Z-Image vs Z-Image Edit: Settings at a Glance

	Z-Image Turbo	Z-Image (base)	Z-Image Edit
Status	Public	Public	Announced, weights pending
Parameters	6B	6B	6B (same architecture)
Sampling	8 steps, no CFG	50 steps, CFG on	50 steps, CFG on
Task	Text-to-image	Text-to-image	Instruction-based editing
Licence	Apache 2.0	Apache 2.0	Not yet stated on a released card
Best use	Volume generation, prompt iteration	Highest-fidelity stills from the family	Edits, once released

A Workflow for This Month: Generate → Edit → Animate

1. Draft in Turbo. 8 steps, no CFG, 1024px. Run four prompt variants in the time a single base-model render takes.
2. Promote the winner to base (or leave it in Turbo). If you need maximum fidelity, re-render the winning prompt on the 50-step base model with CFG on.
3. Edit with a dedicated editor. Qwen-Image-Edit for anything that must preserve the subject; state the preservation clause explicitly.
4. Lock consistency with references. When the same character or product appears in a second image, generate it with the approved frame as a reference rather than re-prompting from scratch.
5. Animate the still. Send the final frame into an image-to-video model — this is where the still pays off, and where consistency across the first frames decides whether the clips cut together.

If you want a hosted Wan-based second opinion on the animation step, Wan 2.7 image-to-video on Pollo is the comparison I would run with the same stills — different motion character, useful when one model keeps flattening a specific move.

The Bottom Line

Z-Image Edit is real as a model design and unreal as a download. The vendor's own model zoo says the weights are pending, the official organisation publishes two repositories today, and everything else carrying the name is a community file, a rename, or a different architecture. Nothing about that is a scandal — it is just a release schedule most tutorials have decided to ignore.

The practical move is to stop waiting and start from what is published. Z-Image Turbo covers generation at a speed that changes how you iterate. Z-Image covers fidelity when speed is not the constraint. Qwen-Image-Edit covers the editing job that the unreleased variant is meant to solve. When the weights do land, the workflow above does not change — one route simply gets shorter.

Try Z-Image Free in the Browser

No download, no graph, no key-mismatch errors. The published Z-Image models run in the browser, which is enough to settle whether the family's look fits your work before you install anything.

• Z-Image Turbo generation in the browser — 8-step speed, no GPU required
• Free credits to test prompts and reference ideas
• Take the result straight into image-to-video when the still is right
• The same model family as the local build, so settings you learn transfer

Run Z-Image free online and keep the local install for the jobs that actually need it — or start from the free Wan video generator if the still is already finished.

FAQ

Is Z-Image Edit released?

Not as of November 10, 2026. The official model zoo lists it with pre-training and SFT complete but both weight columns marked to be released, and the Tongyi-MAI Hugging Face organisation publishes only Z-Image-Turbo and Z-Image.

Can I download Z-Image Edit for ComfyUI?

There is nothing official to download. Files circulating under that name are community uploads of unclear provenance, other architectures renamed, or LoRAs and merges — which is why they either fail to load or behave like the base model.

How do I check whether Z-Image Edit is out?

Open the Tongyi-MAI organisation page on Hugging Face and look for a Z-Image-Edit repository. If it is not there, the weights are not published, regardless of what any download site claims.

What can I use instead right now?

Qwen-Image-Edit for instruction-based edits, Z-Image Turbo plus a community IPAdapter for style transfer, and Z-Image Turbo reference workflows for character and product consistency. All three run in ComfyUI today.

What settings does Z-Image Edit use?

Per the model zoo it is a 50-step, CFG-on editing model — the base Z-Image regime, not Turbo's 8-step, no-CFG regime. That will make it slower per image than Turbo.

Is Z-Image Edit free for commercial use?

Unknown until a released model card states a licence. Z-Image Turbo and Z-Image base are Apache 2.0, which is why building on them today is safe; the edit variant's terms cannot be confirmed from an unreleased card.

Why is my Z-Image Edit file not loading in ComfyUI?

Almost always an architecture mismatch. Place the components in diffusion_models, text_encoders and vae as the base model requires, and if keys still mismatch, the file was not trained against this architecture.

Related guides

• Z-Image Turbo FP8: running the 6B model on an 8GB GPU
• Z-Image IPAdapter in ComfyUI: what actually works
• Z-Image Turbo reference images: keeping a character consistent
• Krea 2 vs Qwen Image Edit vs Z-Image: which one edits

References

• Tongyi-MAI/Z-Image-Turbo on Hugging Face — official model zoo table (Z-Image Edit: 50 steps, CFG, editing task, weights to be released), 8-step Turbo spec
• Tongyi-MAI/Z-Image on Hugging Face — base model card, pipeline example at 50 steps and guidance 4
• Tongyi-MAI/Z-Image on GitHub — architecture notes and inference quick start
• Hugging Face model search: z-image-edit — community uploads under individual accounts, not the vendor org