Hunyuan Image 3.0 Instruct Edit is an instruction, not a generate

hunyuan-image-3-0-instruct-edit is a leftover published row: edit-image / image-to-image, requires a still, 9 credits. Natural-language instruction in, structure of the source out. A prompt-only call will not start. It is not Hunyuan video and not a text-to-image generate.

The slug is still live in the published catalog. That is the whole compliment. A leftover row is allowed to be good at one job. The mistake is inferring a Hunyuan studio from the family name, or treating an instruct pass as a blank-page sample.

A leftover row, not a Hunyuan video

HunyuanImage 3.0 Instruct Edit is Tencent’s instruction-following image editing model. Catalog copy: modify images based on natural language instructions while preserving original content and structure. Nine credits. Categories: edit-image and image-to-image. requires_image is true. Audio is empty. Durations: none. Max output resolution: none. Discounted: false. accepts_video_input is false.

There is no Hunyuan video slug in the published catalog this page is allowed to link. Do not invent one. Do not open this row hoping a still becomes a clip. Motion is a different family, after the edit is signed. There is no Hunyuan provider hub on this site. The model page is the roster. Prompting for this slug: HunyuanImage 3.0 Instruct Edit prompting. The earlier lockup note is change one locked still (9cr). This page is only the leftover-row rule: instruction, not generate.

Nine credits is the catalog sticker, billed per megapixel on a single default option, min 9, max 9. Spend it on a change. Do not spend it on “try again” energy. Trying again from a blank prompt is a text-to-image row, and it will not preserve the structure this model exists to preserve.

Instruction in, structure out

"Make a better version of this campaign" is a generate. "Remove the paper cup, keep the rest" is an instruct edit. HunyuanImage 3.0 Instruct Edit is sold as the second thing. The structure of the photograph is the product. Faces, layout, type, the crop you already signed — those stay unless you name them.

The door is edit a photo with AI: a named agent job on a file you already approved. The tool shape is the AI photo editor. Neither of those is text-to-image. Text-to-image is how you start a picture you do not have. Instruct Edit is how you keep a picture you do.

max_reference is 3. Aspect ratios: 16:9, 4:3, 1:1, 3:4, 9:16, auto. Match the approved still. An edit that silently changes 9:16 into 1:1 is a new picture wearing the old filename. No resolution ceiling is listed. Do not promise 4K. Judge the output against the file you sent. If the still is soft, upscale before the instruct pass, or accept a soft edit.

One object per run is the honest batch. Three unrelated changes in one sentence is how structure slides. Name the object and the change. “Change the jacket to navy, do not touch the face.” Not “fashion editorial, same vibe.” Vibe is a generate. Navy is an edit.

A prompt-only call will not start

requires_image is true. Categories do not include text-to-image. A sentence with no upload is not a job on this slug. The call will not start. That is not a bug. That is the leftover row doing what the flags say.

If you do not have the locked file, you are in the wrong aisle. Generate a still elsewhere — text-to-image is that aisle — then come back with the file you would actually print. Instructing a picture that does not exist is a generate wearing an edit name. The agent can route you to this row only when a still is in the job. A prompt-only brief belongs on a text-to-image model. Pin this for the change. Pin a generate row for the first picture.

When the edit is signed, then you may buy motion on an image-to-video row. Animating the pre-edit file because the edit “mostly worked” is how the cup you removed comes back as a blur. Nine credits to keep the photograph. A new text-to-image to throw it away. The second option is always available and usually wrong once a human has said yes.

Not a text-to-image generate

Text-to-image samples a new picture from a sentence. Instruct Edit takes a picture and a sentence and is supposed to return the same picture with one named change. They fail differently. A generate that “almost” matches the approved still is a cousin. An edit that drifted the crop is a failed instruction. Do not debug one as the other.

This leftover row is not Hunyuan video, not a text-to-image generate, not a 4-credit sketch pad, not a resolution upgrade, and not a soundtrack. Audio is empty. It is an image row. Use it when a human has already said yes to the file, and the note is a change. Do not use it because the word Hunyuan is in the name. Family names do not add a generate that the categories do not list.

The cheap test: if the sentence starts with “change,” “remove,” “keep this and…,” you are on Instruct Edit and you already have the still. If it starts with “a photo of,” you are on text-to-image. If it starts with “a clip of,” you left the leftover row entirely.

FAQ

Can I use this row to make a still from a sentence?

Not as this slug is specced. requires_image is true and the categories are edit-image and image-to-image. A prompt-only call will not start. Generate elsewhere, then instruct.

Why 9 credits if I already have the photo?

Because this is the instruct-edit line, billed at 9 credits in the catalog, not discounted. You are paying to keep structure, not to sample a new picture. A 4-credit generate throws the structure away.

Is this Hunyuan video?

No. The published row is an image edit. There is no Hunyuan video slug in the live catalog this page can link. After the still is signed, pick a video family on a video row. Do not wait for this leftover slug to grow a duration list.

What should the instruction look like?

Name the object and the change. “Remove the cup, keep the table and the type.” Not “same campaign, better.” Better is a generate. The cup is an edit. If you needed a new picture, you needed text-to-image, not this row.