HiDream O1: instruction vs generate

HiDream O1 is two published slugs, both 1 credit, both stills, both up to 2K in their descriptions. They are not a quality tier. HiDream O1 Image is a generate: pixel-native text-to-image that reasons before it draws, optional subject references, requires_image false. HiDream O1 Image Edit is an instruction: edit-image only, image required, instruction-driven and subject-driven changes on a supplied plate. Pick generate when there is no plate. Pick instruction when you can circle the object you are keeping. Using the cheap sticker as a reason to reroll identity is how 1 credit becomes a new asset wearing the old prompt.

HiDream O1 Image Dev is a third card: faster distilled generate, still 1 credit, still text-to-image, still up to 2K. It is not the edit row. This page is generate versus instruction, not full versus Dev.

Two slugs, one family

O1 Image: content type image, category text-to-image, no audio, no duration. Unified model, high-resolution stills up to 2K, optional subject references for personalization. You type a scene. You may attach a subject. You get a drawing.

O1 Image Edit: content type image, category edit-image only, image required, 1 credit, features named image-editing, reasoning, 2K, HiDream. You upload a still. You write a sentence that points. You get a change.

The shared 1-credit price is the trap. Teams treat both as “HiDream, a credit, whatever.” A generate resamples identity even when it is cheap. An edit is scoped to an instruction on a plate you are keeping. The sticker is an argument for editing more often, not for throwing the plate away more often.

Launchers: text to image for O1 Image (and Dev). AI photo editor for O1 Image Edit. Same family page in Slack is not the same form in the app.

Reasoning appears on both descriptions. On generate, it is whether the brief coheres before pixels land. On edit, it is whether the instruction coheres against the upload. Vague “make it premium” on edit lets the reasoner pick a premium and you will not recognise the still. Name the object. Name the change. Name what must not move.

Generate when there is no plate

Use O1 Image when you do not yet have a still you would keep.

Write a complete scene: subject, setting, light, mood, palette, composition. Optional subject references steer identity — a person or product that must recur — they do not replace the scene description. A reference without a prompt is not this product’s default path.

Aspects on the generate card include 1:1, 2:3, 3:2, 4:3, 3:4, 16:9, 9:16. Set the shipping ratio here. Many image-to-video models inherit crop from the still. A 1:1 O1 lock becomes a 1:1 clip.

Generate again when the idea is wrong: wrong location, wrong casting, wrong pack shape. That is a new drawing. It is honest. Generate again when the jacket is the wrong red and everything else is the campaign is dishonest. That is an edit you dressed up as a cheap reroll.

Dev versus full on the generate side is a speed/weight choice inside generate. If you are still exploring, Dev is the distilled tier with strong prompt adherence up to 2K. If you are locking a hero, full O1 Image is the reasoner. Neither Dev nor full is how you change one locked still.

Do not generate a talking clip. There is no duration and no marked audio. A mouth pass is a lipsync row after a plate exists. O1 Image can be that plate. It cannot be the talking take.

Instruct when you can circle the object

Use O1 Image Edit when the approved still is the asset and one attribute is wrong.

The catalog says instruction-driven and subject-driven. Both assume a supplied image. “Make the jacket the red we specified, keep everything else.” “Swap this subject into the scene; keep the lighting.” If your sentence starts with “a photograph of,” you have left the edit and opened a generate.

Keep the instruction to one object when you can. A jacket plus a sky plus a logo is three edits billed as one argument. Do the jacket. Approve. Then the sky. 1 credit each is cheaper than one 1-credit pass that quietly redesigned the product.

There is no aspect-ratio menu published on the edit slug. Do not use it as a reframer for a ratio the pixels do not contain. Crop first if the delivery ratio is the change. Edit second if a subject inside the crop is the change.

2K is the published output size on the edit, not 4K, and not a reason to upsample a still you do not like. If the master needs more pixels than 2K, this row is the look pass, not the print master.

Subject-driven is not a casting call from a blank canvas. You still upload the scene or the subject. No plate, no 1-credit pass. requires_image is true.

Do not spend the credit twice on identity

Wrong: generate O1 Image until the pack is almost right, then generate again to fix the label, then generate again to fix the label. Each generate is a new bottle. The label you “fixed” sits on a cousin SKU.

Right: generate until the pack is the pack. Edit the label. If the label still fails, letter on a typesetting stills row or composite real type. HiDream reasons. It is not Ideogram.

Wrong: edit a still you never approved because “edit is the careful one.” Instruction on a draft you will throw away is still a credit, and it trains you to keep the draft.

Right: approve, then instruct.

After either row, motion is a later slug. Image-to-video wants a plate, which you now have. Do not ask O1 Image to “see if it moves.” Do not ask O1 Image Edit to retouch a clip; pull a frame, edit the still, re-run motion if the sequence has to move again.

Article 50 still applies if you publish a realistic person to an EU audience as authentic-looking media. Editing a real face is not a loophole. Article 50 disclosure is the label.

When you can name whether the plate exists, you can name the slug. Sign up and spend the credit on generate or instruction, not on both as a habit.

FAQ

Does HiDream O1 Image Edit work without an upload?

No. Edit-image, requires_image true. Prompt-only is O1 Image (or Dev), a generate. This slug is the 1-credit change.

Why not always generate at 1 credit if both cost 1?

A generate resamples identity. An edit keeps the plate. The shared sticker is why you should edit the jacket instead of redrawing the campaign. It is not a reason to treat them as one button.

Is O1 Dev the instruction tier?

No. Dev is a distilled generate. Instruction is O1 Image Edit. Full versus Dev is weight inside text-to-image. Generate versus edit is whether a plate already exists.