Imagine is one composer for image and video

Imagine is one composer. Open it at Imagine. You get a prompt box, an Image | Video toggle, a slot for references, and a way to run several models on that same prompt. You pick a mode, you attach a file or you do not, and the job type follows.

Leave the toggle on Image until the still is one you would publish. Then drag that still into the reference slot, switch to Video, and spend the motion credits on a frame you already approved. Attaching a reference is what flips the job: text-to-image becomes image-to-image; text-to-video becomes image-to-video. You do not type those names. The attachment does the flip.

The chat doors for the same work are generate an image from text, edit a photo with AI, generate a video from text, and turn a photo into a video. Use the composer or use chat. Do not fire the same prompt in both and treat the four files as a test.

This is the same account as the Android app on your phone. The compare-and-lock work happens here.

Image or Video, same composer

Sign in. Open Imagine. The toggle is Image | Video.

Image with an empty reference slot is text-to-image. You describe a still. You get stills. Use it for a hook frame, a product plate, a character key, a background you will later animate.

Video with an empty reference slot is text-to-video. You describe a clip. The model invents the first frame and the motion together. That is the expensive way to discover you hate the composition.

Stay on this screen for both jobs. Flip the toggle. The prompt can stay. The references can stay. The models you pick should change, because a stills row and a motion row are different lines in the model catalog. Read the catalog when you do not already know which rows you are comparing. Every plan gets the same models and editor. Credits are volume. Nothing on this screen is a Pro toggle.

Imagine is not the workflow canvas. The canvas is scenes as columns, nodes such as Image → Video → Voiceover, a recipe you save. Imagine is one shot, one prompt, one compare. When the shot is a format you will rerun, leave Imagine and save the recipe. The video editor is a timeline: drop a finished clip there to trim, caption, score, and render. Generate the frame and the motion here. Assemble later.

Attach a reference and the job flips

The toggle is Image or Video. The reference slot is what changes the job inside that mode.

Toggle	No reference	Reference attached
Image	Text-to-image	Image-to-image
Video	Text-to-video	Image-to-video

Attach a photo you already have — a pack shot, a face you have rights to, a room, a still you just generated — and Image stops being a blank-page generate. It becomes an edit of that file.

Attach that same still, switch to Video, and the job is image-to-video. The still is frame one. Your prompt should describe motion, not re-describe the subject. "Slow push-in, steam drifting off the cup, handheld" is a video prompt on a locked still. "A ceramic cup on a linen table in morning window light" is a still prompt. Do not paste the still prompt onto the video run and hope the model ignores the nouns it already has.

There is no second control named image-to-image. The attachment is the control. Take the reference off and you are back to text-to-image or text-to-video, matching the toggle. Useful when a reference is fighting the prompt. Not useful as a habit. Empty Video is how people spend a motion budget on a composition they have not seen yet.

A product you sell should enter as a photo you own, not as a sentence. Image-to-image from the pack shot will hold the label better than a text-to-image guess of "our teal carton." Video from that approved still will hold it better than text-to-video. If you only have a website, that is the Brand screen, not this slot.

Several models, one prompt

Run several models on one prompt. That is the compare. The prompt stays still. The rows change. You look at the files, not at the names.

Two or three rows is a test. Six rows is a shopping trip. Pick candidates from the catalog because you already know what they are for — photoreal still, lettering still, short motion, longer motion — not because the list is long. Every plan can reach every row. The credit cost per row is on the catalog. Read it before you run three expensive motion models against a still you have not locked.

Keep the prompt identical across the set. If you rewrite the lighting for model B, you are no longer comparing models. You are comparing briefs. Change one variable.

Score the stills on the job, not on vibe:

• Is the subject the one you asked for.
• Can you read any type that has to be type.
• Is the crop 9:16 if this is a hook frame, 16:9 if it is a wide plate.
• Would you publish this frame if it never moved.

Pick one. The other files can sit in the library. Drag the winner in as a reference and continue from that file. Do not average two models by running a third prompt that describes both outputs.

A motion row costs more than a stills row for the same idea. Compare stills first. Compare video models only once the frame is the frame. One video model on a locked still beats three video models on a sentence.

Six references, and drag a generation back in

You can attach up to six references. That is a stack, not a pile of retries.

A useful stack is different jobs in the frame: the product, the person, the room, a style still you already approved. Six angles of the same bottle are one reference with extra files. Stay under six. The seventh file is not a hidden slot. If you need a new subject, swap one out.

Anything you already generated can come back in. Drag that still — or a frame you like from a clip — into the reference slot. You do not download it and re-upload it. The library strip is the file. This is how a text-to-image winner becomes the image-to-image base, and how that base becomes the image-to-video first frame, without leaving Imagine.

Iterate on Image first. Generate stills. Pick the closest. Drag it back in and prompt only the change: "same cup, cooler window light, label facing camera." When the still is the one you would ship, keep it in the slot, switch to Video, and prompt only the motion. Dragging is free. Re-generating the subject from a new sentence because you did not drag the winner is how a cheap still becomes three motion runs that still do not match.

If a generation is close and you only need a local fix, stay on Image with that file attached. Do not switch to Video to see if motion hides it. Motion will not hide a crooked label. It will pan across one.

Lock the still before you spend on motion

This is the day-one rule. A still is cheap enough to throw away. A clip is not. Text-to-video spends motion credits on a first frame nobody approved. If the face is wrong, the pack is wrong, or the crop is 1:1 on a 9:16 slot, you paid for seconds of a still you would have rejected in a glance.

The order:

1. Toggle Image. Attach a photo if you have one; leave the slot empty if you do not.
2. Write a still prompt. Subject, setting, light, crop. One scene.
3. Run two or three stills models on that prompt.
4. Pick the frame you would publish on mute, as a poster, at phone size.
5. If it needs a fix, drag it back in and edit. Repeat until the still is locked.
6. Keep that locked still in the reference slot.
7. Toggle Video. Prompt the camera and the motion only.
8. Run one video model. If the motion is wrong, change the motion prompt, not the subject. If the subject drifted, go back to Image.

Do not keep rolling video until the pack looks right. Image-to-video cannot invent a better bottle than the still you gave it. Give it a better still.

Once the clip is the shot you wanted, stop. If you need a second shot, lock a second still. A multi-scene piece is a workflow: scenes as columns, Image → Video in each scene, save and reuse that recipe so next week is a run. Imagine will not remember the series for you. It will remember the files in your library, which you can drag back in.

Credits are the same currency on every plan. A free account is enough to start. Weekly is 240 credits, Standard Monthly 774, Pro Monthly 1,600, with annual buckets at 9,600 and 19,200. They roll over while you are subscribed. Read the quote before you confirm a run. Do not invent a number in the prompt.

Three moves on the first still

1. Open Imagine. Leave the toggle on Image. Type one still. Run two or three models from the catalog on that prompt. Pick the frame you would ship.
2. Drag that generation back in as a reference. If it needs a fix, stay on Image and prompt only the change. Do not switch to Video yet.
3. When the still is locked, keep it attached, toggle Video, describe only the motion, and run one video model. That is image-to-video. You already paid to like the first frame.

If the first session is a product you sell, start with your own pack shot in the reference slot so step 1 is already image-to-image. If you needed a series, take the locked still and the motion prompt to the canvas and save them there. Do not keep re-typing the subject into an empty Video box.