How to Make a UGC Product Video with the Google Flow API

Introduction

A UGC product video can be built end to end through the useapi.net Google Flow API for 45 Flow credits, with nothing processed locally. One invented presenter shows three invented products to the camera.

UGC is the user-generated-content style that fills social feeds and ad libraries: one person talking to a phone camera, to sell a product, to entertain, or just for fun.

This is the full A-to-Z guide to producing that kind of video at a professional level: the products, the presenter, her lines and the final cut all under your control, in six steps of API calls you can copy as they are.

Here is the result, three 10-second Omni 1.1 Flash clips upscaled to 1080p, every frame and every sound produced through the Google Flow API. The same upscale call goes to 4K when you need professional quality, see Step 5:

The result — 29 seconds, 1080p, three products, three clips
45 credits in total

If you would rather watch than read, here is the whole guide in 75 seconds:

The 75-second walkthrough · watch on YouTube · narration and music from the MiniMax API

Pricing

Omni 1.1 Flash video needs a paid Google AI plan plus a flat $15/month to useapi.net for API access, with no per-call surcharge. Every image, portrait, character and the concatenate below cost nothing. The three 10-second clips cost 15 Flow credits each, 45 in total, which is 45 cents on the Google AI Ultra $99 plan. Full per-tier credit costs are on the Google Flow overview.

About the ids in the requests

Every request below refers to earlier results by their mediaGenerationId. The ids are long, so they are shortened here to user:1234-email:…-image:b0bdf3e8-…, and a line under each request says which earlier result each one is. An image id comes back from POST /images under media[0].image.generatedImage.mediaGenerationId, a character id in the character field of POST /characters, a video id under media[0].mediaGenerationId of the finished job, and an upscaled clip's id is its source's id with _upsampled on the end.

Step 1: a sheet for each product

A product sheet is one image that shows the item from four sides, so the model has the whole object and not one lucky angle. POST /images, square, one request per product. The prompt spells out the camera for every panel, because "front, side, back" on its own gave two identical panels for the mug and no true side for the keyboard. A flat object like the keyboard needs a top-down panel, a cube-shaped one does not. No hands, no captions, no logos, so nothing but the product can leak into the clips.

The same works for a real product. Upload photos of it with POST /assets and pass their ids as reference_1 to reference_4 on this request, and the model turns them into a clean four-panel sheet. The same trick makes a sheet for a package, a logo or a set.

The mug — Nano Banana 2 Lite, id b0bdf3e8
0 credits

The desk robot — Nano Banana 2 Lite, id 7aaed056
0 credits

The keyboard — Nano Banana 2 Lite, id 4f861c23
0 credits

Images cost nothing, so ask for count of 2 to 4 and keep the best sheet. The keyboard took a few rounds before its prompt described a real macro pad with a top-down panel, the cube worked first time.

Step 2: the presenter

Three calls. A front portrait from a description. A three-quarter portrait generated from the first one with reference_1, so the two agree on the face. Then POST /characters with both portraits, which returns a character reference. From then on she is @character_1 in any prompt, and Google pins her through its own entity system instead of through a description.

Front — Nano Banana 2 Lite, id a236ef66
0 credits

Rendered a few looks, picked this one. If a detail needs changing afterwards, the chain say, re-render her with her own portrait as reference_1 and a prompt that names only the change, and the face stays put.

reference_1 is the front portrait, a236ef66. This prompt still mentions freckles, a sweater and earrings left over from an earlier presenter, and it did no harm: with a reference image in the request, the reference wins over the words.

Three-quarter — id d891301c
0 credits

imageReference_1 is the front portrait a236ef66, imageReference_2 the three-quarter d891301c. The response's character field is her id, 53ca1036 below. The voice is one of Google's system presets and rides along with the character, but it cannot reach the clip mode used in Step 4, see What to expect.

Step 3: the still, the whole scene in one frame

One request puts everything together. The character goes in as character_1, the three sheets as reference_1 to reference_3, and the prompt places them by tag, left to right on the table in front of her. This frame becomes both ends of every clip, so it is the one place where the tags do their work.

character_1 is the character from Step 2, 53ca1036. reference_1 to reference_3 are the three sheets from Step 1: the mug b0bdf3e8, the desk robot 7aaed056, the keyboard 4f861c23.

The still — her face from the character, each product from its sheet, id 034cddc8
0 credits

The mug has its glow, the robot its face and antenna, the keyboard its three colours and its port, and she is the woman from the portraits.

Step 4: one clip per product, all from the still

Omni's first-plus-last-frame mode with the still in both slots, one request per product, the same POST /videos each time. This mode takes nothing but the two images, no sheets, no character, no voice reference. That is fine, because the still already holds all of it. Every item she picks up came out matching its sheet from the still alone.

The prompt does three jobs. It pins both ends to the frame, with a quiet beat on each side. It gives her the line and the voice. And it times the choreography: she starts talking as she lifts the item, the item is back on its spot by the eight-second mark, and the last two seconds are a still, silent hold. Without that last part she ran late and her arm snapped into the end frame.

startImage and endImage are both the still from Step 3, 034cddc8. The same two ids go into all three clips.

Clip 1, the mug — 10 seconds, id 55808825, shown after the 1080p upscale of Step 5
15 credits

Clip 2, the desk robot — 10 seconds, id c4d0e6b1, shown after the 1080p upscale of Step 5
15 credits

Clip 3, the keyboard — 10 seconds, id 3cc8ea2f, shown after the 1080p upscale of Step 5
15 credits

Same woman, same table, same light, and each clip opens and closes on the still, so any of them can follow any other. The lines are around 20 words, which is what 10 seconds holds at this pace with room for the beat at the end.

Step 5: upscale to 1080p

Omni 1.1 Flash renders at 720p. POST /videos/upscale takes a clip to 1080p for free on any paid plan, in about 40 seconds, and answers in the same call with the upscaled clip's id and URL. One request per clip, no email, the account comes from the id.

mediaGenerationId is the mug clip from Step 4, 55808825. The same call with c4d0e6b1 and 3cc8ea2f upscales the robot and the keyboard. Each response carries the new clip under media[0].mediaGenerationId, which is the source id with _upsampled on the end, so the three upscaled ids are 55808825-…_upsampled, c4d0e6b1-…_upsampled and 3cc8ea2f-…_upsampled. The clips shown in Step 4 are these upscaled versions. 4K is also available, for 50 credits and only on a Google AI Ultra plan.

Step 6: join them

POST /videos/concatenate takes the three upscaled ids in order and returns the joined file as base64, at the clips' resolution. The trims decide whether the joins feel like one take. Each clip opens with a short hold on the still and closes with a settle back onto it, and untrimmed those stack into a freeze at every cut. But she finishes each line with her hands already resting, so a trim measured from the picture alone cuts into the last words. Cut after the last sound plus a beat, never earlier. On these clips that meant leaving clip 1 whole, since its laugh runs to the last frame, taking just over half a second off the end of clip 2, and a few frames off the start of clips 2 and 3.

The three ids are the upscaled clips from Step 5 in order: the mug, the desk robot, the keyboard.

Joined — 28.8 seconds, 1080p
0 credits

She finishes a line with her hands on the table, the cut lands, and she reaches for the next item. The joined file was re-encoded to a smaller size for hosting, the three clips are as the API returned them.

What to expect

• Generate, do not edit. The UGC talking head post extended a clip with video-to-video edits, and that still works when the lines are short. Here it did not. An edit keeps its source's speech on the source's own timeline, so with a 10-second line filling the seed, a new line got braided into the old one instead of replacing it, and the products left on the table drifted between takes. Generating each clip from the still gave complete lines, intact products and cleaner joins, for fewer credits.
• The still carries the tags. The first-plus-last-frame mode accepts only the two images. Do the tagging in the still, with the character and the sheets, and the clips inherit it.
• Time the choreography. A 10-second clip with a full line needs the put-down called out by the clock, "by the eight-second mark", and a still, silent hold at the end. Left to itself the model ran the line long and jumped into the end frame.
• Describe the voice the same way in every prompt. Nothing carries a voice between separate generations, and a character can hold a voice preset but cannot reach this mode. With one identical description the three clips came back close enough to pass as one take.
• Sheets need per-panel cameras. Name what each of the four panels shows. Give flat objects a top-down panel. Keep hands, captions and logos out, because whatever is on the sheet can end up in the clip.
• Trim to the sound, not the picture. She goes still before she stops talking. A cut placed at the first still frame took the end of a line with it.
• The presenter is invented. Google's moderation is stricter with real, identifiable people. Start from a generated character.

What it cost

Step	Mode	Credits
Three product sheets, 1:1	Nano Banana 2 Lite	0
Two presenter portraits, 3:4	Nano Banana 2 Lite	0
The character	—	0
The still, 9:16	Nano Banana 2 Lite, character + 3 references	0
Clip 1, the mug, 10s	I2V, first + last frame	15
Clip 2, the desk robot, 10s	I2V, first + last frame	15
Clip 3, the keyboard, 10s	I2V, first + last frame	15
Upscale, three clips to 1080p	—	0
Concatenate	—	0
Total		45

A 10-second Omni clip costs 15 credits at 720p, so the whole video is 45 cents on the Google AI Ultra $99 plan. Images, characters, the 1080p upscale and the concatenate cost nothing, so iterate freely on the sheets, the presenter and the still before spending a credit on video. 4K is the one upgrade that costs, 50 credits a clip on Ultra.

Run it yourself

Every step above is one API call, and the six of them chain into a script. The complete Node.js pipeline, a Python port, and the prompts.json holding every prompt on this page live in useapi/google-flow-api/ugc-product-video. Clone it, put your own products and lines in prompts.json, and one command builds the whole video.

It checkpoints after every step, so an interrupted run resumes instead of paying for the clips again, and you can redo a single step — a better still, a different presenter — without regenerating what came before it.

Frequently asked questions

• Can I use my own product photos instead of generated sheets? Yes. Upload them with POST /assets and pass the returned ids as reference_1 to reference_3 on the still, exactly as the generated sheets are passed here. The four-panel sheet is a way to give the model every side of an object, and a real product benefits from the same treatment.
• Why not edit one clip into the next, like the talking-head post? Because a video-to-video edit keeps its source's speech on the source's timeline. With short lines that replacement works. With a 10-second line filling the clip, the new line got braided into the old one and the products on the table drifted. Generating every clip from the same still avoids both, and costs 15 credits instead of 20. See What to expect.
• Why do all three clips start and end on the same frame? So they can be joined in any order with no visible cut. The first-plus-last-frame mode pins both ends of each clip to the still, and the concatenate trims the short holds around each join. The mechanism is explained in the UGC talking head post.
• Does the voice stay the same across the clips? Nothing carries a voice between separate first-plus-last generations, since that mode takes no audio reference, and a character's voice preset cannot reach it either. Describing the voice identically in every prompt kept these three close enough to pass as one take. If you need a guaranteed voice, the reference-to-video mode accepts referenceAudio_1, at the cost of the pinned frames.
• How much does it cost? You keep your paid Google AI plan and add the flat useapi.net subscription. The whole video is 45 Flow credits, 45 cents on the Ultra $99 plan, and every image, portrait, character and the concatenate are free. See Pricing.

Conclusion

Visit our Discord Server or Telegram Channel for any support questions and concerns.

Check our GitHub repo with code examples.