Pika H3 Flex Gives Creators More Control Over the Second Take

Getting a promising first clip is one thing. Keeping the camera move the client liked while changing the backdrop is another. Pika H3 Flex gives you a way to ask for that second version. Pika released the model through its developer API and as downloadable components on October 29. The company also announced November 2 for its creation app. As of our November 3 check in Sydney, its public catalog still did not show Flex.

Flex builds on MiniMax H3 and generates five to 15 seconds of video with synchronized sound. For a creator, its appeal is being able to show the model which parts of a shot matter, compare rough takes and pay to finish the one worth keeping. This early assessment draws on documentation, published demonstrations and an independent creator's local tests. Choosely did not generate clips.

Give the camera move its own reference

Imagine a ten-second film for a bottle of perfume. You have a product photo, a mood board for the lighting and five seconds of phone video showing a slow camera move around a bottle. In Pika's Omni Reference guide, you tag each file in the prompt and give it a job. For this illustrative shot, that might read, “Use @Image1 for the bottle and cap, @Image2 for the warm window light, and only the camera movement from @Video1.” Now the model has a reason to ignore the kitchen counter in your phone clip.

The reference route accepts 12 files in total, with ceilings of nine images, three videos and three audio files. Include at least one image or video. Each video or audio reference can run from two to 15 seconds, and the combined limit is 15 seconds for each of those two media types. If all you need is a particular opening and closing composition, keyframes let you give the model those two images directly.

Pika's prompt guide shows how to direct a sequence of shots. Its launch demos illustrate a different task, changing a prop, outfit or setting around an existing performance. In Pika's own blind preference tests, competing models still had advantages on physics, multi-scene storytelling and some video-guided tasks. For the perfume film, watch the cap as the camera passes behind it. Keeping that detail recognizable matters more than how closely one still matches the mood board.

Keep the take and finish it once

Flex generates at 480p or 768p. Ask the hosted API for 1080p, 2K or 4K directly and it generates at 768p, then upscales within the same job. You can also compare drafts first. Give the finishing endpoint a completed 480p or 768p Flex job from your account and it upscales that take with its existing sound.

Consider a text-generated draft with ten billable seconds at 768p. At Pika's listed US launch rates, the draft costs $0.20 and the equivalent 4K job costs $0.80. Because Pika deducts the eligible draft price, finishing that take would add $0.60. The endpoint also accepts an external 480p or 768p video, so someone working locally can send just the selected clip for finishing. For an outside clip, Pika charges the target rate minus the native rate per output second. At the same ten-second duration and launch text rates, that also comes to $0.60.

Sound can follow the same pattern. The original clip includes a stereo mix. Pika's standalone separation service accepts audio or video up to 120 seconds, including outside files. It returns three WAV tracks for dialogue, music and effects. Three ten-second stems cost $0.03 at $0.001 per output stem second. If you need footsteps isolated from room tone, take care with the product demonstration: Pika identifies its separate Foley and ambience tracks as part of an earlier five-stem preset.

What took seven seconds?

Pika reduced the model's denoising process from 50 steps to six. In its controlled benchmark, the transformer performed its part of a 15-second 768p generation in as little as seven seconds. A creator waiting for a finished file also has to allow for the queue, reference processing, decoding and any finishing or separation. Pika hasn't supplied an end-to-end delivery time for our perfume shot.

One creator measured a different workflow on an RTX 5090 with 32GB of GPU memory. Stuubzzz Creative Studios ran 75 ComfyUI renders across three setups. For four 15-second scenarios, the fastest Flex run of three seeds took 238 to 252 seconds per scenario. The author heard slightly clearer sound and judged the picture slightly behind the other two setups. Across the 15-second takes, the test counted 233 held frames across 12 long Flex takes, against 91 and one for the other two setups. Half the Flex takes had none. It was a nonblind local test by a creator who sells ComfyUI training.

What the references add to the bill

That five-second camera reference has a price. At 768p the text, image and keyframe routes charge for output seconds. The mixed-reference route has a higher rate and counts input-video seconds too. Image and audio attachments add no input-duration charge.

These examples assume the exact billable durations shown. A campaign budget also has to cover the takes you discard, any paid finishing and your plan. API Club lists a $10 monthly membership that includes $10 in API credits. The model page offers members 15 free generations per day from October 29 to November 5. Pika's press release describes two weeks of free daily generations and a 50% discount for members and subscribers. Check the offer attached to your account before counting those generations in a budget.

For creators using the app, Pika has a guide to its reference, keyframe, draft and audio controls. At about 8:38 a.m. Sydney time on November 3, or 1:38 p.m. on November 2 in California, Flex was absent from the public video-tools catalog. Its direct app link led to sign-in. Those public pages did not establish which accounts could use Flex or what a clip cost in app credits.

What you can run locally

Pika released two adapters, small sets of changes that work with the much larger MiniMax H3 base. Together they support text, endpoint frames and mixed references, producing native video with mixed sound. Pika's upscaler, stem separation, optional prompt enhancement and fast serving system run on its servers. Its official local configuration needs at least 211GB for the base components alone. Pika tested representative jobs on an H100 with 80GB of GPU memory.

The inference code carries an Apache-2.0 license. The adapters carry the MiniMax H3 Community License, whose grant excludes the US, UK, EU and Republic of Korea. MiniMax invites users in those territories to seek a separate license. For API users, Pika says its agreement licenses outputs for commercial use. Check the terms for the route you plan to use.

Choosely's verdict

Flex is worth a look if client feedback regularly sounds like “keep the move, change the look.” Use keyframes when the start and finish are enough. Pay for a video reference when the movement itself is the brief. Approve the motion at draft resolution, then finish the keeper. If you already generate locally, buying just the upscale or the soundtrack split may cover the part your workflow lacks.

For someone choosing a consumer subscription, the decision turns on whether Flex appears in their account and what the app charges for that shot. For a wider comparison of motion, audio and editing workflows, start with Choosely's AI video generator guide.