4K Finishing & Audio Stems
Explaining native 480p/768p latent generation versus cloud upscaling to 1080p/2K/4K, the two documented draft-to-video input paths, and five-stem WAV audio separation.
Native 480p/768p Generation vs. Hosted 1080p/2K/4K Upscaling
The boundary between open-weight execution and cloud finishing centers on the resolution of the diffusion latents:
Native Latent Generation (480p & 768p)
The underlying MiniMax H3 / Pika H3 Flex diffusion transformer computes latents natively where the shorter dimension is 480 pixels or 768 pixels (e.g. 1344×768 or 768×1344). In local open-weights ComfyUI pipelines, the video VAE decodes directly from these raw latents into 480p or 768p video frames.
Hosted Cloud Finishing (1080p, 2K & 4K)
Outputs at 1080p, 2K, and 4K are generated through Pika's cloud finishing pipelines. These stages upscale and enhance native 480p/768p latents or videos to production resolutions while maintaining temporal continuity. These upscaling pipelines are not distributed in the open weights.
The Documented Draft-to-Video API Paths
According to Pika's official API reference (dev.pika.art/models/pika/pika-h3-flex/draft-to-video), the Draft-to-Video endpoint upgrades an existing 480p or 768p video to 1080p, 2K, or 4K. The schema documents two distinct input mechanisms:
Path 1: In-Platform Draft Job ID
Parameter: draft_job_id (string)
Used when the initial draft was generated directly on Pika’s platform. Passing the ID of a completed draft job upgrades that take to the target resolution. Pricing is calculated as the differential between the draft tier and the finishing tier rate.
Path 2: External Video URL
Parameter: video (string URL)
Used when providing an external 480p or 768p video URL. When uploading a local file (such as a local ComfyUI generation) through Pika's documented upload flow, the endpoint returns two values: upload_url is the temporary pre-signed destination for the file PUT request, whereas url is the persistent media URL that must be passed to the model as video. Do not pass upload_url to the model endpoint.
These two input paths reflect the documented request schema published in Pika’s developer documentation. This guide documents published API specifications; we have not executed an empirical test of either path. Limits documented by Pika require external input videos to match supported 480p or 768p dimensions.
How to Obtain Stems & What Is Returned
While local open-weights inference produces a combined stereo mix, Pika’s hosted cloud platform offers two documented ways to separate soundtracks into isolated audio tracks:
Method 1: In-Generation Mode
Parameter: output_mode: "audio-stems"
Request stems directly alongside an initial Pika H3 Flex video generation request. The pipeline outputs the generated video together with the separated audio stems.
Billing: The stem extraction fee ($0.001 per output-stem second across the sum of decoded output-stem durations) is added to the base video generation cost.
Method 2: Standalone Endpoint Mode
Endpoint: dev.pika.art/models/pika/pika-h3-flex/audio-stems
Call the standalone audio-stems endpoint on an existing media file up to 120 seconds in length, providing either:
request_id: The ID of a completed Pika H3 Flex generation owned by your organization.input_url: A publicly accessible media URL pointing to any audio or video file with sound. For local files, upload via Pika's upload flow and pass the returnedurl(not the temporaryupload_url).
Billing: Billed for the stem separation job at $0.001 per output-stem second across the sum of decoded output-stem durations.
Both methods above reflect published vendor documentation from pika.art/pika-h3-flex and dev.pika.art/models/pika/pika-h3-flex/audio-stems. We have not empirically executed either stem extraction path.
Returned ZIP Archive & The Five Stem Categories
The official audio-stems reference confirms that the job outputs a ZIP archive containing uncompressed WAV audio files spanning five distinct categories (without prescribing fixed member filenames in the documentation):
| Stem Category | Format | Content Description | Post-Production DAW / NLE Role |
|---|---|---|---|
| Dialogue | WAV (in ZIP) | Isolated spoken character dialogue and voiceovers. | Allows surgical EQ, de-essing, leveling, or full dialogue dubbing without affecting sound effects or score. |
| Score | WAV (in ZIP) | Non-diegetic background musical composition. | Allows ducking music beneath dialogue, re-timing musical cues, or replacing the score with a licensed track. |
| SFX (Sound Effects) | WAV (in ZIP) | Dynamic impact and mechanical sound effects. | Enables spatial panning across surround sound channels or layering with sound effects libraries. |
| Foley | WAV (in ZIP) | Subtle physical interactions (footsteps, clothing rustle, object handling). | Preserves physical realism and acoustic weight when the musical score is modified or muted. |
| Ambience | WAV (in ZIP) | Continuous room tone, weather elements, wind, and environmental background. | Maintains consistent acoustic atmosphere across cut points, preventing jarring room-tone dropouts. |
Audio-Stem Billing Mechanics: Sum of Decoded Output Durations
According to Pika’s official pricing documentation (dev.pika.art/models/pika/pika-h3-flex/audio-stems/pricing), stem extraction is billed usage-based at $0.001 per output second based on the sum of decoded output-stem durations, not simply the input video duration:
Billing Calculation Example
When you submit a 10-second video for stem separation and all five 10-second stems are generated and returned in the ZIP archive:
5 stems × 10 seconds = 50 output-stem seconds; 50 × $0.001 = $0.05.
This $0.05 stem extraction fee is billed in addition to any base video generation costs. Note that the documented API processes and outputs all five stems together rather than offering individual track subsets.
Verified API Pricing Schedule
The pricing table below reflects official published rates verified on October 30, 2026 from pika.art/pika-h3-flex:
| Output Resolution | Regular Rate | Launch Promo Rate | Billing Notes |
|---|---|---|---|
| 480p | $0.02 / second | $0.01 / second | Standard preview tier; fast draft iteration. |
| 768p | $0.04 / second | $0.02 / second | Matches maximum open-weight local native resolution. |
| 1080p | $0.07 / second | $0.035 / second | Full HD cloud finishing tier. |
| 2K | $0.08 / second | $0.04 / second | High-density cinematic finish. |
| 4K | $0.16 / second | $0.08 / second | Ultra-high-definition production master. |
| Audio Stems (Add-on) | $0.001 / second | $0.001 / second | Billed by sum of decoded output stem durations ($0.05 per 10s clip for five 10s stems). |
Launch Promotion Window: Verified as of October 30, 2026, members of the Pika API Club receive 15 free Pika H3 Flex generations per day during the promotional period running from October 29 to November 5, 2026.
Ongoing Usage Note: This daily free generation quota is a temporary launch event. It does not make 4K finishing or audio stem separation permanently free. Once the promotion expires or daily limits are exceeded, standard per-second rates apply.
Additional Usage & Billing Rules
- Video References: Video-to-video reference generation is billed for input-video seconds plus output-video seconds at the target resolution rate. Image and audio references do not incur extra per-second surcharges.
- Draft-to-Video Differential Billing: When upgrading an existing draft generated via Pika Flex to 1080p, 2K, or 4K using
draft_job_id, billing reflects the difference between the base draft tier and the target finishing tier rate. - Prompt Rewriter: Automatically included for generations processed through Pika's platform; standalone calls to the prompt rewriter API cost $0.02 per request.