This generator runs Pika H3 Flex — Pika Labs' post-trained MiniMax H3 video model, released 2026-10-29. Use FL2VA for text or first-and-last-frame input, or Ref2VA for up to 12 reference images, clips or audio stems. Six sampling steps, up to 4K, up to 15 seconds, synced audio. Free waitlist tier launching soon; paid plans listed on the pricing page.
First 5 images are included; each additional image uses 3 credits.
MP4 / MOV / MKV, up to 3 files, 50MB each
MP3 / WAV / AAC / FLAC, up to 3 files, 15MB each
Recommended for testing prompts and first drafts.
2K output is available on the 2K tier only.
Checking your account…
Pick a result, make it yours
Choose an example and your prompt, model, and settings load instantly.
Cinematic“Pika H3 Flex (FL2VA direct) — 2K text-to-video sample from the official Pika Labs release. 10 seconds at 2560x1440, six-step LCM sampler, synced audio.”
Pika H3 Flex (FL2VA direct) — 2K text-to-video sample from the official Pika Labs release. 10 seconds at 2560x1440, six-step LCM sampler, synced audio.
Cinematic“Pika H3 Flex (I2VA direct) — 2K image-to-video sample from the official Pika Labs release. 8 seconds at 2560x1440 generated from a single reference frame with synced audio.”
Pika H3 Flex (I2VA direct) — 2K image-to-video sample from the official Pika Labs release. 8 seconds at 2560x1440 generated from a single reference frame with synced audio.
Cinematic“Pika H3 Flex (Ref2VA direct) — 2K multimodal reference-to-video sample from the official Pika Labs release. 5 seconds at 2560x1440 with identity preservation and synced audio.”
Pika H3 Flex (Ref2VA direct) — 2K multimodal reference-to-video sample from the official Pika Labs release. 5 seconds at 2560x1440 with identity preservation and synced audio.
Free tier launching soon — join the waitlist above. Paid tiers are listed on the pricing page.
This page runs the hosted Pika H3 generator. It supports two modes — FL2VA for text or first-and-last-frame input, and Ref2VA for reference-conditioned generation — both at the model's recommended six-step sampling.
Everything below the tool is written to be readable by search engines and language models without JavaScript. If you need the open weights, pull them from huggingface.co/Pika-Labs/Pika-H3-Flex.
The Pika H3 generator runs Pika H3 Flex, a post-trained version of MiniMax H3 released by Pika Labs on 2026-10-29. The hosted backend calls the same FL2VA and Ref2VA adapter weights that Pika Labs published on Hugging Face, with the model's recommended six-step sampler and the official 15-second maximum duration.
| Base model | MiniMax H3 |
|---|---|
| Recommended steps | 6 |
| Max resolution | 4K (tiled) |
| Max duration | 15 seconds |
| Synced audio | Yes — up to 5 exportable stems |
Both adapters sit on top of the same base model but accept different inputs. Pick the one that matches what you have.
| FL2VA | Details |
|---|---|
| Input | Text, or first frame + last frame |
| Best for | Concept shots, clean interpolation, title cards |
| Typical params | Steps 6, duration 8–12s, 1080p |
| Common failures | Over-detailed prompts; wrong aspect ratio between frame pair |
| Ref2VA | Details |
|---|---|
| Input | Up to 12 references: images, clips, audio stems |
| Best for | Identity preservation, soundtrack-driven motion, product shots |
| Typical params | Steps 6, duration 10–15s, 3–6 references |
| Common failures | Too many references over-constrain motion; low-contrast refs ignored |
| Parameter | Recommended | Range | If you push it |
|---|---|---|---|
| Steps | 6 | 4–20 | Latency grows linearly; no visible gain past 8. |
| Resolution | 1080p | 720p–4K | 4K enters tiled mode; much slower. |
| Duration | 10s | 1–15s | Hard cap at 15s; audio stems clip at 15s. |
| References | 3 | 0–12 | More than 6 can over-constrain motion. |
| Seed | Fixed while iterating | any 64-bit int | Random seeds make prompt debugging impossible. |
As of November 2026, the hosted Pika H3 free tier is in waitlist mode. Published policy on launch: a limited number of generations per day, no watermark on paid tiers, and MP4 export at the resolution you selected. The open-weights release itself has no watermark at all when run locally.
Side-by-side with the paid alternative at pikah3.com, which sells one-time credit packs:
| This page | pikah3.com | |
|---|---|---|
| Price | Free tier (waitlist) + paid plans | One-time packs at $19.90 / $49.90 / $99.90 |
| Signup | Required for free credits | Required for purchase |
| Watermark | None (paid tiers) | None |
| Parameter transparency | Steps / seed / references exposed | Prompt + mode only |
Commercial-use rights follow whichever path you chose. Open-weights outputs are governed by the Hugging Face model card. Hosted-service outputs follow the service's own terms — check the pricing page for the exact grant.
A free tier is launching soon. Join the waitlist from the panel above; once live, free users will get a limited number of generations per day with the same FL2VA and Ref2VA modes paying users have. Hosted usage limits are posted on the pricing page.
No. This page runs the model in your browser against a hosted backend, so there is nothing to install. ComfyUI is only required if you want to run the open weights on your own GPU; the model itself is identical.
FL2VA generates video from text or from a first-and-last frame pair. Ref2VA generates video conditioned on up to twelve reference inputs — images, short clips or audio stems. Pick FL2VA for text-driven ideas or clean frame-to-frame interpolation; pick Ref2VA when identity, motion or soundtrack must come from a reference.
Fifteen seconds per clip. The sampler runs at six denoising steps regardless of duration, so a fifteen-second render takes roughly the same wall-time as a shorter one at the same resolution.
Yes. Pika H3 Flex is the full product name of the model Pika Labs shipped on 2026-10-29; "Pika H3" is the short name used in docs and most references. There is only one model.
As of November 2026, outputs from the open-weights release are licensed for commercial use subject to the model card terms. Hosted-service outputs follow the hosting service's own terms, which are usually looser for paid tiers; check the pricing page for the exact grant.
720p, 1080p, 2K and 4K. 4K requires tiled resampling on all but the largest GPUs; the hosted generator above exposes 1080p and 2K by default, with 4K available for longer-running jobs.
On Hugging Face at huggingface.co/Pika-Labs/Pika-H3-Flex. FL2VA and Ref2VA are published as separate adapters so they can be loaded independently in ComfyUI.
Related: Pika H3 · Pika H3 ComfyUI · Pika H3 Prompts · Pika H3 vs Pika H3 Flex · MiniMax H3
