H3 Flex guide: specs, modes and prompt tips
Updated 2026-10-30. Every number on this page comes from Pika’s public documentation, linked in Sources. Independent site, not affiliated with Pika or MiniMax.
What is H3 Flex?
Pika H3 Flex is an open-weight AI video model from Pika. According to its model card, it is a post-trained version of MiniMax H3 that generates video together with synchronized stereo audio in six sampling steps, and accepts up to 12 reference files for shot direction. Pika lists it for multi-shot storytelling, motion graphics, product films and reference-guided animation.
ComfyUI’s tutorial describes how it was made: Pika fine-tuned H3 on video, audio and reference examples, distilled the 50-step generator to six steps, and refined it with reward feedback and preference optimization. It ships as two LoRA adapters on the MiniMax H3 base.
H3 Flex specs
| Duration | 5–15 s (whole seconds); 5 s → 124 frames ≈ 5.17 s, 15 s → 362 frames ≈ 15.08 s |
|---|---|
| Frame rate | 24 fps |
| Resolution | Native 480p / 768p; 1080p, 2K, 4K via Pika API upscaling |
| Aspect ratios | 21:9, 16:9, 3:2, 4:3, 5:4, 1:1, 4:5, 3:4, 2:3, 9:16, 9:21 (API also accepts auto) |
| Audio | 32 kHz stereo mixed into the video; optional stems via API |
| Output file | MP4 (H.264 + AAC) plus JSON metadata |
| Prompt length | 1–8,000 characters |
| Model size | 33B parameters (Hugging Face listing) |
| License | MiniMax H3 Community License Agreement (adapters are model derivatives of MiniMax H3) |
The five H3 Flex modes
- Text to video and audio — prompt only (
t2va). - First-frame (image to video) — prompt plus an opening frame (
i2va). - Last-frame — prompt plus an ending frame (
l2va). - First-and-last-frame — prompt plus both endpoint frames (
fl2va). - Multimodal references — up to 9 images, 3 videos and 3 audio clips, 12 files total, at least one image or video (
ref2va). Each video or audio clip must be 2–15 s, and combined video and combined audio must each be ≤15 s.
H3 Flex prompt tips
Pika’s showcase prompts follow a consistent pattern. Put these in order:
- Lock the references. If you use an image, say what must stay the same: “Image 1 is the exact two players, outfits and pool: keep them exactly as in Image 1.”
- Style and look. e.g. “Photoreal 35mm film look, true colour, light natural grain.”
- Timed shots. One line per time range with one camera move each: “0–3 s: top-down overhead… 3–6 s: smooth crane down to table height…”. For cuts, label them [Shot 1], [Shot 2].
- Speed. Pika’s examples say it outright: “everything moves at normal real-time speed… NO slow motion.”
- Sound, separately. A “Sound:” line for diegetic sound (footsteps, wind, a knock) and a “Music:” line — or say “No music, no dialogue” when you want silence. Put dialogue in quotes with “Only these two lines.”
- Negatives at the end. “No text, no captions, no logos.”
Practical notes from ComfyUI’s testing of the released adapters (source):
- Camera amplitude is advisory: describe the move inside the shot sentence and always give it a stopping point. Explicit degrees and cut timestamps can land 0.2–0.3 s off.
- “Wide shot” alone tends to render as a medium shot; describing the subject as small inside a large empty room holds the framing better.
- Fine contact constraints are unreliable (a hand meant to stop short usually touches).
- Short rendered text on props and products stays legible across camera moves.
Copy-paste prompt template
[Subject and setting in one or two sentences.] [Style: e.g. photoreal 35mm film look, soft daylight, light grain.] 0–3 s: [shot size + camera move], [what happens], [where the camera stops]. 3–7 s: [next camera move], [action]. 7–10 s: [next camera move], [action]. Everything moves at normal real-time speed, no slow motion. Sound: [diegetic sounds]. Music: [genre and tempo, or "no music"]. No text, no captions, no logos.
Where to use H3 Flex today
- Pika API — text to video, image to video, reference to video, draft to video (upscale), prompt rewriter and audio stems (dev.pika.art).
- Your own GPU — weights and inference package on Hugging Face; Pika’s tested cases ran on H100 80GB, 768p/15 s needed
--cpu-offload, and you need about 211 GB of disk for base components (hardware notes). - ComfyUI — official R2V, FL2VA and T2V workflows (tutorial).
- This site — a no-install, pay-per-video generator is coming soon; it is not live yet.
H3 Flex API pricing (Pika, read 2026-10-30)
- Text to video: 480p $0.01/s · 768p $0.02/s · 1080p $0.035/s · 2K $0.04/s · 4K $0.08/s; only successful generations are charged (source).
- Image to video from $0.01/s; reference to video from $0.03/s; prompt rewriter $0.02/request; audio stems $0.001/s (source).
Example: a 10-second 1080p text-to-video clip is 10 × $0.035 = $0.35 at those rates.
H3 Flex FAQ
How long can an H3 Flex video be?
You request a whole number of seconds from 5 to 15. The output follows the model frame grid, so a 5-second request gives 124 frames (about 5.17 s) and a 15-second request gives 362 frames (about 15.08 s), at 24 fps.
What resolution does H3 Flex output?
Native 480p or 768p. Pika’s hosted API can upscale to 1080p, 2K or 4K; it generates at 768p first and then upscales.
Does H3 Flex generate sound?
Yes. Output is an MP4 with H.264 video and AAC audio; locally the audio is 32 kHz stereo mixed into the video. The API can also return separate Dialogue, Score, SFX, Foley and Ambience stems.
Can I run H3 Flex on my own GPU?
The weights are on Hugging Face. Pika’s tested configurations are on an H100 80GB; 480p/5 s used about 66–69 GiB of GPU memory, and 768p/15 s needed CPU offload. Plan for about 211 GB of disk for base components plus about 5.5 GB for the adapters.
How much does H3 Flex cost on Pika’s API?
Pika’s text-to-video page lists $0.01/s at 480p, $0.02/s at 768p, $0.035/s at 1080p, $0.04/s at 2K and $0.08/s at 4K, charged only for successful generations (read on 2026-10-30; check the source for current prices).
Get notified when our H3 Flex generator opens
Sources
- Pika-Labs/Pika-H3-Flex model card (Hugging Face) — specs, modes, reference limits, license.
- Pika API: H3 Flex text to video — parameters, resolutions, prices.
- Pika API: H3 Flex models — endpoint list and starting prices.
- Pika-H3-Flex hardware notes (GitHub) — tested GPUs, memory, disk.
- ComfyUI: Pika H3 Flex workflows — training summary and prompting notes.
All sources read on 2026-10-30. Specs and prices can change; the official pages win.