Pika H3 Flex character consistency with a turnaround sheet

From turnaround sheet to Pika H3 Flex in five steps

1. Get a character sheet: download a pack from the library or a free character. The front, side and back are already separate files.
2. Upload the five images below as references, in this order, so the prompt can name them @Image1 to @Image5. Image references cost nothing extra on the API.
3. Add a short video of the movement you want as @Video1 if the character has to dance, fight or walk a specific way. Keep it 2–15 seconds; it is billed by its length.
4. In the prompt, say what every reference supplies, keep the identity line identical in every clip, and write each shot with a time range and a stopping point for the camera.
5. Draft at 480p. Finish only the take you keep at 1080p, 2K or 4K with draft-to-video, which keeps the same take and audio and charges only the price difference.

Browse character turnaround sheets

1. @Image102_turnaround_views/view_front.pngFace and front of the outfit
2. @Image202_turnaround_views/view_side.pngProfile and hair from the side
3. @Image302_turnaround_views/view_back.pngHair and outfit from behind
4. @Image401_character_reference/full_body.pngHeight and proportions
5. @Image505_scenes/scene_workspace.pngThe location

Pika's own showcases do both: one prompt calls a single character sheet "front, three-quarter and back views of ONE character", another uses an opening frame plus a separate character sheet. Separate files let you name which image holds the face, the profile and the back, which matters when a shot shows the character from the side or from behind.

Prompt template: character sheet + scene + motion video

Adapted from Pika's "Record-shop dance" showcase, which uses exactly this setup: a character sheet, an environment image and a motion-reference video. The key line is "take NOTHING else from @Video1": without it the person in the motion video can leak into the result.

Prompt template: multi-shot scene without a motion video

Pika and ComfyUI both direct shots with time ranges and hard cuts inside one prompt. Showing the side in one shot and the back in another gives each turnaround view a job. ComfyUI's testing notes say cut timestamps can land 0.2–0.3 seconds off, so do not rely on frame-exact cuts.

Running the open weights in ComfyUI or locally

• Use the "Pika H3 Flex: Reference to Video" template in ComfyUI's Template Library (Video). It loads the H3 ref2va base plus the Ref2VA adapter; the text-to-video and first/last-frame templates use the fl2va files instead.
• Keep Pika's six-step schedule (lcm sampler, ManualSigmas, CFG 1.0) and the adapter strength at 1.0. ComfyUI warns not to stack a Turbo or Lightning LoRA on top: the adapter is already distilled, and a second acceleration LoRA breaks the output.
• Pika's own inference package was tested on one H100 80GB per request and downloads about 210 GB of base files plus 5.5 GB of adapters. Long clips and large reference sets can run out of GPU memory, so Pika suggests starting at 480p and 5 seconds and using --cpu-offload when needed.
• There is no prompt rewriter locally. Write the full prompt yourself, or call Pika's prompt rewriter API ($0.02 per request) if you host the model.

Community repacks with extra speed-up LoRAs have appeared on X; they are third-party, not part of Pika's release, and ComfyUI's official notes advise against stacking acceleration LoRAs on the Flex adapter.

What one attempt costs on the API

Reference-to-video launch prices from Pika's pricing page (29 October to 5 November 2026 for Pika API Club members). Image and audio references are free; a reference video is billed per input second at the same rate as the output.

• 10 seconds at 480p (a draft): $0.30
• 15 seconds at 768p with five pack images: $0.90
• 15 seconds at 768p plus a 5-second motion video: $1.20 (20 billed seconds)
• After the launch window the listed regular price doubles at 480p and 768p ($0.06 and $0.12 per second), so the same 15-second 768p clip costs $1.80

Only successful generations are charged, and the API has no free generation tier. Prices change; check Pika's pricing page before a batch.

Pika H3 Flex or MiniMax H3?

Both use the same reference limits and the same pack images, so you can switch without redoing the character. Pika H3 Flex is faster (six steps), finishes up to 4K and returns separate audio stems, and its image references are free. MiniMax H3 gives the first five images free and adds 2K regeneration. Pika's own human evaluation reports Flex leading on visual style and camera framing in reference-to-video, while other models were preferred for action-heavy reference tasks and multi-scene narrative. That is Pika's evaluation, not ours, so test your character on both before a long project.

MiniMax H3

Face drift and other problems

The face changes between clips.

Upload the same images in the same order and paste the same identity line into every prompt. On the API, turning the prompt rewriter off (rewrite: false) keeps your exact wording; it is on by default.

The person from my motion video shows up instead of my character.

Say that @Video1 is the motion reference only and list what not to take from it: face, body, clothes, location, colours. Pika's own showcase prompt does exactly this.

My illustrated character turns photoreal.

Name the art style and tie it to the sheet: "render the whole video in the art style of @Image1, never photoreal". A live-action motion video pulls the look toward photo unless you say so.

The wide shot comes out as a medium shot.

ComfyUI's testing notes say "wide shot" alone tends to render medium. Describe the framing as content instead, for example "she is small in a large empty hall".

Will more sampling steps make the face sharper?

No. Six steps is the trained schedule: according to ComfyUI more steps do not add detail, and a shorter sigma list breaks it. For a sharper face, finish the kept take at a higher resolution instead.