FriendsforHumanity/Pika-H3-Flex-FRef2VA

Pika H3 Flex — FRef2VA (first/last frame + Ref2VA references)

Powered by MiniMax H3. Pin a clip's first frame, last frame, or both, while keeping Pika H3 Flex Ref2VA's image, video and voice references, in one request. Code only: no new weights.

Why

Pika ships first/last frame (FL2VA) and references (Ref2VA) as two separate adapters, and its diffusers pipeline runs one or the other. Chaining clips needs both: each clip should start exactly where the last one ended, and keep the same faces and voices. With FL2VA alone the cast drifts; with Ref2VA alone the clips don't join.

How it works

No merging or retraining. H3 reads text, references, pinned frames and the clip being made as one sequence. A pinned frame is a clean block placed at frame 0 or the last frame and held fixed while sampling. The Ref2VA model already follows pinned frames placed beside its references (ComfyUI does this with MiniMaxH3AddGuide next to MiniMaxH3ReferenceToVideo). These blocks build the same sequence in diffusers:

Part of the sequence	From
Text, references, target audio and video	Pika's own Ref2VA layout, unchanged
Pinned first and/or last frame	added beside the references, at the same positions ComfyUI uses

tests/test_cpu.py checks both rows on CPU in a few seconds, with no weights: without pinned frames the sequence equals Pika's, and with them it equals ComfyUI's.

Usage in ComfyUI

Stock nodes only; three API workflows are in comfyui/ (first, last, first + last, each with two image references and a voice):

• UNETLoader → the H3 Ref2VA checkpoint
• MiniMaxH3ReferenceToVideo with your references, plus MiniMaxH3AddGuide for each pinned frame (frame_idx 0 for the first, -1 for the last)
• LoraLoaderModelOnly → pika_h3_flex_ref2va_comfy_bf16, strength 1.0
• Sampler lcm, BasicGuider (CFG 1.0)
• ManualSigmas: 1.0, 0.972973, 0.923077, 0.8, 0.571429, 0.333333, 0.0 (6 steps)

Each workflow saves the clip and its last frame, ready to pin as the next clip's first frame.

Usage in diffusers

pip install -r requirements.txt
python example.py --prompt prompt.txt --first start.png --last end.png \
    --reference character.png --reference place.png --voice voice.wav --out clip.mp4

Give --first, --last or both; --chain 3 makes three clips, each starting on the previous clip's last frame. In Python, load the blocks with ModularPipelineBlocks.from_pretrained("FriendsforHumanity/Pika-H3-Flex-FRef2VA", trust_remote_code=True) and pass image= and/or last_image= beside references= (see example.py).

Requires diffusers at Pika's commit d0b9b1b5, the one these blocks were tested on, and Pika-Labs/Pika-H3-Flex revision 3ff16644 (both pinned). Before upgrading either, run python tests/test_cpu.py; if it fails, stay on the pins. Voice references may total at most 15 s (Pika's limit).

Testing so far

On one B200, 15 s clips at 480p, four stories with two image references and a voice each. One, Maddox, comes from a 54-clip story made in the Daydream World app.

• Every clip landed on its given frames, first, last or both, within 2–4/255 (mean absolute difference).
• The ComfyUI version ran the whole 54-clip, 13.5-minute Maddox story with no burn, softening or darkening; every clip stayed within 2.4/255 of its pinned frame.
• About 140 s per clip in diffusers (CPU offload), 81 s in ComfyUI (everything on the GPU).

Known limits: faces from small or crowded references can drift in close-ups; motion and sound restart at each join; large changes between the first and last frame are untested; only 480p and 768p were tried.

License and notices

This code: Apache-2.0 (see LICENSE). It downloads, and needs, MiniMax H3 and the Pika H3 Flex adapters, which are under the MiniMax H3 Community License Agreement, its Acceptable Use Policy and territorial scope. Check that licence before running them.

• MiniMax H3 © 2026 MiniMax. All Rights Reserved.
• Pika H3 Flex adapters © 2026 Mellis, Inc.
• FRef2VA blocks and workflows © 2026 Friends for Humanity.