Pika 2.0 (Official)

Par Pika Labs

Dibināts 2022 · Palo Alto, USA

Pika Labs was founded in 2022 by Demi Guo (CEO) and Chenlin Meng (CTO), both Stanford PhD candidates in generative AI (Meng is a co-author of foundational diffusion-model papers including SDEdit and Diffusion ODE samplers). The company shipped its initial Discord-based product in mid-2023, launched the Pika web app and Pika 1.0 in November 2023, and rolled out Pika 2.0 in late 2024 with Scene Ingredients (compose a clip from several reference images) and Pikadditions / Pikaffects (object insertion and visual effects). Investors include Lightspeed Venture Partners, Spark Capital, Homebrew and Greycroft, with total funding above $135M and a $470M valuation in 2024. Pika is widely used by short-form social-media creators.

Arhitektūra

Latent video diffusion model (transformer denoiser)

Pika 2.0 is a closed latent-video-diffusion model with a transformer-based denoiser operating on a learned spatio-temporal latent space. Conditioning includes a multilingual text encoder, image embeddings for image-to-video, and the headline 'Scene Ingredients' system that composes a single output clip from multiple reference images of distinct subjects, props and scenes -- a richer multimodal-conditioning surface than Pika 1.0. The model generates 5-second clips at 720p with optional upscaling, and supports special-effect modes (Pikaffects: melt, squish, levitate, etc.) trained as conditioned variations of the base diffusion model. Pika has published no formal paper; product behaviour and team papers (e.g., Meng's diffusion sampler work) suggest a DiT-style architecture with classifier-free guidance and a curated training corpus of stock and web video.

Parametri

Undisclosed

Iespējas

• Text-to-video and image-to-video
• 'Scene Ingredients' multi-image conditioning (subjects, props, scene)
• Pikaffects / Pikadditions for one-shot visual effects (melt, squish, etc.)
• 5-second 720p clips with upscale
• Web app and Discord workflow
• Strong creative-effects vocabulary aimed at social-media
• Easy UX for non-technical creators
• Affordable consumer pricing
• Best for: short-form social video, creative effects, meme content, ad ideation.

Apmācība un licence

Closed corpus of licensed stock footage, web video and synthetic data with dense captions; exact size undisclosed.

Licence: Proprietary commercial licence via pika.art terms; commercial use on paid plans.

Drošības pārbaudes: NSFW, CSAM, public-figure and trademark filters at prompt and output stages; basic moderation playbook with no public RSP.