Based on the self-developed video diffusion model, Haiper provides AI video creation capabilities such as text to video, picture to video, keyframe control, video redrawing and motion brushing, supports up to 4K resolution output, and opens API and ComfyUI custom nodes.

haiper

Haiper’s core parameters and statistics

Haiper is a new generation AI video generation platform from London, UK. It was co-founded in 2022 by former DeepMind core members Yishu Miao, Alexandre Davies and Ziyu Wang. Its technical route revolves around the self-developed video diffusion model, and its product form covers the web, iOS and API, establishing differentiated competitiveness in "naturalness of motion" and "consistency between frames".

Projects	Public Information
Official positioning	AI video generation and creation platform
Core Technology	Self-developed video diffusion model (Haiper Video 2.x / 1.5)
Core capabilities	Text to video, picture to video, key frame control, video redrawing, video enhancement and extension, text to image
Output resolution	540p / 720p / 1080p / 4K (depending on the plan and model version)
Video duration	Basic generation is 4~6 seconds, the maximum can be extended
Generation mode	standard (high fidelity), smooth (large motion), enhanced (high fidelity + motion)
Prompt word enhancement	Built-in Prompt Enhancer, supports positive and negative prompt words
Supported platforms	Web, iOS, API (RESTful)
Third Party Integrations	ComfyUI Custom Nodes (Open Source MIT License)
API frequency control	500 HTTP requests per minute, up to 40 concurrent generation tasks
Affiliated Company	Haiper.ai (London, UK)
Founding Team	Former DeepMind and ByteDance AI researcher

Core Difference: Haiper's technical pedigree comes from the DeepMind vision team, which gives it an innate advantage in the "naturalness of motion" of video generation - the facial expressions and body movements of characters are more coherent between different frames, and the frame flicker problem is significantly reduced compared to early video generation models. Compared with contemporary products such as Runway Gen-3 and Pika 2.0, Haiper has a balanced performance in prompt word compliance and style diversity, and the integration of keyframe control and ComfyUI is its unique plus point for professional users.

Haiper’s users and market recognition

Haiper has not disclosed the specific number of users, number of paid subscriptions or list of enterprise customers. Market recognition is mainly reflected in the following dimensions:

Team Background Endorsement: The three co-founders all come from the core teams of DeepMind and ByteDance AI, which provides Haiper with implicit credit in terms of technical credibility - investors and early users are more likely to trust its model path rather than marketing rhetoric.

Reaction from the developer community: Haiper has open sourced the ComfyUI custom node (MIT license) on GitHub. The repository provides four core nodes: Text-to-Video, Image-to-Video, Text-to-Image and Keyframe Conditioning, making it easier for ComfyUI users to directly call the Haiper API in local workflows. This dual-track strategy of "platform products + developer tools" helps it accumulate reputation among the AI ​​creator community.

iOS application market coverage: Haiper has launched the iOS version (Haiper AI Video & Photo Creator) on the Apple App Store, extending its video generation capabilities to mobile terminals and lowering the threshold for mobile creators.

Competitive Landscape: The AI ​​video generation track has gathered strong competitors such as Runway (Gen-3), Pika (Pika 2.0), Kuaishou (Kling), Google (Veo), OpenAI (Sora), and domestic Keling Vidu. As a startup, Haiper has not released comparable data in terms of market share or financing scale, but its "DeepMind series" technology trump card and differentiated functions (key frame control, video redrawing) are the key to establishing its recognition in the cracks.

Haiper’s cost advantage

Haiper's cost structure is divided into three layers: C-side subscription API pay-as-you-go and enterprise customization. Different entrances correspond to different cost logics.

C client/individual users: Adopt a points system + monthly subscription model. Users who register for free can get a certain amount of initial points, and each video generation consumes 1 to 20 points (depending on the model version, resolution and duration). After the points are used up, you can get a small amount of free credits through daily check-in or directly purchase a subscription plan. Typical monthly plans cover hundreds to thousands of points per month, and watermark removal, high resolution, and priority generation queues require payment to unlock. For individual creators, the monthly fee ranges from about $10 to $30, which is significantly lower than the outsourcing or software licensing costs of traditional video production. However, it should be noted that the "unlimited" plan is usually subject to the principle of fair use, and the actual production volume is still limited.

API Developers: The API is billed based on the duration and resolution of the generated video/image, in units of one second:

• Haiper Video 2.x: 720p video $0.05/sec, 540p video $0.033/sec
• Haiper Image 2.x: 720p image $0.04/image, 1080p $0.08/image, 4K $0.12/image
• Video enhancement (1080p upscale): $0.05/second

Developers also need to bear the limit of API call frequency control (500 requests per minute and 40 concurrent tasks). If the limit is exceeded, you need to contact the business for adjustment. Compared with Runway's API pricing, Haiper's per-second billing has certain cost flexibility in short-term and high-concurrency scenarios.

Enterprise/privatization: The enterprise plan requires contacting the business owner for confirmation. The public pricing page provides a Contact Sales entrance. In enterprise scenarios, you need to pay attention to the following hidden costs - whether the concurrent quota of API requests is negotiable, whether the data is only used for current inference and will not enter the secondary training of the model, and whether there is a dedicated SLA to ensure the stability of the generated response time.

Cost Tiers	Payment Methods	Typical Price Ranges	Key Limitations
C-side/individual	Points + monthly subscription	$10～$30/month	Points amount, watermark, resolution limit
API developer	Billed by video duration/number of pictures	$0.033~$0.12/second or/picture	500 requests per minute 40 concurrent
Enterprise/Private	Commercial Quotation	Unpublished	Subject to confirmation of SLA, concurrency and data terms

Main functions of Haiper

Haiper's product capabilities revolve around "multi-modal input-driven video generation", covering the complete link from text to video, picture to video, key frame control, video redrawing to image generation.

• Text-to-Video: Enter a natural language description, and AI will generate a 4-6 second video clip corresponding to the scene. Supports positive and negative prompt words, and the built-in Prompt Enhancer can automatically optimize the prompt word syntax and details. The generation mode provides three options: standard (high fidelity details), smooth (large motion) and enhanced (high fidelity + large motion), adapting to different creative intentions. For users, the focus of acceptance is the restoration of the visual characteristics of the subject in the prompt words and the rationality of movement - a typical rollover scenario is "multi-subject interaction" that leads to chaotic action logic.

• Image-to-Video: Upload a JPG/PNG/WebP picture as the starting frame, and AI generates subsequent dynamic pictures based on the prompt words. Supports Base64 encoding upload, suitable for e-commerce product image animation, illustration conversion to short film and other scenarios. The hidden cost is: the more complex the image content (multiple people, multi-layered depth of field), the greater the dynamic expansion quality fluctuation of the model. It is recommended to give priority to input images with clear themes and simple backgrounds.

• Keyframe Conditioning: Upload 2 to 6 pictures as key frames of the video, specify the position (frame index) of each picture on the timeline, and AI will automatically generate a transition animation between frames. This is Haiper's differentiated capability from most competing products - it is not simply a "picture-based video", but allows creators to accurately control key postures and compositional turning points in the video. The API side exposes this capability through the /v1/jobs/gen2/afc endpoint, and the duration supports two gears: 4 seconds (31 frames) and 6 seconds (47 frames).

• Video Restyle / Video-to-Video: Upload an existing video, and AI will re-render its visual style (watercolor, cyberpunk, oil painting, cartoon, etc.) while keeping the original motion trajectory unchanged. This feature is of great value in scenarios of rapid iteration of advertising creative proposals and concept designs - designers do not need to redraw frame by frame, they only need to select a style template to generate a stylized version. Public pricing information shows that this feature uses the Video 2.x model link, which is the same price as Image-to-Video.

• Video Enhance & Extend: Improve the resolution (up to 1080p/4K) or extend the duration of the generated video. Enhancements are billed standalone at 1080p (10 credits/sec web, $0.05/sec API), and extensions are priced at the same price as the base model. The key points for acceptance are the consistency of the painting style and the smoothness of motion connection in the extended part - style drift is still a common problem in long extended scenes.

• Text-to-Image: Enter the prompt word, and AI will generate 4 images (one group at a time). It supports three resolutions: 720p, 1080p, and 4K. The Haiper Image 2.x model is billed at 5 to 20 points/group (Web) or $0.04 to $0.12/picture (API). This function exists more as an auxiliary module for video generation rather than an independent image generation tool.

Haiper’s model and version evolution

Since its public launch at the end of 2024, Haiper has mainly gone through three version stages:

v1.0 (~2024-12)

The first public version of Haiper. Supports basic text-to-video conversion, the output is 540p/720p resolution, and the video duration is 2 to 4 seconds. The main value of this stage is to verify the feasibility of the product direction of "DeepMind video generation" and allow early users to intuitively experience its advantages in natural motion. The model capabilities are concentrated in single-agent static scenes, and the frame consistency in complex multi-agent scenes is limited.

v1.5 (~2025-08)

Based on v1.0, the ability to convert images to videos and key frame control has been added, marking the product's leap from "pure text generation" to "multi-modal driver". At the same time, API internal testing starts and some developers are invited to access. Movement fluency and prompt word compliance have been significantly improved, and 1080p output has been supported. This version is a transitional node for Haiper from a "novel experience" to a "productivity tool".

v2.0 (~2026-04)

The introduction of Haiper Video 2.x and Haiper Image 2.x dual model series is the largest capability upgrade since the product was released. Core changes:

• Resolution upper limit increase: video output supports 4K (paid version), image generation supports 4K
• Generation mode refinement: Added three new generation modes: standard/smooth/enhanced, allowing users to weigh fidelity and motion range according to the scene.
• Video-to-Video is officially launched to complete the stylized migration capability
• Prompt Word Enhancer is enabled by default to improve Prompt stability
• ComfyUI custom node open source (MIT license), lowering the threshold for professional users to call Haiper in local workflows
• iOS App launched, expanding mobile coverage

Version	Approximate time	New capabilities	Maximum resolution	API availability
v1.0	~2024-12	Text to video	720p	Not yet open
v1.5	~2025-08	Image to video, key frame control	1080p	Internal beta
v2.0	~2026-04	Video redrawing 4K, three generation modes ComfyUI node iOS App	4K	Officially open

Haiper’s technical advantages

Haiper's technical advantages focus on three aspects: the self-developed diffusion model's algorithm leadership in motion naturalness, the unified reasoning architecture for multi-modal input, and its integrability for professional users.

Algorithmic Root Causes of Motion Consistency: Haiper’s technical team comes from the DeepMind Vision Group, and its core papers and technology accumulation revolve around the temporal consistency (temporal consistency) in the video diffusion model. Common pain points of traditional video generation models are inter-frame flicker, subject morphing and motion jitter. By optimizing the temporal attention mechanism and conditional injection strategy, Haiper shows better stability than competing products in the same period in these dimensions. The effect is as follows: the facial expressions of the characters remain consistent in consecutive frames, and the body movement trajectories are smooth instead of sudden changes. This is particularly critical for scenarios that require high identity consistency, such as dynamic display of e-commerce products and short videos broadcast by characters.

Multi-modal driven unified architecture: Haiper does not build independent generation pipelines for text, pictures, and key frames. Instead, it processes the input conditions of different modalities under the same diffusion framework. This means users can mix Text-to-Video, Image-to-Video, and Keyframe Conditioning in a single session without switching models or reasoning contexts. For workflow, this architecture reduces the connection friction caused by "multi-tool patchwork" - for example, first use Keyframe to set the starting and ending postures, and then use Text-to-Video to supplement the intermediate transition description, all within one platform.

ComfyUI integration and developer ecosystem: Haiper has open sourced the ComfyUI-HaiperAI-API custom node on GitHub under the MIT license, providing four core nodes (Text2Video, Image2Video, Text2Image, Keyframe Conditioning). This means that ComfyUI users can insert Haiper's generation capabilities into local complex AI workflows - for example, first use Stable Diffusion to generate character images, then pass them to Haiper's Image2Video node to generate animations, and finally control the transition through Haiper's Keyframe Conditioning node. This hybrid architecture of "local preprocessing + cloud inference" is more efficient than a pure Web UI in professional creation scenarios that require frequent iteration of prompts and parameters.

API Engineering: Haiper API adopts RESTful design, providing standard creation tasks (POST), polling status (GET) and obtain result links. The interface parameters cover key dimensions such as prompt, negative_prompt, seed, aspect_ratio, resolution, duration, gen_mode, etc. It supports is_public to control the visibility of generated content, and is_enable_prompt_enhancer to automatically optimize prompt words. The rate limit is set to 500 req/min and 40 concurrent tasks, which is a moderately loose level among similar APIs. It should be noted that the polling interval of the generated task is recommended to be set to 20 seconds (API document example), otherwise unnecessary frequency control may be triggered.

How to use Haiper

Haiper provides three usage paths, covering ordinary creators, mobile users and developer/automation scenarios:

Entrance	Applicable people	Features	Cost
Web client (haiper.ai)	Individual creators and operators	No installation required, direct use in the browser, visual operation interface	Free initial quota + subscription
iOS App (App Store)	Mobile Creator	Mobile shooting + AI generation, creating at any time	Same as above, including in-app purchases
API + ComfyUI node	Developer AI workflow user	RESTful API or ComfyUI graphical node, which can be integrated into existing pipelines	Pay-as-you-go billing

Get started quickly on the Web: After registering a haiper.ai account, the system automatically allocates initial points. Select Text-to-Video or Image-to-Video on the "Create" page, enter the prompt word (or upload an image), set the resolution and duration, and click Generate. After the generation is completed, you can view, download or share it in "Gallery". Paying users can upgrade their subscription plan on the "Billing" page.

API Quick Access: Obtain HAIPER_KEY after registering a Haiper API account and call it through the RESTful endpoint. The following is a Curl example of Text-to-Video (the parameters are subject to the official documentation):

curl -X POST https://api.haiper.ai/v1/jobs/gen2/text2video \
  -H "Authorization: Bearer <YOUR_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "A golden retriever running on a sunny beach",
    "negative_prompt": "blurry, low quality",
    "settings": {
      "seed": -1,
      "aspect_ratio": "16:9",
      "resolution": 720,
      "duration": 6
    },
    "gen_mode": "standard",
    "is_public": false,
    "is_enable_prompt_enhancer": true
  }'

The response returns generation_id, and polls GET /v1/jobs/{generation_id}/status to obtain the generation status. After the status is succeed, call POST /v1/creation/{generation_id}/watermark-free-url to obtain the non-watermark video URL.

ComfyUI integration: Clone https://github.com/Haiper-ai/ComfyUI-HaiperAI-API.git in the custom_nodes directory of ComfyUI, install dependencies, set HAIPER_KEY in the .env file, and restart ComfyUI to bring up Haiper's four custom nodes in the node panel. The API Key acquisition page is subject to the official real-time entrance.

Haiper product pricing

Haiper's pricing system is based on two parallel tracks: "points" (on the Web/iOS side) and "pay-as-you-go" (on the API side), and the overall structure presents a hierarchical structure.

Web version and iOS version (points system):

• Free users receive initial points (subject to the official registration page), and each generation consumes 1 to 20 points, depending on the model, resolution and duration.
• Free output includes Haiper watermark, watermark removal requires payment
• The paid plan provides fixed points, watermark removal and high-resolution rights based on monthly subscription
• 4K resolution is only available on paid plans, 1080p enhancement is available for an additional charge (10 points/second)

API side (pay-as-you-go):

• Video 2.x model: 720p $0.05/sec, 540p $0.033/sec
• Video 1.5 model: 720p/540p $0.05/sec vs. $0.033/sec
• Image 2.x model: 720p $0.04/picture, 1080p $0.08/picture, 4K $0.12/picture
• Video enhancement (upscale to 1080p): $0.05/second
• Video extension: Same as the base price of the corresponding model

Enterprise Plan: Pricing is not disclosed and needs to be confirmed by contacting the official Contact Sales or [email protected].

Taken together, Haiper's pricing is at a mid-range level in the AI ​​video generation track - lower than Sora (undisclosed but expected to have a high threshold) and Runway's enterprise-level plans, and close to Pika's subscription level. However, for high-frequency API callers, per-second billing means that the marginal cost of long video generation is higher, which is more suitable for batch generation scenarios of short clips of 4 to 6 seconds.

Haiper application scenarios

Haiper's capability boundaries determine that it is most suitable for "short-duration, high visual density, and strong stylization" video creation tasks. The following are four typical scenarios:

• Social media short video material generation: Creators of TikTok/Reels/Shorts can use Text-to-Video to quickly generate background concept videos, transition animations or product display short films. The traditional method requires screening from the material library or shooting by yourself. Haiper's one-click generation can compress the time to obtain a single piece of material from 15 to 30 minutes to 2 to 5 minutes. Key points to check: Whether the visual impact of the first 2 seconds of the video is enough to attract people to stay in the information flow.

• Dynamic display of e-commerce products: The operator uploads the main image of the product (bags, sneakers, cosmetics), and generates a 4-second dynamic display video through Image-to-Video, which can be used for product details pages, circle of friends ads, or information flow delivery. Compared with traditional shooting, the AI ​​solution does not require a studio or model, and the cost of a single video has been reduced from hundreds of yuan to almost zero. Key points of verification: Whether the edges of the product remain clear, whether there is any deformation or distortion, and whether the movement trajectory conforms to real physical logic.

• Advertising creativity and concept design: Designers convert live samples or reference videos into watercolor, cyberpunk or cartoon styles through Video Restyle, and quickly produce multiple versions of creative solutions for customer review. The core value of this scene lies in "multi-style iteration" - traditional methods require frame-by-frame re-rendering or special effects synthesis, Haiper compressed the style switching time from days to tens of minutes. Key points to check: whether the original motion details are retained after redrawing, and the uniformity of style transfer (whether there are local leaks or over-rendering).

• Dynamic online education content: Educational content creators use Text-to-Video to transform abstract concepts ("photosynthesis process", "cell division", "Newtonian mechanics") into short visual videos of 4 to 6 seconds, and embed them in courseware or explanation videos. Dynamic demonstrations significantly help learners understand complex concepts compared to static illustrations. Verification focus: Whether the scientific accuracy meets the standards - the "renderings" generated by AI may deviate from the real scientific illustrations and require manual review before use.

Applicable groups of Haiper

• Social media content creator: Self-media operators who need to produce short video materials at a high frequency. Haiper's Text-to-Video and Image-to-Video can significantly reduce material production costs, but creators need to have basic prompt engineering capabilities to stably output the target screen. It is not suitable for creators who require extremely high video accuracy (such as movie-level image quality).

• E-commerce and Marketing Operations: E-commerce operations and advertising pitchers who need to produce product display videos in batches. Haiper's Image-to-Video is highly efficient in product dynamic scenarios, but the style consistency during batch generation still requires manual screening. Not suitable for product videos that require complex multi-angle presentations (such as 360° rotating presentations).

• Advertising Creative and Designer: Designers who need to quickly produce multi-style video proposals during the conceptual design stage. Video Restyle and Keyframe Conditioning are core value features for this role. Not suitable for final-level video delivery that requires strict brand customization.

• AI Developers & Workflow Builders: Developers who need to integrate video generation into automation pipelines. Haiper's API and ComfyUI nodes provide flexible integration capabilities for this role. Not suitable for users who want to complete complex video narrative editing with "zero code" - Haiper is positioned as a generation module rather than a complete video editor.

Summary and Outlook

Haiper's core competitive barrier is the combination of "DeepMind's motion consistency algorithm + multi-modal input architecture + ComfyUI developer ecosystem". On the AI ​​video generation track, it is not the player with the largest parameters, nor is it the star startup with the highest financing, but it has established a group of loyal users with its differentiated function selection (keyframe control, video redrawing) and pragmatic product path (Web + iOS + API three-terminal coverage).

Current limitations: The video duration is still locked in the basic window of 4 to 6 seconds, and the controllability of the extension function is limited. The generated content exceeding 10 seconds fluctuates greatly in terms of style consistency and physical rationality; the availability of 4K resolution output depends on the complexity of the scene - simple static scenes perform well, but details may still be lost in high-speed motion scenes; API documentation and SDK maturity are slightly insufficient compared to Runway, and error handling and retry strategies need to be improved by developers themselves.

Follow-up focus: Audio synchronization generation (lip-sync, background music generation) is a natural extension of the video generation tool chain. Whether Haiper can complement this will directly affect its competitiveness in the oral short video scene; the openness of the API and the completeness of the documentation will determine whether it can gain long-term reputation in the developer ecosystem; the continuous iteration of mobile apps (such as real-time shooting + AI generation) may become a new growth point.

Not suitable for boundaries: Haiper is not suitable for the following scenarios - video creation that requires more than 30 seconds of continuous narrative (such as short plays, documentary clips), scientific visualization that has rigid requirements for physical accuracy, and commercial video projects that require multi-shot switching and a complete editing process. In these scenarios, traditional video production software or more professional 3D rendering pipelines are still irreplaceable.

Procurement/Adoption Risk Assessment: As a vertical video generation startup, Haiper faces direct competition from competing products of similar size such as Runway and Pika. It also needs to deal with dimensionality reduction attacks from well-funded major players such as Google Veo and OpenAI Sora - the latter has structural advantages in terms of model parameter scale and training data. For purchasers, it is recommended to position Haiper as a "short video clip generation module" rather than a "complete video production platform". First use the free Web-side quota to verify the generation quality and stability in 1 to 2 core scenarios, and then evaluate whether to upgrade to a paid plan or API integration; before purchasing, enterprises need to confirm with the business the data usage terms (will not be used for model training), concurrency quotas, SLA and other key terms.

Related tools: runway, pika

Version Info

• Haiper v2 ：There is no official precise date yet. Introducing the Haiper Video 2.x and Haiper Image 2.x dual model series, which greatly improves video resolution (up to 4K), motion consistency and prompt word compliance. （~2026-04）
• Haiper v1.5 ：There is no official precise date yet. Added image conversion and keyframe control capabilities, optimized motion smoothness, and launched API internal testing. （~2025-08）
• Haiper v1 ：There is no official precise date yet. The first public version of Haiper supports basic text-to-video conversion and 540p/720p output. （~2024-12）