HappyHorse 1.1 Reference-to-Video
HappyHorse 1.1 Reference-to-Video is Alibaba's next-generation AI reference-guided video generation model. Supporting 1 to 9 reference images (covering character identity, outfits, product silhouette, and scene styles), it achieves high-precision multi-image fusion and subject locking to prevent visual drift across shots. It generates native 720P/1080P HD videos up to 15 seconds long per run.
Related Models
HappyHorse 1.1 Reference-to-Video Spicy
HappyHorse 1.1 Reference-to-Video Spicy is Alibaba's high-expressiveness, high-dynamism variant in the HappyHorse model series. Building upon multi-image subject locking and native audio-visual synchronization, the Spicy mode is a uncensored and fine-tuned for high-intensity action, aggressive camera tracking, strong visual impact, and dramatic visual effects. It leverages reference images to drive bold, highly expressive motion sequences and complex camera maneuvers while maintaining multilingual lip-sync and ambient audio—ideal for action-packed short dramas, viral social video, game CG FX, and high-impact commercial ads.
HappyHorse 1.1 Text-to-Video
HappyHorse 1.1 Text-to-Video is Alibaba's next-generation AI text-to-video model. Built on an integrated audio-video joint generation architecture, it generates native 720P/1080P HD videos directly from text, supporting up to 15 seconds of rendering. It natively supports multilingual lip-sync, ambient sound effects, and audio-visual synchronization without extra dubbing, while delivering exceptional motion smoothness, subject consistency, and camera control—ideal for short dramas, commercial ads, and social media video production.
HappyHorse 1.1 Image-to-Video
HappyHorse 1.1 Image-to-Video is Alibaba's next-generation AI image-to-video model. Supporting first-frame driving, first-to-last frame transitions, and short video extension, it generates native 720P/1080P HD videos up to 15 seconds per run. Built on an integrated audio-video joint generation architecture, it natively supports multilingual lip-sync, ambient sound effects, and audio-visual synchronization, while delivering exceptional motion smoothness, subject consistency, and camera control—ideal for e-commerce, short drama VFX, and social media production.
HappyHorse 1.1 Image-to-Video Spicy
HappyHorse 1.1 Image to Video Spicy transforms a single starting image into a short cinematic video, combining rock-solid temporal consistency with smooth, highly expressive character movement.
Read Me
HappyHorse 1.1 Reference-to-Video API
HappyHorse 1.1 is Alibaba's next-generation multimodal video generation model released on June 22, 2026, built on a 15B-parameter unified multimodal Transformer architecture (Transfusion single-stream). It supports four modes: text-to-video (T2V), image-to-video (I2V), reference-to-video (R2V), and video editing. This document covers its Reference-to-Video endpoint: generating high-quality video from multiple reference images plus a text prompt, supporting up to 9 reference images for character, product, and style consistency control, with native audio generated in sync.
On iCreat, developers call the ali/happyhorse-1-1/reference-to-video endpoint through an asynchronous task-based REST API. Pricing: 720P $0.14/sec, 1080P $0.18/sec, billed by output duration.
Model Positioning
HappyHorse 1.1 Reference-to-Video targets multi-reference-driven video creation — the standard endpoint in the HappyHorse 1.1 family for multi-image reference-to-video. Unlike T2V (pure text) and I2V (single-image animation), it accepts up to 9 reference images simultaneously to lock character faces, clothing, product details, brand elements, and scene style. It is ideal for e-commerce livestream-style videos, multi-character short dramas, and other tasks demanding strong subject consistency.
Core Capabilities
Multi-reference consistency control
Accepts 1–9 reference images simultaneously. The model fuses characters, products, clothing, and background elements from multiple references with a text prompt describing scene and action, producing videos with consistent style and controllable characters.
Enhanced dynamic expression
Optimized motion modeling and temporal consistency over version 1.0 — resolves sluggish pacing and weak rhythm, significantly improving continuity and power in complex action scenes (fighting, dancing, running, vehicle movement, tracking shots).
Native audio generation
Audio and video are generated together (not added in post). Dialogue speech rate, pauses, and emotional tone adjust dynamically to the scene; background music and ambient sound can be described in the prompt. Audio-visual sync precision is further improved.
Visual texture refinement
Mitigates the "oily" and "over-sharpened" look reported in 1.0, preserving real skin details (blemishes, smile lines, pores) while reducing the smeared appearance. Enhanced understanding of professional camera language (shot/reverse-shot, tracking).
24 FPS cinematic output
Supports 720P / 1080P resolution, 3–15 second single clips, 24 FPS, flexible aspect ratios (1:1 / 9:16 / 16:9 / 3:4 / 4:3).
Pricing
| Resolution | Unit Price (USD/sec) | 5-Second Cost |
|---|---|---|
| 720P | $0.14 | $0.70 |
| 1080P | $0.18 | $0.90 |
Total = unit price × output video duration. Billed by actual generated duration with no minimum-length threshold.
Use Cases
- E-commerce and livestream-style videos: combine spokesperson, product, outfit, and livestream-room reference images into a consistent short ad clip
- Multi-character short dramas: lock faces and outfits with multiple character references, generate plot segments from scripts
- Brand promotional clips: product + brand-element reference images for style-consistent brand videos
- Character dynamic testing and pre-visualization: generate dynamic shots from character references for pre-production
- Social media creative: multi-reference generation of vertical (9:16) short-form content
Model Comparison
HappyHorse 1.1 R2V vs. HappyHorse 1.1 T2V
| Dimension | R2V | T2V |
|---|---|---|
| Endpoint | ali/happyhorse-1-1/reference-to-video |
ali/happyhorse-1-1/text-to-video |
| Input | Reference images (1–9) + text prompt | Text prompt only |
| Reference images | Supported (up to 9) | Not supported |
| Subject consistency | Strong (multi-image locking of characters/products/style) | Text description only |
| Resolution | 720P / 1080P | 720P / 1080P |
| Duration | 4–15 seconds | 3–15 seconds |
| Frame rate / audio | 24 FPS / native audio | 24 FPS / native audio |
| Aspect ratios | 1:1 / 9:16 / 16:9 / 3:4 / 4:3 | 1:1 / 9:16 / 16:9 / 3:4 / 4:3 |
| Best for | Multi-reference consistency, e-commerce, multi-character drama | Pure text-driven creative generation, no reference visuals |
HappyHorse 1.1 R2V vs. HappyHorse 1.1 I2V
| Dimension | R2V | I2V |
|---|---|---|
| Endpoint | ali/happyhorse-1-1/reference-to-video |
ali/happyhorse-1-1/image-to-video |
| Input | Multiple reference images (1–9) + text prompt | Single first-frame image + text prompt |
| Reference image count | Up to 9 | 1 (first frame) |
| First-frame control | Not applicable (multi-image fusion) | Supported (image animation) |
| Subject consistency | Multi-character / multi-element consistency | Single-image subject preservation |
| Resolution | 720P / 1080P | 720P / 1080P |
| Duration | 4–15 seconds | 3–15 seconds |
| Best for | Multi-character drama, multi-product e-commerce, brand-element locking | Single-image animation, product display, artistic creation |
Why Choose HappyHorse 1.1 R2V?
- Up to 9 reference images: lock multiple characters, products, outfits, and scene elements in one pass
- Native audio sync: visuals and sound generated together; dialogue rate/pauses/tone adapt to emotion
- Cinematic dynamics: fighting, dancing, and tracking shots feel natural and continuous — no slow-motion artifacts
- Visual texture upgrade: reduced oily/over-sharpened look, real skin detail preserved
- Transparent per-second billing: 720P $0.14/sec, 1080P $0.18/sec, predictable cost
API
This interface is an asynchronous task-based REST API. Submit endpoint: https://api.icreat.ai/v1/task/submit/ali/happyhorse-1-1/reference-to-video. The flow has three steps: POST the task parameters to the submit endpoint and receive a task_id; poll the query endpoint with the task_id until the status is SUCCEEDED; then call the result endpoint to fetch the generated video.
The request body uses an input + parameters structure. input contains prompt (required), negative_prompt (optional), and media (reference media array, each element with type and url). parameters contains resolution (480P/720P/1080P), ratio (aspect ratio), duration (integer 4–15), and watermark (default false).
Specifications
| Item | Description |
|---|---|
| Endpoint | ali/happyhorse-1-1/reference-to-video |
| Authentication | API Key (Authorization: Bearer) |
input.prompt |
string, required, prompt describing the video to generate |
input.negative_prompt |
string, optional, content that should not appear in the video |
input.media |
object[], optional, reference media list |
input.media[].type |
string, required, media type (use reference_image for first-frame image) |
input.media[].url |
string, required, reference image URL |
parameters.resolution |
string, required; 480P, 720P, 1080P |
parameters.ratio |
string, required; 1:1, 9:16, 16:9, 3:4, 4:3 |
parameters.duration |
integer, required; 4–15 seconds |
parameters.watermark |
boolean, optional; default false |
| Output | Video file; 24 FPS, native audio, up to 15 seconds |
| Model specs | 15B, unified multimodal Transformer (Transfusion single-stream), 40 layers |
Architecture
HappyHorse 1.1 R2V uses an asynchronous task-queue architecture. Requests enter the task queue via the submit endpoint and are processed by the HappyHorse 1.1 unified multimodal generation system. The model uses a Transfusion single-stream Transformer architecture (15B parameters, 40 layers), processing text, image, video, and audio tokens in one model rather than stitching separate modules. Reference images are passed via the media array; the model fuses character/product/style features from multiple references and generates a video frame sequence with native audio guided by the text prompt. Inference uses DMD-2 distillation (8 denoising steps) with CFG removed for efficiency — a 5-second 1080P video takes about 38 seconds on a single H100.
Notes
durationmust be an integer between 4 and 15; out-of-range values return an errorpromptis required;negative_promptis optional, used to exclude unwanted content- Reference images should have a short side greater than 400 pixels; keep multi-image aspect ratios consistent for better consistency
- Supported image formats: JPEG/JPG/PNG/BMP/WEBP; single image ≤20MB; side length ≥300 pixels; aspect ratio between 2:5 and 5:2
resolutionsupports480P/720P/1080P, but the pricing table only lists 720P and 1080P; 480P billing is subject to the platform- 1080P costs more per second than 720P; estimate cost in advance for long videos (total = unit price × duration)
- Fetch results via the result endpoint promptly after completion
FAQ
What is the difference between R2V and T2V?
R2V accepts up to 9 reference images to lock character/product/style consistency; T2V takes only a text prompt for creative generation without reference visuals. Both share the same resolution, frame rate, and audio capabilities.
What is the difference between R2V and I2V?
R2V supports multiple reference images (1–9) for fusion-based generation, suited for multi-character/multi-product scenes; I2V takes a single first-frame image and animates it. Choose R2V for multi-element consistency, I2V for single-image animation.
How many reference images are supported?
Up to 9. Recommended short side greater than 400 pixels; keep multi-image aspect ratios consistent for best consistency.
How is billing calculated?
Billed by output duration: total = unit price × output video duration. 720P is $0.14/sec and 1080P is $0.18/sec; a 5-second clip costs $0.70 and $0.90 respectively.
Is native audio supported?
Yes. HappyHorse 1.1 generates audio and video together. You can describe dialogue, background music, and ambient sound in the prompt; speech rate/pauses/tone adapt dynamically to the scene emotion.