Alibaba: HappyHorse 1.1

Alibaba HappyHorse 1.1 is a video-generation edition with separately named text-to-video, first-frame image-to-video and reference-to-video profiles. The manufacturer emphasizes improved visual detail, movement, character identity across clips, text stability and audio-visual synchronization. The result is a short video with sound, useful for a product shot, a portrait animation or a small character scene. Select the input mode from what should define the scene: a written concept, an exact opening image or several visual references.

For text-to-video, describe the subject, location, action, framing, camera movement and sound intent. For a short action scene, show a cyclist stopping beside a riverside café while the camera moves gently toward the parked bicycle. Keep the action appropriate to a short clip and describe the final moment. The maker’s current text guide permits up to 5,000 non-Chinese or 2,500 Chinese characters and truncates excess text. Organize the brief so its essential subject and action are explicit rather than buried in repeated adjectives.

First-frame image-to-video uses exactly one image as the opening. Prepare the subject, composition and intended crop in that still, then describe what should move or how the camera should change. The documented image formats are JPEG/JPG, PNG and WEBP, up to 20 MB, with both dimensions at least 300 pixels and an aspect ratio between 1:2.5 and 2.5:1. The output approximates the selected resolution tier while retaining the source image’s proportions. A portrait result therefore starts with a suitable portrait composition rather than an unrelated aspect-ratio instruction.

Reference-to-video accepts one to nine images. The prompt uses [Image 1], [Image 2] and subsequent positions to identify them in the input order. State the role of each reference and name the relevant object or person, so a scene image is not mistaken for a character source. The reference guide accepts JPEG/JPG, PNG or WEBP up to 20 MB, requires the shortest side to be at least 400 pixels and recommends clear references at 720p or higher. Avoid a heavily compressed reference whose identifying detail is already missing.

For a scene with several references, place the person from one image in the room from another while retaining a coat or product supplied by a third. Describe the interaction and camera in the same brief. The 1.1 reference profile highlights stability of subjects, scene style and cinematic motion; inspect the actual result for the intended identity, product markings and spatial placement. Several images give the brief more information, but the review still needs to check which features carried into the video.

The maker lists 480p, 720p and 1080p output for the 1.1 generation modes, with whole-second duration from three to fifteen and MP4 at 24 fps. The API response describes H.264 encoding. Text and reference generation offer shapes including 16:9, 9:16, 1:1, 4:3, 3:4, 4:5, 5:4, 9:21 and 21:9; first-frame generation follows the input image. Choose the shape for the destination and leave room for later captions or layout elements. Inspect the returned video’s actual framing at the intended display size.

The model family’s output includes sound. Give the performance or atmosphere a clear role in the brief, and review any dialogue, effects or background track with the picture. Keep spoken material short enough for the action and pauses rather than compressing a long script into one short shot. Check lip movement, speaker attribution and the actual wording when speech is present. Prompt-language acceptance does not itself define a guaranteed pronunciation or a separate voice-reference input.

The documented maker workflow is asynchronous: a task identifier tracks processing, and a successful result supplies the video location. The result link and task queries are documented with a twenty-four-hour validity period. Retain the completed clip, references and prompt as the revision record. The guides also document seed and watermark controls. A fixed seed can aid comparison, but the maker explicitly says it does not guarantee identical results, so judge the returned artifact rather than assuming one setting fixes every frame.

HappyHorse 1.0 and 1.1 are separate editions. The 1.1 output table adds 480p to the 720p and 1080p options of 1.0, and the 1.1 model pages describe upgrades in detail, motion and consistency. The separately named HappyHorse 1.0 video-edit profile has its own inputs and limits and is not an editing mode to silently assign to 1.1. Review the complete 1.1 clip for identity, motion, framing, text and sound, then make a specific correction while retaining the successful parts of the brief.