新模型：OmniHuman 2、Meshy T2.5、Krea 3 及更多

Six new models cover image generation and editing, video with audio, alpha mattes from video and clean 3D meshes.

Image Generation & Editing

Krea 3: Generate high-quality images from text at 1K or 2K resolution. Choose the Berry variant when image quality is the priority, or the Jam variant for speed.
Krea 3 Edit: Edit or combine up to eight images from text instructions. Set an optional high reasoning level that plans complex edits and keeps unedited areas intact. Output is 1K or 2K resolution.

Minimax H3 Fast: Generate videos from 5 to 15 seconds long from text, first and last frame images, or multimodal references. Every generation includes native stereo audio, at 480P or 768P and 24 FPS. Combine up to 12 reference files across images, video and optional audio to guide the characters, motion and style.
OmniHuman 2: Animate a person from a single image so they speak or sing along to an audio track. Supply an optional prompt to direct the scene, movements and camera. The model outputs native 720p video and supports durations from 30 to 300 seconds.
LTX-2.5 Alpha Gen: Generate a frame-aligned grayscale alpha matte from an RGB video, with no text prompt and no green screen. The matte holds soft and partially transparent edges, capturing hair, smoke, fire and glass. White stays opaque, black turns transparent and gray gives partial opacity. Input goes up to 1920x1088 and 145 frames.

Meshy T2.5 Smart Topology: Generate clean, production-ready 3D meshes from text prompts or images, with no secondary remesh pass. Set the target polycount during generation, up to 30,000 faces, to balance asset quality and rendering performance. Output is triangle topology with sharper detail from your source material.