探索 AI 模型

[Core Function] Veo 3.1 I2V is Google's cinematic image-to-video generation model. [Strengths] It generates high-fidelity 4K video from a starting image. It supports advanced features like first-and-last frame conditioning and referencing up to three images. [Best For] Highly recommended for: animating concept art, creating cinematic transitions between images, and high-end video production. [Limitations] When using first+last frame or reference-only modes, the `duration` parameter must strictly be 8. Negative prompts are not supported in reference-only mode. [Routing] Use this model by default for high-quality image-to-video tasks or when multiple reference images are provided.

[Core Function] Seedream 5.0 Lite Edit is a reasoning-enhanced, smart image editing model. [Strengths] It features superior cross-modal understanding and reasoning, allowing for highly accurate, interactive multi-turn image editing with real-time knowledge enhancement. [Best For] Highly recommended for: complex image editing tasks, structural modifications, and edits requiring deep semantic understanding. [Limitations] As a 'Lite' model, raw visual rendering might not match the 4.5 tier. [Routing] Use this model by default for complex, reasoning-based image editing tasks.

[Core Function] Nano Banana 2 Edit is a high-speed image editing model. [Strengths] It rapidly modifies existing images or extracts image frames from videos based on text prompts. [Best For] Highly recommended for: rapid style transfer, quick image modifications, and fast creative edits. [Limitations] Do NOT use this model for meticulous photorealistic retouching. [Routing] Use this model by default for fast, creative image editing tasks.

[Core Function] Seedance 2.0 I2V is ByteDance's flagship unified multimodal video generation model. [Strengths] It supports complex mixed references (multiple images, audio clips) and generates up to 15s of multi-shot audio-video output with dual-channel audio. [Best For] Highly recommended for: high-end complex video generation, multi-shot narratives, and mixed-reference cinematic production. [Limitations] Do NOT use this model if you only need a very basic legacy generation without complex references. [Routing] Use this model by default for any complex, multi-reference, or high-fidelity image-to-video tasks.

[Core Function] Seedance 2.0 V2V is ByteDance's flagship multimodal video-to-video model. [Strengths] It allows powerful editing and stylization of input videos by supporting mixed references (text, images, video, and audio) and producing multi-shot 15s outputs. [Best For] Highly recommended for: complex video-to-video transformations, restyling existing footage, and creating dynamic multi-shot edits. [Limitations] Do NOT use this model for simple still-image generation (use Seedream instead). [Routing] Use this model by default for any video editing or video-to-video generation tasks.

[Core Function] Grok Imagine Video 1.5 I2V animates a single starting image into a video using the Grok Imagine 1.5 generation backbone. [Strengths] It excels at producing motion from one starting frame with the improved 1.5 model. [Best For] Highly recommended for: animating a photo or illustration when the 1.5 generation backbone is preferred. [Limitations] Do NOT use this model for text-only generation, for resolutions above 1080p, or for clips longer than 15 seconds; it requires a starting image and is limited to 480p/720p/1080p and 15s. [Routing] Use this model when the user provides one starting image and prefers the 1.5 backbone.

[Core Function] Grok Imagine Image Edit (Quality) is xAI's high-fidelity image editing model. [Strengths] It excels at applying detailed, prompt-guided edits and style transformations to one or more source images while preserving fine detail. [Best For] Highly recommended for: high-quality restyling, detailed inpainting-style edits, and combining up to 3 source images. [Limitations] Do NOT use this model when latency is critical, as it is slower than the standard edit variant; a maximum of 3 source images is supported. [Routing] Use this model by default for quality-sensitive edits. For faster edits, route to Grok Imagine Image Edit (standard).

[Core Function] Wan 2.7 Video Editing is an instruction-based video modification model. [Strengths] It supports complex video editing tasks like content replacement using reference images, and replicating actions, effects, and camera movements. [Best For] Highly recommended for: modifying existing video footage, style transfer on videos, and targeted element replacement. [Limitations] Do NOT use this model to generate a brand new video from scratch; it requires an input video. [Routing] Use this model by default whenever a user wants to edit, alter, or restyle an existing video.

**[Core Function]** SkyReels Lip Sync (retalking) re-drives a talking video so the subject's lips match a given audio track. **[Strengths]** Accurate lip re-synchronization on an existing talking-head video. **[Best For]** Dubbing, re-voicing talking-head video, and localizing spoken video. **[Limitations]** Do NOT use this to generate new motion or content from scratch; it only re-times lips on an existing video. Requires video_url and audio_url. Output resolution is fixed at 720p. **[Routing]** Provide the source video_url and the target audio_url; optionally provide reference_char_url to guide the driven face.

**[Core Function]** SkyReels Shot-Switching Extension continues a video while transitioning to a new shot or camera angle. **[Strengths]** Cinematic shot transitions (cut-in, cut-out, reverse-shot, multi-angle, cut-away) when extending footage. **[Best For]** Adding a new shot after existing footage and cinematic transitions. **[Limitations]** Do NOT use this for a plain single-shot continuation (use video-extension-single-shot) or for generation from scratch. Requires a prefix_video (mp4 URL); adds 2-5s. **[Routing]** Provide prompt and prefix_video; choose cut_type for the transition style, or Auto to let the model decide.

**[Core Function]** SkyReels Single-Shot Extension continues an existing single-shot video, generating additional seconds guided by a text prompt. **[Strengths]** Seamless single-shot continuation of the existing motion and scene. **[Best For]** Lengthening clips and continuing an action within one continuous shot. **[Limitations]** Do NOT use this to create a video from scratch (use skyreels-t2v) or to switch shots / add transitions (use video-extension-shot-switching). Requires a prefix_video (mp4 URL); adds 5-30s. **[Routing]** Provide prompt and prefix_video; set duration for how many seconds (5-30) to append.

**[Core Function]** SkyReels Restyle re-renders an existing video into a preset visual style. **[Strengths]** Consistent style transfer across all frames into a chosen named art style. **[Best For]** Turning footage into simpsons, lego, paper-cutting, amigurumi, animal-crossing, van-gogh, or pixel-art looks. **[Limitations]** Do NOT use this to change content, motion, or add new scenes; it only restyles an existing video (input <=30s). Output resolution is fixed at 720p. **[Routing]** Provide the source video_url and a style_name from the supported list.

**[Core Function]** SkyReels Omni is a reference-driven video model that generates or edits video using reference images (@image) and/or a reference video (@video), bound by tags in the prompt. **[Strengths]** A single endpoint covers motion reference, subject/background replacement, object insertion/removal, local editing, and video extension. **[Best For]** Video subject or background swap, motion transfer onto an image, object add/remove, local video edits, and extending a reference video. **[Limitations]** Do NOT use this for pure text-to-video (use skyreels-t2v) or simple single-image animation (use skyreels-i2v). Each ref tag must appear in the prompt as @tag; ref_videos supports only one video (<=15s); a reference-type video may combine only with image-type ref_images, while an extend-type video cannot combine with ref_images. When ref_videos is provided, aspect_ratio is ignored (output matches the video). **[Routing]** Provide ref_images (type grid or image) for image references and/or a single ref_videos entry (type reference for motion/edit, type extend for continuation); the reference tags must be used in the prompt.

**[Core Function]** SkyReels Segmented Camera Motion (audio-to-video) generates a talking-avatar video with directed camera movement across time segments. **[Strengths]** Combines an audio-driven avatar with per-segment camera trajectories such as push, pan, crane and rotation. **[Best For]** Dynamic presenter clips and cinematic avatar shots with controlled camera motion. **[Limitations]** Do NOT use this when you need a completely static camera (use single-actor-avatar). It requires first_frame_image and one audio segment; use camera_control_pro for multi-segment or compound moves. mode=std outputs 720p, mode=pro outputs 1080p. **[Routing]** Set a single traj_type plus camera_control_strength for a simple move, or supply camera_control_pro (a list of per-segment {start_time, end_time, traj_type, ...}) for compound motion; choose mode=pro for 1080p.

**[Core Function]** SkyReels Single-Actor Avatar (audio-to-video) drives a talking-avatar video from a single portrait image and one audio track. **[Strengths]** Lip-synced single-speaker talking-head video generated from an image plus audio. **[Best For]** Virtual presenters, single-speaker dubbing, and talking avatars. **[Limitations]** Do NOT use this for multi-speaker scenes (use the multi-actor flow) or when you only have text. It requires first_frame_image and exactly one audio segment (<=200s). mode=std outputs 720p, mode=pro outputs 1080p. **[Routing]** Provide a portrait first_frame_image and one audio URL in audios; choose mode=pro for 1080p output.

**[Core Function]** SkyReels Reference-to-Video (multiobject) generates a video from a prompt while preserving the subjects from 1-4 reference images. **[Strengths]** Multi-subject identity preservation, placing specific characters or objects into a newly generated scene. **[Best For]** Putting given characters/products into a new scene, multi-subject composition from reference photos. **[Limitations]** Do NOT use this to animate a single fixed frame (use skyreels-i2v) or for text-only generation (use skyreels-t2v). It requires 1-4 reference images and produces clips up to 5s. **[Routing]** Provide 1-4 subject reference images in ref_images; set aspect_ratio and duration (1-5s) as needed.