What are Models

Models are the AI systems that do the actual generation. Get3W exposes models from 25 providers behind one API and one addressing scheme.

Model Slugs

Every model is addressed by a three-part slug:

{provider_id}/{model_id}/{run_type}

Examples:

• google/nano-banana-pro/text-to-image
• bytedance/seedance-2/reference-to-video
• minimax/music-2.6/text-to-music
• tripo/p1/image-to-3d

The slug is also the request path, so the model you want is the endpoint you call:

POST /v1/google/nano-banana-pro/text-to-image

The same model_id can support several run types. alibaba/wan-3 for instance handles text-to-video, first-to-video, first-last-to-video, and reference-to-video — the run type you pick determines which inputs are required.

Run Types

Run type	Input → Output
text-to-image	Prompt → image
image-to-image	Image + prompt → image (edit, upscale, restyle)
text-to-video	Prompt → video
first-to-video	First frame + prompt → video
first-last-to-video	First and last frame + prompt → video
reference-to-video	Reference images + prompt → video
edit-to-video	Video + prompt → edited video
extend-to-video	Video + prompt → new continuation segment
digital-human	Portrait + audio → talking avatar video
text-to-speech	Text → audio
text-to-music	Prompt → music
speech-to-text	Audio → transcribed text
text-to-3d	Prompt → 3D model
image-to-3d	Image → 3D model
text-to-panorama	Prompt → 360° panorama
image-to-panorama	Image → 360° panorama
chat	Messages → text (OpenAI-compatible)

Reference, edit, and extend

Three run types take a video as input, and picking the wrong one is the most common mistake. They differ in what happens to the source footage:

Run type	What the source video is for	duration	aspect_ratio
reference-to-video	Style / subject guidance for brand-new footage	Your choice	Your choice
edit-to-video	The footage being rewritten in place	Set by the source, not accepted	Follows the source
extend-to-video	The footage new frames continue from	Length of the new segment	Follows the source

reference-to-video does not reuse the source footage at all — it generates a new video guided by it. Use edit-to-video to change what is in an existing clip, and extend-to-video to keep an existing clip and add to it.

One thing to get right on extend: duration is the length of the new segment, not the combined length, and the response contains only that new segment. Joining it to the original is up to you. Extending a 10s clip with duration: 5 returns 5 seconds of footage, which you concatenate to get 15.

On Seedance 2.5 these map to one upstream endpoint with three task modes, so a couple of constraints follow from the model rather than from Get3W:

• The edit source video must be 4–30 seconds. Shorter or longer is rejected at submit.
• Edit output length always matches the source, which is why duration is not offered. Billing reflects the probed source length, not a fixed 5 seconds.
• Extend accepts duration between 4 and 30 seconds for the new segment.
• Both edit and extend keep the source aspect ratio, so aspect_ratio is ignored.
• Anything with a video input is billed on (source length + output length), so a long source costs more even when you generate only a few seconds from it.

Write the prompt in the voice of the task you picked — "remove the car from the video" for edit, "continue the shot as the camera pulls back" for extend. The model cross-checks your prompt against the run type and fails the task outright when the two disagree.

Providers

Models come from Alibaba, Anthropic, Bilibili, Black Forest Labs, ByteDance, ElevenLabs, Google, Ideogram, JD, Kling, Lightricks, Midjourney, MiniMax, OpenAI, PixVerse, Recraft, Runway, SkyReels, Tencent, Topaz, Tripo, Vidu, xAI, Ace Step, and Get3W's own models.

The provider_id in a slug identifies who made the model, not which infrastructure serves it. A single slug can be served by several upstream platforms, and Get3W fails over between them within one task — see How Get3W Works.

Channels

Many models accept a channel in the request body to select a service tier:

{
  "prompt": "A sunset over mountains",
  "channel": "stable"
}

The common set:

Channel	Trade-off
economy	Cheapest, best for bulk work
stable	Balanced speed and reliability
official	Direct official provider access

Some model families use their own values instead — for example turbo / balanced / quality on Ideogram, turbo / pro on Vidu, pro / fast on Lightricks, and standard / human where a real-person path exists. Omit channel and Get3W applies that model's default.

Channel affects price, latency, and which upstream providers are eligible, so it is worth benchmarking before settling on one.

Parameters

Each model has its own input schema. Parameters that recur across many:

Parameter	Description	Typical use
prompt	Text description of the output	Most models
image_url / image_urls	Input image(s)	Image and video models
video_url / video_urls	Input video(s)	Edit-to-video, extend-to-video, reference-to-video
aspect_ratio	Output framing, e.g. 16:9	Image and video models
resolution	Output resolution, e.g. 1k, 720p	Image and video models
duration	Video length in seconds	Video models
output_format	File format, e.g. png	Image models
seed	Seed for reproducibility	Many models
channel	Service tier	Where supported

TIP

Parameter names, allowed values, and defaults vary per model. Each model's page on get3w.com/models lists its full schema. Sending an unsupported value returns 1401 Invalid Parameter in the task result.

Pricing

Price depends on the model, the channel, and output characteristics such as resolution and video duration. Your account tier can discount it further, per model. Every request is priced before it is queued and your balance is checked, so an underfunded call fails fast with 402 rather than running and overdrawing. See How Pricing Works.

Finding Models

Browse the full catalog with filters at get3w.com/models. Each model page shows its slug, supported run types, channels, parameters, and price.

For chat models specifically, there is an OpenAI-compatible listing endpoint:

curl "https://api.get3w.com/v1/models" \
  -H "Authorization: Bearer YOUR_API_KEY"

This returns only the chat models, matching what OpenAI clients expect from /v1/models. Generation models are not included — use the models catalog on the website for those.

Next Steps

• What are Tasks
• How to Choose the Right Model
• Get Started with API