Helios 14B Base
A 14B autoregressive video model from PKU-YuanGroup, fine-tuned from Wan 2.1, that covers text-to-video, image-to-video and video-to-video and builds clips up to a minute long in 33-frame chunks.
Overview
Helios 14B Base is one of three released checkpoints of Helios, a 14B video model from PKU-YuanGroup fine-tuned from Wan 2.1 T2V 14B. It is an autoregressive diffusion model: video comes out 33 frames at a time, each chunk conditioned on the frames before it, so clip length depends on how many chunks you ask for.
Text, image and video conditioning share one input representation, which lets the same weights handle text-to-video, image-to-video and video-to-video. The authors report minute-scale clips that stay coherent without the usual anti-drifting heuristics such as self-forcing or keyframe sampling, and attribute this to training that deliberately simulates drift.
The other two checkpoints are Mid and Distilled. The authors rate Base highest of the three for quality; it uses v-prediction with standard classifier-free guidance.
What you can build
- Clips tens of seconds long from a prompt alone, such as a tracking shot or a music-video style performance.
- Motion added to a still, for example a product photo or a piece of concept art, with the prompt describing what moves.
- Video-to-video runs, where an uploaded clip conditions the new footage alongside the prompt.
Inputs
The prompt is the only required field. Like the authors' own examples, the sample prompts here are full paragraphs naming subject, setting, light and camera angle. You can also add one image (JPG, JPEG, PNG, GIF, WEBP or HEIC), a video file and a negative prompt.
Frames default to 512 by 512 and can reach 767 px wide and 768 px tall. Length runs from 1 to 1449 frames (default 99) at 1 to 24 fps (default 20), so the defaults give about 5 seconds of video. Steps go from 1 to 50 (default 40), guidance from 0 to 20 (default 5.0), and a seed can be set.
Outputs
An MP4 video with no audio track. The four samples are 512 by 512 at 20 fps. A galloping horse, a giant tree seen from below and a kangaroo in purple overalls on a rainy Mumbai street each run just under 5 seconds, and a concert scene runs about 36 seconds.
Recommended settings
- Use a frame count that is a multiple of 33, such as 99, 132, 264 or 726. The model card says other values round up to the next multiple, so 1449 becomes 1452 frames, about 60 seconds at 24 fps.
- The authors' Base examples use guidance 5.0, 50 steps, a 640 by 384 frame and 24 fps. The step default here is 40.
- Fill in the negative prompt, which starts empty. The authors' default covers overexposure, static frames, blurred detail, subtitles, malformed hands and limbs, and crowded backgrounds.
Limitations
- Training was text-to-video, so the authors expect image-to-video and video-to-video to be slightly weaker, and the first chunks of those runs can come out static.
- Output tops out at 767 by 768 pixels, so 720p and 1080p are not possible.
- Long clips can still drift. In the 36-second concert sample the dress stays consistent while the singer's face and hair color change from shot to shot.
Safety & compliance
Helios is released under the Apache-2.0 license. If an uploaded image or video shows a real person, make sure you have that person's consent.
Example prompts
Great starting points for Helios-14B-Base.
API quick start
Run Helios-14B-Base with a single API call.
{
"prompt": "A bright yellow Lamborghini Huracn Tecnic…",
"inputVideo": "https://your-cdn.com/input.mp4",
"inputImage": "https://your-cdn.com/input.png",
"negativePrompt": "..."
}curl -X POST "https://api.wiro.ai/v1/Run/bestwishysh/helios-14b-base" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "A bright yellow Lamborghini Huracn Tecnic…",
"inputVideo": "https://your-cdn.com/input.mp4",
"inputImage": "https://your-cdn.com/input.png",
"negativePrompt": "..."
}
JSONOther models you might like
Explore similar video generation models.
