We’re proud to announce unlimited video generation with no per-use “token” or “credits” pricing. Unlimited video (Wan 2.1, SkyReels, HunYuan and LTX Lighttricks) is available today for Graydient Plus Video members which also includes unlimited images with Flux, unlimited SDXL lora training, unlimited bot creation, and unlimited LLMs like DeepSeek and Llama 3.3 and many other popular models. It’s the best value in AI, bar none!
How to create AI videos
This feature works best with our chat app PirateDiffusion, in both private and group channels. Here’s a quick overview of the video commands, along with prompt templates.
The easiest way to learn how to create videos is by copying others’ published videos and prompts, and there’s already thousands of examples! Members should join the PirateDiffusion PlayRoom Telegram channel to see hundreds of example videos already created by our community.
UNLIMITED WAN 2.1 VIDEO GENENERATOR
Look in our Workflows for unlimited Wan 2.1 video generation in both the 1.3B and 14B models – your choice of Q4, Q6, Q8 optimized versions. The smaller number means a more compressed model, which is less beautiful but creates longer videos. Use the Q8 to take the quality to the limit. Videos complete in just 200 seconds (more or less) depending on the prompt.
WAN 2.1 IMAGE TO VIDEO IS ALSO HERE
You can upload a photo and transform it into a video easily with our new animate-wan21 workflow, give it a try! It’s great for both realistic photos and also anime/illustrations
Fast SkyReels Hunyuan Video hosted online (new!)
We can now transform your photographs into moving images with HunYuan-based SkyReels video. The aspect ratio of the video depends on your input photo, so please crop accordingly using the photo editor tool on the My Graydient dashboard before creating your video. Prompting is easy and natural, as shown below.
You can upload a real photograph, or generate one with our Flux workflows like this one:
Then “reply” to the image with the animate-skyreels workflow and your prompt, like this:
/workflow /run:animate-skyreels The video shows a young woman standing in front of a beige wall. She is wearing a white crop top and light blue jeans with ripped details on the knees and thighs. She has short dark hair styled in a bob with bangs and is wearing white sneakers. The woman is posing with one hand on her hip and the other resting on her thigh. She appears to be confident and relaxed. The camera is zooming out
Optional features – Developer API specification
The following parameters are is intended for developers that want to bring video into their own apps. You don’t have to learn any of this to use our video product, but for the adventurous enthusiast we think you’ll find it interesting!
HUNYUAN LORA VIDEO TIPS
The new workflow ‘hunylora‘ now supports up to six video loras and weights at once. You can use more than one, but loras are seldom made by the same person, and subject to style and aspect ratio conflicts. These are open source models, after all. Try 1-2 at first for starters. In oury experience, giving the character a weight 1.2 and the pose .7 or less works well.
Use the concepts browser in your MyGraydient dashboard to see all of the available Hunyuan aka HUNY loras. Only HUNY concept types can be used: it does not support SDXL or Flux loras, it will treat them like a token.
PROMPTING
Unlike LTX, Hunyuan is very easy to prompt for. It seems to have a short context window, so word order is very important. Put the most important things at the front of the prompt, don’t waste tokens on things like ‘masterpiece’ at the front.
Resolution
/size – Control the resolution like /size:480×320. This workflow cannot handle 720p or 1080p yet. We advise testing around 320×320 or 320×512 at first to see if it will understand your prompt before going bigger. A video upscaler to go much higher is on our roadmap.
Length
/slot1 controls how long the video is. Hunyuan is locked at 24 frames a second. A safe number is around 73 frames but you can go as high as ~200 at low resolution, depending on how big the loras are, etc.
Guidance and Model Sampling Shift
/guidance is what controls the prompt adherence. The default is set to 10. Going lower makes the video look softer and gives the AI more freedom to interpret your ideas, whereas higher guidance may stick closer to your prompt in a more literal sense, but may render graps or glitch if it cannot dream up the specific thing that you requested. You can use one digit of decimal precision like 7.5
/slot2 is mapped to Model Sampling shift, which is a new parameter creates a small but noticeable change in image stabilization and variability. The default is set to 17. Experiment between 0-30. It’s interesting to adjust this and guidance together.
/steps is of course what controls the overall frame image quality – recommended around 10-12 but feel free to go higher at lower rest and shorter videos. Stay under 30 steps to avoid timeouts
/strength controls denoise. Its set to 1 by default but 0.5 – 0.9 can produce nice results too.
/slot3 controls the frames per second. A number between 8 (anime) to 24 (realistic) is a good choice. Hunyuan tops out at 24, so anything higher will look fast-forwarded.
VAE CONTROLS
These parameters control the speed vs. quality when creating each keyframe. Going too high will make the video time out, fair warning. (An unhandled error message will now tell you if it timed out, look for that if the video doesn’t come back)
/slot4 = VAE Tile size. Think of this like the size of the tiles in making a mosaic of each video keyframe. The default is 64. You can lower it and go as high as 128 (maybe). It is a question of speed/efficiency to balance with your other settings
/slot5 = VAE Overlap. How many pixels should it feather as it moves across each keyframe? The default is set to 32, you can try 0 and 64 alternative
/slot6 = Temporal Size – The amount of frames to decode at a time. The default is 12, you can go up in 4 frame increments but stay under 64
/slot7 = Temporal Overlap – The amount of frames to overlap. The default is 8. You can go up in 4 frames at a time as well
Of course, all of the above is optional for enthusiasts and tinkerers, and totally optional. You can just use it without learning any of this stuff. Enjoy!