Most free AI video tools give you five seconds and a watermark. SkyReels-V2 is different: it is an open-source video model from Skywork AI that can generate film-like clips of 10, 30 or even 60 seconds — and its small 1.3B version runs on a free Google Colab GPU. Here is what it does, and exactly how to run it for free.
What is SkyReels-V2?
SkyReels-V2 is an open-source text-to-video and image-to-video model, released by Skywork AI (the research arm of Kunlun Wanwei) in April 2025 and described in the paper arXiv:2504.13074. Its headline trick is Diffusion Forcing: instead of generating a whole short clip at once, it builds the video in blocks, carrying what came before into each new block. That is how it keeps characters and scenes consistent across long takes while other free models fall apart after a few seconds.
- Text-to-video: describe a scene and get a moving clip.
- Image-to-video: upload a still image and animate it, with optional first-and-last-frame control.
- Long video: the Diffusion Forcing models support 10, 15, 30 and 60-second generations in one run.
- Camera language: it understands cinematic prompts — shot composition, camera movement, actor expressions.
- Open weights and code: the repository has about 7,800 stars on GitHub, and the models are free to download from Hugging Face.
Which version to use
Two sizes matter for free use. The 1.3B 540P model is the free-tier hero: with CPU offloading and TeaCache it runs in about 14.7 GB of VRAM, which fits a free Colab T4 GPU (16 GB). The 14B 720P model is sharper but needs 24 GB or more — a paid Colab GPU, a rented A100, or a high-end card of your own. Start with the 1.3B. It also ships in a Diffusers-ready form, so you can run it with a few lines of Python instead of the full repository.
Method 1: Free on Google Colab (no PC needed)
This is the easiest free route. You need only a Google account.
- Go to Google Colab and start a new notebook. In Runtime → Change runtime type, choose T4 GPU (free tier).
- Install the libraries in one cell:
!pip install -q diffusers transformers accelerate torch
- Paste this code and run it. It loads the 1.3B Diffusion Forcing model and makes a 4-second clip (97 frames at 24 fps):
import torch from diffusers import AutoModel, SkyReelsV2DiffusionForcingPipeline, UniPCMultistepScheduler from diffusers.utils import export_to_video model_id = "Skywork/SkyReels-V2-DF-1.3B-540P-Diffusers" vae = AutoModel.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32) pipe = SkyReelsV2DiffusionForcingPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16) pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=8.0) pipe = pipe.to("cuda") frames = pipe( prompt="A small blue bird takes off from the ground, flapping its wings, blue sky with white clouds, camera follows the bird upward", num_inference_steps=30, height=544, width=960, num_frames=97, base_num_frames=97, ar_step=5, causal_block_size=5, ).frames[0] export_to_video(frames, "my_video.mp4", fps=24, quality=8) - Download my_video.mp4 from the file browser on the left. That is your first clip.
For a longer video, change num_frames to 257 for about 10 seconds or 737 for about 30 seconds, and add overlap_history=17 and addnoise_condition=20 — both keep long clips consistent. Expect several minutes of generation per clip on a free T4, and remember Colab's free tier has daily usage limits and can disconnect idle sessions.
Method 2: Run it on your own PC
If you have an Nvidia GPU with at least 16 GB of VRAM, run the official repository:
- Clone it and install the requirements:
git clone https://github.com/SkyworkAI/SkyReels-V2 cd SkyReels-V2 pip install -r requirements.txt
- Generate a 10-second clip (the model downloads from Hugging Face automatically):
python3 generate_video_df.py \ --model_id Skywork/SkyReels-V2-DF-1.3B-540P \ --resolution 540P \ --ar_step 0 \ --base_num_frames 97 \ --num_frames 257 \ --prompt "A graceful white swan swimming in a serene lake at dawn, mist rising from the water" \ --offload --teacache
- To animate a photo instead, use the image-to-video model (
Skywork/SkyReels-V2-I2V-1.3B-540P) and add--image your_photo.jpgto the command.
Prompt tips that actually help
- Write like a director: subject, action, setting, light, camera move — "a street food vendor flips noodles at night, steam rising, warm neon light, slow dolly-in".
- One scene per clip: long prompts with scene changes confuse the model; chain short clips instead.
- Use the captioner: the same team released SkyCaptioner-V1, a model that writes detailed video captions — good prompts in, good videos out.
Honest limits
- The free 1.3B route is 540p — fine for social clips, not cinema.
- Generation is slow on a free T4: budget several minutes per clip and keep sessions short.
- Long clips can show small jumps where blocks join; the project's own notes and community reports flag this in asynchronous mode.
- Check the Skywork licence in the repository before using clips commercially — the weights are free to download, but the licence is the project's own, not a standard permissive one.
- A newer SkyReels-V3 exists (released January 2026), but it is heavier — the V2 1.3B is still the one that runs free.
Bottom line
If you want long AI video clips without paying for Sora, Runway or Kling credits, SkyReels-V2's 1.3B model is the best free option right now: real open weights, a genuine long-video trick in Diffusion Forcing, and a route that runs on a free Colab GPU. Open the notebook, paste the code, and your first clip is minutes away.
Links: SkyReels-V2 on GitHub · 1.3B model on Hugging Face · Technical report (arXiv:2504.13074) · Google Colab. Hero photo: a RED Digital Cinema camera, by Tvcameraop, CC BY-SA 4.0, via Wikimedia Commons.
