Model reference · open weights
SkyReels-R2V is an open-weight video model from Skywork. SkyReels-V3-R2V-14B (BF16) weighs 51.8 GB; the smallest configuration that runs it is RTX 5090 32 GB.
Summary of the Skywork/SkyReels-V3-R2V-14B model card, 2026-10-01
What it is
| Released by | Skywork |
|---|---|
| Released | 2026-01-19 |
| Parameters | 14.3B |
| VRAM | 51.8 GB for the weights |
What it runs on
| Card | Weights | Memory |
|---|---|---|
| RTX 3060 12 GB … RTX 4090 24 GB | does not fit | |
| RTX 5090 32 GB | tight (encoders offloaded) | 31.0 GB |
| L40S 48 GB | fits (encoders offloaded) | 44.0 GB |
| A100 80 GB | fits | 78.2 GB |
| H100 80 GB | fits | 78.1 GB |
| RTX PRO 6000 Blackwell 96 GB | fits | 93.8 GB |
| DGX Spark (GB10) 128 GB unified | fits | 107 GB |
| H200 141 GB | fits | 138 GB |
| B200 180 GB | fits | 176 GB |
From the model card
👋 Playground . 🔧 API Platform · 🤗 Hugging Face · 🤖 ModelScope · 📑 Technical Report
Welcome to the SkyReels V3 repository! This is the official release of our flagship video generation model, built upon a unified multimodal in-context learning framework. SkyReels V3 natively supports three core generative capabilities: 1) multi-subject video generation from reference images, 2) video generation guided by audio, and 3) video-to-video generation.
The demos above showcase videos generated using our SkyReels-V3 unified multimodal in-context learning framework.
# Clone the repository
git clone https://github.com/SkyworkAI/SkyReels-V3
cd SkyReels-V3
# Install dependencies (Recommended: Python 3.12+, CUDA 12.8+)
pip install -r requirements.txt
Models are available on Hugging Face and ModelScope:
| Model Type | Variant | Links |
|---|---|---|
| Reference to Video | 14B-720P | 🤗 Hugging Face / 🤖 ModelScope |
| Video Extension | 14B-720P | 🤗 Hugging Face / 🤖 ModelScope |
| Talking Avatar | 19B-720P | 🤗 Hugging Face / 🤖 ModelScope |
Note: By default, the script automatically downloads models from Hugging Face. To use a local path, specify it via the
--model_idflag.
Reference-to-Video is a model that synthesizes coherent video sequences from 1 to 4 reference images and a text prompt. It excels at maintaining strong identity fidelity and narrative consistency for characters, objects, and backgrounds.
python3 generate_video.py --task_type reference_to_video --ref_imgs "https://skyreels-api.oss-accelerate.aliyuncs.com/examples/subject_reference/0_1.png,https://skyreels-api.oss-accelerate.aliyuncs.com/examples/subject_reference/0_2.png" --prompt "In a dimly lit, cluttered occult club room adorned with shelves full of books, skulls, and mysterious dolls, two young Asian girls are talking. One girl has vibrant teal pigtails with bangs, wearing a white collared polo shirt, while the other has a sleek black bob with bangs, also in a white polo shirt, conversing under the hum of fluorescent lights, a high-quality and detailed cinematic shot." --duration 5 --offload
torchrun --nproc_per_node=4 generate_video.py --task_type reference_to_video --ref_imgs "https://skyreels-api.oss-accelerate.aliyuncs.com/examples/subject_reference/0_1.png,https://skyreels-api.oss-accelerate.aliyuncs.com/examples/subject_reference/0_2.png" --prompt "In a dimly lit, cluttered occult club room adorned with shelves full of books, skulls, and mysterious dolls, two young Asian girls are talking. One girl has vibrant teal pigtails with bangs, wearing a white collared polo shirt, while the other has a sleek black bob with bangs, also in a white polo shirt, conversing under the hum of fluorescent lights, a high-quality and detailed cinematic shot." --duration 5 --offload --use_usp
💡 Notes:
- The
--task_typeparameter must be set toreference_to_video.- The
--ref_imgsparameter accepts 1 to 4 reference images. When providing multiple images, please separate their paths or URLs with commas.- The recommended output specification for this model is a 5-second video at 720p and 24 fps.
Extends existing videos while preserving motion continuity, scene coherence, and subject identity.
python3 generate_video.py --task_type single_shot_extension --input_video https://skyreels-api.oss-accelerate.aliyuncs.com/examples/video_extension/test.mp4 --prompt "A man is making his way forward slowly, leaning on a white cane to prop himself up." --duration 5 --offload
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.
Running it yourself
Rent a machine by the hour. Runs as it is with diffusers, on the machine, in Python.
# on your rented machine: pip install diffusers transformers accelerate ftfy
import torch
from diffusers import DiffusionPipeline
from diffusers.utils import export_to_video
pipe = DiffusionPipeline.from_pretrained("Skywork/SkyReels-V3-R2V-14B", torch_dtype=torch.bfloat16).to("cuda")
frames = pipe(prompt="a drone shot over a forest at sunrise").frames[0]
export_to_video(frames, "/workspace/out.mp4", fps=16)