Video generation
Video — Mochi, CogVideo, WAN, upscale pipelines
Text-to-video and video upscaling without the dependency tar pit.
Launch standard video
Pay by the hour · no contracts · terminate anytime from the dashboard
What this is
Open-weight video models move fast — Mochi-1, CogVideoX-5B, WAN 2.1, HunyuanVideo — and each one has its own CUDA / diffusers / xformers version matrix. The instance boots a clean CUDA 12.1 / Ubuntu base — nothing extra installed yet — but we publish the combination that actually works so you spend your first minutes on `pip install`, not build errors.
Our value-add
We track the working CUDA / xformers / diffusers combination for current-gen video models and update the recommended stack when it breaks.
Environment & recommended stack
- Base image: nvidia/cuda 12.1 devel (Ubuntu 22.04)clean boot — no ML framework preinstalled
- PyTorch + Flash Attention 2recommended — the base video stack, one pip install
- Diffusers (latest)recommended for Mochi-1, CogVideoX, AnimateDiff
- FFmpeg with NVENC/NVDECapt install — hardware-accelerated encode/decode
- RIFE + Real-ESRGANrecommended for interpolation and upscale
You stay in control
- Full root SSH — pull any video model weights or custom ComfyUI workflows
- Snapshot the instance before you terminate it — relaunch from that snapshot later; local disk doesn't survive termination otherwise
- Auto-destroy deadline so an overnight render can't run to $500
- tmux / nohup friendly — kick off a long render and come back
Example workflow
Render a 5-second CogVideoX clip
# pip install torch diffusers transformers accelerate — see the recommended stack above
from diffusers import CogVideoXPipeline
pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype="auto").to("cuda")
video = pipe("A slow push-in on a robot guitar, shallow depth of field", num_frames=49).frames[0]
pipe.export_to_video(video, "clip.mp4", fps=8)Run inside the instance after `ssh root@<instance-ip>` or via the web terminal.
Launch options
Pick a GPU for video generation
Recommended start
Standard video — Mochi-1 / CogVideoX-5B
80 GB HBM3 handles Mochi-1 + long-form CogVideo at native resolution
$4.41/ hour
Launch with this configScale option
Heavy multi-GPU — HunyuanVideo / long clips
Tensor parallelism for ~10s+ clips without OOM
$31.10/ hour
Launch with this configHourly rates shown are current target rates — the authoritative price is surfaced on the /launch page at the moment you provision. Billing is per 5 minutes of runtime.