FramePack — Long Image-to-Video, Explained & Hosted
FramePack lets a 6GB laptop GPU generate minute-long video from a single image — slowly. Here you get the same job done in about a minute on hosted frontier engines: upload a still, describe the motion, download a clean MP4.
FramePack is an open-source project by Lvmin Zhang (lllyasviel), released under Apache 2.0. SeedDance is an independent platform and is not affiliated with the project — generation on this page runs on comparable hosted engines.
FramePack-style long takes, made here
Dancing, talking, walking — single stills carried into continuous motion. Click any card to view the prompt.
The long-take idea, without the wait
Minutes of render, not hours
A FramePack minute-long clip is an hour-plus on an RTX 4090. Hosted engines here return 5–10s clips in about a minute — iterate instead of queueing.
Similar engines, hosted
Generation runs on Seedance 2.0 and Kling 3.0 — frontier i2v with strong identity consistency, the same quality FramePack's anti-drift design chases.
Image in, motion out
Upload a portrait, sketch or product still and describe the motion — the core FramePack workflow, minus the install.
Animate an image in 3 steps
Upload your still
Portrait, full-body shot, sketch or product photo — clear subjects animate best.
Describe the motion
FramePack-style prompts work here too: 'the girl starts to dance gracefully', 'the man begins speaking with natural gestures'.
Generate & download
About 60–90 seconds for a clip. Clean MP4, no watermark, full rights.
What is FramePack?
FramePack is an open-source video generation method by Lvmin Zhang (lllyasviel — the researcher behind ControlNet). Instead of rendering a whole clip at once, it predicts video frame-by-frame while compressing the growing context to a constant length — so generating minute-long video costs no more memory than generating seconds. Built on a 13B model (HunyuanVideo-based), it made long image-to-video possible on a 6GB laptop GPU when it landed in April 2025, with the F1 and P1 variants following.
The catch is speed, not memory: roughly 1.5–2.5 seconds per frame on an RTX 4090 — so a 60-second, 30fps clip is well over an hour of rendering, and laptop GPUs run 4–8× slower. Anti-drift tricks keep identity stable over long takes, but local FramePack remains a patient hobbyist's tool: install, queue, wait.
That's why most people who just need the result use a hosted service. SeedDance runs comparable frontier engines — Seedance 2.0 and Kling 3.0, marked "similar" in the generator above — turning a still image into motion in about a minute: no GPU, first two videos free, no account.
| Author | Lvmin Zhang (lllyasviel), ControlNet creator |
|---|---|
| Type | Image-to-video via next-frame prediction |
| Min VRAM (local) | 6GB (RTX 30/40/50 series) |
| License | Apache 2.0 |
| Local speed | ~1.5–2.5 s/frame on RTX 4090; laptops 4–8× slower |
FramePack releases
| Version | Released | What changed |
|---|---|---|
| FramePack | Apr 2025 | Initial release: next-frame prediction with constant-length context; long i2v on 6GB VRAM |
| FramePack-F1 | May 2025 | Forward-only prediction variant — larger motion range, less drift accumulation |
| FramePack-P1 | Jun 2025 | Planned-anti-drifting variant announced |
FramePack is a research method with a demo GUI, not a maintained product — expect DIY setup and community forks if you go local.
FramePack vs Wan vs Kling
| FramePack | Wan 2.5 | Kling 3.0 | |
|---|---|---|---|
| Developer | lllyasviel | Alibaba | Kuaishou |
| Strength | Minute-long i2v on 6GB VRAM, Apache 2.0 | Native synchronized audio | Motion control, cinematic camera |
| Open-source | ✅ | ❌ | ❌ |
| Best for | Patient self-hosters with modest GPUs | Sound-on clips, ComfyUI workflows | Choreographed action, hosted speed |
Wan and Kling-class engines run hosted on SeedDance — try the same image and prompt side by side.
FramePack, answered
What is FramePack?
FramePack is an open-source image-to-video method by lllyasviel (ControlNet's author) that predicts video frame-by-frame with a constant-memory context — enabling minute-long clips on GPUs with as little as 6GB VRAM. Full explainer above.
Is FramePack free?
Yes — it's Apache 2.0 open source; you pay only in hardware and time (over an hour per minute of video on an RTX 4090). On SeedDance your first videos on comparable engines are free to try — no card required.
Does this page run FramePack itself?
No — generation here runs on comparable frontier engines (Seedance 2.0, Kling 3.0), marked 'similar' in the model picker. To run FramePack itself, use the lllyasviel/FramePack GitHub repo.
Is this the official FramePack site?
No — FramePack's official home is the lllyasviel/FramePack repository on GitHub. SeedDance is an independent platform.
What hardware does FramePack need locally?
An NVIDIA RTX 30/40/50-series GPU with 6GB+ VRAM, on Windows or Linux. It runs on laptop GPUs, just 4–8× slower than desktops — budget real hours for long clips.
How is FramePack different from normal video models?
Most models render a whole 5–10s clip in one pass, so length is capped by memory. FramePack generates frame-by-frame with constant memory, trading speed for arbitrary length and using anti-drift scheduling to keep identity stable.
Do videos here have a watermark?
No. Free and paid videos both download as clean MP4 files.
Can I use the videos commercially?
Yes — videos generated on SeedDance come with full rights. Self-hosted FramePack output is Apache 2.0-friendly too; mind the base model's own license terms.
Can hosted engines do minute-long videos like FramePack?
Hosted clips here run 5–10 seconds per generation — the trade for frontier quality in a minute of waiting. For longer pieces, generate scenes and cut them together; extend features vary by engine.
FramePack vs Wan — which should I self-host?
FramePack if VRAM is your constraint and long takes are the goal; Wan if you want the bigger ecosystem, audio (2.5+) and faster whole-clip rendering on a beefier card. Both reward tinkerers more than deadline-holders.