Seam GitHub

Two shots in. One transition out.

Seam makes a short video that travels from a first shot to a last one through an effect. Images, clips or 3D models in; open-weights LTX-2.5 on your own GPU; every take is a ComfyUI workflow you can open.

Run it locally Watch the demo Read the code
First shotimage
First shot: a white sneaker on wet concrete
Last shotimage
Last shot: a ceramic cup on a mossy rock
EffectLTX-2.5
  1. Whip Pan
  2. Smoke
  3. Water Bend
Output5 s · 1280×720 · rendered on an RTX 5090

Real takes, seed 1, no retouching. The input stills are fictional, made for testing.

Three ways in

Each shot takes whatever you have. Everything becomes the frame the transition starts from or lands on.

  1. An image

    PNG, JPEG or WebP. Seam reads what is in it with a small local model and writes that into the prompt; you can correct the line.

  2. A 1–6 s clip

    The transition leaves from the first clip's last frame and lands on the second clip's first frame, then Seam joins clip, transition and clip into one video.

  3. A 3D model (.glb)

    Frame it in the browser and record a 3 s camera orbit: the orbit becomes the clip. Or keep one angle as a still. The model never leaves your machine.

3D → Smoke → 3DAn egg orbit and a brachiosaurus orbit, both from .glb models, joined by one real take.

Only effects that pass ship

Each effect is our own prompt, routed to the model that renders it best and played on two image pairs, two seeds, then on unseen pairs. Pick one, or let Seam surprise you.

Smoke41 s median render
Water Bend40 s median render
Whip Pan41 s median render

Render time from submit to finished 5 s video on one RTX 5090, including the prompt enhancer and model loading. Seven more effects have since passed (Lava, Beast, Ink, Light Burst, Zoom Through, Melt, Wind Vortex), and these three also render 10 s takes (~76 s).

Open by default

Open weights
LTX-2.5 first/last-frame video with its in-graph prompt enhancer, and Gemma-4 E2B to read the shots. No partner nodes, no closed APIs, no credits.
Every take is a workflow
Seam archives the exact graph it submitted. "Open in ComfyUI" downloads it with both shots; the receipt says how the take was made.
Comfy Developer Platform
Every render goes through the official Comfy SDK to the Comfy API v2: served locally by comfy-api-proxy in front of your ComfyUI. The same client can point at Comfy Cloud, which we have not tested yet.
Agents included
A skill and an MCP server let Claude Code or any MCP agent plan, render, review and refine transitions, packaged as a Claude Code plugin and an Agent Plugins plugin.

Run it

A 32 GB NVIDIA GPU, 64 GB RAM and ComfyUI. The full guide, written so an agent can follow it too, is install.md.

  1. Get the models (~44 GB, open weights)
    git clone https://github.com/jdnichollsc/seam && cd seam
uv run python scripts/download_models.py --models-dir "<ComfyUI>/models" --only l25
  2. Start the Comfy API proxy in front of ComfyUI
    uv tool install comfy-api-proxy==0.1.6
comfy-api-proxy start --comfyui http://127.0.0.1:8188
  3. Check, then open the studio
    uv run seam doctor
uv run seam serve        # http://127.0.0.1:7860

Prefer the terminal? seam make first.png last.png --preset smoke. Prefer an agent? Install the plugin and ask for a transition.