Seam is an open-source take on the "transitions" apps (Higgsfield, Kling, Pika), built for Comfy's Open Source Higgsfield challenge. No accounts, no credits: the model runs on your machine and every take can be reopened, inspected and re-run in ComfyUI.
- Web studio: shots on the left, the take on the right, your gallery below. Images, 1–6 s clips or 3D models as input, ten qualified effects (5 s takes, 10 s for Smoke, Water bend and Whip pan) or Surprise me, and three examples so you can try it in one click.
- Sequences: more than two shots? Put them in order, pick an effect per hand-off, and Seam renders each transition and joins them into one video (clips play in full at their place). A bad hand-off gets a new take on its own; the others are reused. Or grow a story shot by shot: Continue a finished two-shot take with a new shot, then keep adding; only the new transition renders, and every transition starts from your own shot, so the story stays as consistent as the shots you bring.
- CLI:
seam make first.png last.png --preset smoke,seam sequence make 1.png 2.png 3.png --preset smoke,whip_pan, thenseam sequence extend <id> 4.png --preset ink, andseam sequence edit <id> --order 1,3,2,4 --transition 1=ink --transition 2=melt --transition 3=smoketo reorder (only the new hand-offs render). - Agent plugin: a skill + MCP server so Claude Code (or any MCP agent) can plan, render, review and refine transitions for you. Packaged as a Claude Code plugin and as an Agent Plugins 1.0 plugin.
Why it's different
| Generative, not a 2D blend | Code tools (ffmpeg xfade, Remotion) cross-fade two flat frames. Seam asks a video model for what is between the shots: smoke that really hides the subject, water that reshapes into the next scene, a pan with real parallax and motion blur. |
| Any shot in, even 3D | Images, clips, or a .glb you frame in the browser and record as a camera orbit. Everything becomes the frame the transition leaves from or lands on, so clips and orbits stay continuous. |
| Open by default | Open weights (LTX-2.5, Gemma-4 E2B), core ComfyUI nodes, the official Comfy SDK. Each take archives the exact graph it submitted; Open in ComfyUI downloads it with both shots, and a receipt says how it was made. |
| Only qualified effects ship | An effect is exposed only after it plays well on two image pairs × two seeds, then on unseen pairs. Ten passed; more ship only if they pass too. Longer takes are qualified the same way, per effect. |
| Seam reads your shots | A small local model describes each shot ("a cracked eggshell with jagged openings…") and that line goes into the prompt. You can correct it before you render. |
| Agents included | The same server backs the web app, the CLI and an MCP server, so an agent can run the whole loop: plan, render, look at the take, refine. |
Tour
The studio is styled after comfy.org: Comfy ink and yellow, condensed type, and ComfyUI's own node-and-wire language (socket colours by type). Design notes: DESIGN.md.
Takes
All real renders, seed 1, no retouching. The input stills are fictional and were generated for testing.
A person (first shot → Seam → last shot)
| First shot | Smoke | Water bend | Whip pan | Last shot |
|---|---|---|---|---|
A product (sneaker on wet concrete → cup on a mossy rock)
| Smoke | Water bend | Whip pan |
|---|---|---|
3D models — the egg and the brachiosaurus at the top of this page: two .glb files, two recorded orbits, one take.
Measured (RTX 5090 32 GB, Windows 11, ComfyUI Desktop 0.41.3; 5 s clips at 1280×704 or 704×1280):
| Effect | Takes | Render time, median (min–max) |
|---|---|---|
| Smoke | 10 | 41 s (30–46 s) |
| Water bend | 9 | 40 s (30–41 s) |
| Whip pan | 9 | 41 s (30–56 s) |
| Lava, Beast, Ink, Light burst, Zoom through, Melt, Wind vortex | 4 each | 38–43 s (35–47 s) |
| Smoke, Water bend, Whip pan at 10 s | 18 | 76 s (71–86 s) |
Render time runs from submit to finished video, including LTX-2.5's prompt enhancer and any model loading (the slowest takes were the first after a load). Reading the two shots (~1 s each, cached) comes on top. Peak VRAM was ~31 GB on a 32 GB card; 24 GB cards should work through ComfyUI's offloading (64 GB RAM) but are untested. The qualification record is dev/PRESET_REVIEW.md.
3D models in
Drop a .glb on either shot. Seam opens it in a small 3D studio in the browser (three.js): neutral set, studio
light, the model framed for you; drag to change the angle.
- Record orbit renders a 3 s, 120° camera orbit and uploads it as a clip. The transition leaves from where the first orbit ends and lands where the second orbit begins, and Seam joins orbit + transition + orbit into one video. The orbit is paced by the clock, so a slow GPU drops frames but the clip still lasts 3 s.
- Use this angle keeps one view as a still.
The model never leaves your machine; only the recorded orbit or still goes to the Seam server, like any upload.
Hackathon submission
| Category fit | Open Weights Only: every render is LTX-2.5 (+ Gemma-4 E2B to read the shots) on your own GPU through core ComfyUI nodes; no partner nodes, no closed APIs. VFX Workflow built with LTX: LTX-2.5 first/last-frame is the core model of every effect. |
| Comfy Developer Platform | Every render goes through the official Comfy SDK to the Comfy API v2 (comfy-api-proxy locally). |
| Workflow files | src/seam/workflows/ (API format, slot-mapped), and any take exports its exact graph via Open in ComfyUI. |
| Run it locally | install.md, checked end to end by uv run seam doctor. |
| Try it | The app ships three example pairs: open the studio and pick one. There is no hosted endpoint: Seam runs on your GPU. |
| Demo | docs/media/seam-demo.mp4 (1:56, real renders, waits sped up and labelled). |
Quick start
You need an NVIDIA GPU with 32 GB (24 GB untested), 64 GB RAM, ComfyUI ≥ 0.41.3, Python 3.12 with uv, Node 20+
and ffmpeg. The full guide, written so an agent can follow it too, is install.md.
git clone https://github.com/jdnichollsc/seam && cd seam
uv run python scripts/download_models.py --models-dir "<ComfyUI>/models" --only l25 # ~44 GB, open weights
uv tool install comfy-api-proxy==0.1.6
comfy-api-proxy start --comfyui http://127.0.0.1:8188 # Comfy API v2 on :8189
uv run seam doctor # checks ffmpeg, the Comfy API, nodes, models and the web build
uv run seam serve # http://127.0.0.1:7860
Agents. In Claude Code:
claude plugin marketplace add jdnichollsc/seam
claude plugin install seam
Then ask for a transition ("make a smoke transition from a.png to b.png and show me a preview"). The MCP server
exposes seam_status, list_presets, create_transition, wait_job, preview_job, refine_job, get_job,
list_jobs and cancel_job, plus create_sequence, wait_sequence, get_sequence, preview_sequence,
refine_segment, cancel_sequence, continue_take, extend_sequence and edit_sequence (remove / reorder
shots, a story line for the whole film) for multi-shot videos; the skill tells the agent how to plan shots, judge a take and refine it.
How it works
browser / CLI / agent (MCP)
│ REST + live events
Seam server (FastAPI) ── presets → model-specific prompt + slot-mapped workflow
│ Comfy SDK (comfy-sdk, Comfy API v2)
comfy-api-proxy → ComfyUI (core nodes only) → LTX-2.5 on your GPU
- Every generation goes through the official Comfy SDK against the Comfy API v2, served locally by
comfy-api-proxy. The same client can point at Comfy Cloud or a deployed endpoint throughCOMFY_BASE_URL(not tested yet). - Each effect is a preset (our own prompt text) routed to the model that renders it best: LTX-2.5 first/last-frame with its in-graph prompt enhancer (every launch effect), or LTX-2.3 with an open transition LoRA.
- Seam reads both shots first with the same small local model as the enhancer (Gemma-4 E2B), so the prompt says what is in them. You can correct that line before rendering.
- Workflows are the official ComfyUI templates exported to API format; Seam only touches them through a named slot map, with a test that fails if the template drifts.
- Jobs persist in SQLite and survive a server restart (they re-attach to the running ComfyUI job, and a job that was mid-submit is reported, never silently resubmitted).
- Clips in, clips out: either side can be a 1–6 s clip (normalized to one clock on upload). The transition starts from clip A's last frame and lands on clip B's first, and Seam joins clip A + transition + clip B with ffmpeg.
- Every take is reproducible: a receipt (input hashes, effect, model, seed, final prompt, graph hash) and a one-click Open in ComfyUI bundle: the exact API workflow plus both stills, runnable in plain ComfyUI.
- The server only answers on loopback and refuses cross-site writes.
Built with
| Tool | What it does in Seam |
|---|---|
| ComfyUI | runs the graphs (core nodes only, the official LTX-2.5 first/last-frame template) |
| Comfy SDK + comfy-api-proxy | every submit, upload and download, over the Comfy API v2 |
| LTX-2.5 | the video model, with its in-graph prompt enhancer |
| Gemma-4 E2B | reads each shot so the prompt describes it |
| FastAPI · SQLite · ffmpeg | the Seam server, job store, clip normalization and joins |
| React · GSAP · three.js | the web studio, its motion, and the 3D-model studio |
| MCP · Agent skills | the agent plugin (Claude Code, Agent Plugins) |
Models and licenses
| Model | Role | License |
|---|---|---|
| LTX-2.5 22B distilled (int8) + VAEs | default video model | LTX-2.x Community License (gated) |
| Gemma-4 12B (LTX-2.5 text encoder) · Gemma-4 E2B (prompt enhancer) | text encoding / prompt enhancement | Gemma terms |
| LTX-2.3 22B distilled fp8 + Gemma-3 12B | optional video model | LTX-2 Community License |
| LTX-2.3 Transition LoRA | optional transition style | see model card |
| Z-Image-Turbo | only used to generate the fictional test images in dev/ | see model card |
Seam's code is MIT (LICENSE). The models keep their own licenses; read them before commercial use.
The effect presets are our own text, inspired by common transition concepts — no third-party prompts or media.
The 3D models (.glb) in the demo are the author's own assets, released under MIT.
Status: tested on Windows 11 with one RTX 5090. Linux, macOS and smaller GPUs are untested.
Responsible use
Use images you own or made. Don't put a real, identifiable person into a transition without their permission. Keep it SFW.
Credits
Built on ComfyUI, the Comfy SDK and comfy-api-proxy, LTX by Lightricks, and the open transition LoRA by valiantcat. Made for #OpenSourceHiggsfield by @jdnichollsc.