GPT-Image-2-Seedance-2.5-Workflow: a guide to image-to-video prompting

A curated guide, not a code library

This repository is a workflow guide for combining GPT Image 2 image generation with Seedance 2.5 video generation, and the useful thing to understand about it is what it is not. There is no pipeline to install and no library to import. The stated purpose is to collect proven workflows, prompt templates and real creator examples, and the README says most cases are curated from X and Twitter creators, community experiments and real production workflows. The division of labour between the two models is described plainly: GPT Image 2 handles what to draw and visual consistency, while Seedance 2.0 handles how to move, animating those images into video. The repository layout matches that framing, with the README in a dozen languages, an images folder for the figures, and a scripts directory. The primary language is recorded as Python, which reflects the small amount of scripting that ships alongside the guide rather than a Python package you would depend on.

One API call turns a storyboard into a clip

The concrete part of the guide is a request to the EvoLink API. The key is exported into the environment and then a POST goes to the video generations endpoint:

export EVOLINK_API_KEY="your_key_here"

curl --request POST \
  --url https://api.evolink.ai/v1/videos/generations \
  --header "Authorization: Bearer ${EVOLINK_API_KEY}" \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "seedance-2.0-image-to-video",
    "prompt": "Animate this GPT Image 2 product storyboard with a slow dolly-in, realistic reflections, and smooth commercial camera motion",
    "image": "https://example.com/gpt-image-2-storyboard.png",
    "duration": 5,
    "quality": "720p",
    "aspect_ratio": "16:9"
  }'

Four parameters in that payload are worth naming because they are the levers a caller actually has. The model is the image-to-video variant, the image field points at a URL rather than carrying an upload, and the prompt describes the camera move rather than the subject, which matches the split of responsibilities between the two models. Duration, quality and aspect ratio are set as plain numbers and strings. The image reference being a URL is the practical constraint of the workflow: the storyboard produced by the image model has to live somewhere the API can fetch it.

Two npx installers that place skills in your agent

The guide also points at two companion skills. The GPT Image 2 Gen Skill adds GPT Image 2 image generation and editing to your agent, and the Seedance 2 Video Gen Skill adds Seedance 2 video generation. Both are installed with npx, and the terminal versions are:

npx evolink-gpt-image@latest
npx evolink-seedance@latest

The installers auto-detect a known skills directory when possible and otherwise prompt you where to install. That behaviour is the reason there is a second family of commands. Each agent host gets its own path so an agent can install without asking follow-up questions, with the path flag pointing at the skills directory of the host:

npx evolink-gpt-image@latest -y --path ~/.opencode/skills
npx evolink-seedance@latest -y --path ~/.opencode/skills

The host list covers Claude Code with a .claude/skills path, OpenCode with .opencode/skills, Cursor with .cursor/skills, and OpenClaw, where the commands carry the -y flag without an explicit path. The guidance is to use the plain terminal form when you are installing manually, and the host specific form when you want an agent to do it unattended. The package names are published on npm under evolink-gpt-image and evolink-seedance, and both are pinned to the latest tag, so a re-run picks up a new version without any version management on your side.

The case library is the real content

The table of contents is organised as storyboard techniques with numbered cases, each credited to the creator it came from. The first entries are a standard storyboard to video case by @kiyoshi_shin, a three by three grid storyboard method by @servasyy_ai, and a multi-frame reference to fast-cut video case by @heygentlewhale. Reading those attributions tells you something about the provenance of the advice, which matters when you are copying prompts. The methods are structural rather than stylistic, a grid storyboard is a layout decision that affects how the model reads motion, and a multi-frame reference is a different input shape from a single still. That is the value of collecting cases from working creators, since the prompt text alone does not carry the layout. The guide is also mirrored across a set of sibling repositories under the same organisation, an awesome list of Seedance 2.5 prompts, a Seedance 2.5 Gateway Service with API examples, a Seedance 2.5 guide collection, and an awesome list for the GPT Image 2 API and prompts.

The licences and the release history need reading carefully

The metadata and the badge disagree here. The repository is listed as having a custom licence that GitHub cannot classify, with a LICENSE file at the root, while the README carries a Creative Commons Attribution 4.0 badge. For a guide made of curated prompts and screenshots, the difference between an open content licence and an unclassified custom licence is not cosmetic, and the LICENSE file is the only authority. That is worth reading before you republish any of the cases. The release history is thin: a single tag, v1.0.0 with the note Initial Release, dated 2026-04-23, and there is no tag after it. The last push was on 2026-06-24 and the repository is not archived, so the content is still being edited, but there is no versioning scheme behind it. In practice that means you track the master branch, not a release, and the guide is documentation you read rather than a dependency you pin.

Early access gates the video side of the workflow

The Seedance half of the pipeline is not simply open. The README carries a note that instructs the reader to build images with GPT Image 2 and then bring them into the Seedance 2.5 early access workflow for cinematic video generation, with a link to join the rollout. So the workflow has three dependencies rather than two: an image model, a Seedance 2.5 early access invitation through the launch page, and an EvoLink API key obtained through signup. The documentation site at docs.evolink.ai is linked for API details, and the launch page for Seedance 2.5 Early Access is promoted from the top of the README. That structure explains the repository's shape. The curl example uses the seedance-2.0 image-to-video model, which is the version the public endpoint documents, while the 2.5 early access path is where the newer capabilities are being offered, so a reader following the guide today is on the 2.0 API with 2.5 as the next step.

Editorial conclusion

Adopt this repository if you are producing image-to-video clips and want proven prompt patterns rather than a starting guess, and read the creator cases before writing your own. Do not treat it as a maintained library, since the only release is v1.0.0 from 2026-04-23, and note that the images have to be reachable at a public URL for the API call to work.

Frequently asked questions

Official sources

1. EvoLinkAI/GPT-Image-2-Seedance-2.5-Workflow on GitHub
2. Issues
3. Project website
4. README
5. Releases