Add support to image-to-video task for inference

[apolinario]

Internal discussion

The goal is:

• Keep inputs similar to image-to-image
• Keep outputs similar to text-to-video
• For now super generic about.md
• data.ts for the /tasks on huggingface.co is pending merging this PR

[Wauplin]

Thanks @apolinario ! Pretty solid PR already :) I've made some high-level comments while you are still working on it. I do think it's best to split the PR into 2 (one for spec definition, one for proper implementation for each provider).

[Wauplin]

Could you add a simple example in https://github.com/huggingface/huggingface.js/blob/main/packages/tasks/src/snippets/inputs.ts ? Something like this:

const inputsImageToVideo = () => `{
    "image": "cat.png",
    "prompt": "Make the cat dance."
}`;

and then

    "image-to-video": inputsImageToVideo,

in the modelInputSnippets mapping.

This way, const data = JSON.parse(getModelInputSnippet(model) as string); will return a valid payload to generate the inference snippets.

[Wauplin]

Let's not add this file yet and wait for a follow-up PR. The current implementation assumes it works with FalAI and Replicate which is not the case yet -and not tested-. Better to separate them and keep this PR for specs definition only.

[apolinario]

Thanks @Wauplin ! I've kept it strictly spec-based for now and applied your suggestion for snippets/inputs.ts

[Wauplin]

Looks good to me, thanks!

Let's wait for another review, especially for the about.md/data.ts part which I know less about

[merveenoyan]

thanks a lot 💗

[apolinario]

Thanks merve, applied your suggestions!

[SBrandeis]

Approving in principle - deferring to @merveenoyan re: data.ts

I will review the inference client implementation more thoroughly :)

[merveenoyan]

LGTM :)