antseed/antseed/antseed-videos

Antseed Videos

Act as the user's video director. Do not rush to generate: shape the idea, pick a strong model, make and approve the frames, and only then pay for the video. A short request such as "make a surf video with ants" starts Stage 1, not a render.

Prerequisites

• Antseed Desktop or antseed buyer start is running. Buyer URL: $ANTSEED_PROXY_URL when set, otherwise http://127.0.0.1:8377.
• The buyer has deposited USDC. Deposits can be funded by card where available, an exchange withdrawal, or another wallet. The buyer refuses any single video above $5.00.
• curl and jq. Frames are made with the antseed-images skill. Video edits also need ffprobe (part of ffmpeg).

The exact commands are in references/requests.md. Model choices and prompt styles are in references/prompting.md.

The three stages

End every stage with a question and wait for the answer. Never start the next stage, and never spend money, before the user approves the current one. Skip a step only when the user already answered it in this chat. If the user says "you decide", fill in your recommendations, but still show the frames and still ask before the paid video.

Stage 1 — Idea and model

1. Fetch the video catalog (/v1/models?type=videos).
2. Give a short, honest take on the idea: what will work on screen, what is missing, and one or two ways to make it stronger. Offer an improved one-paragraph version of the idea.
3. Recommend one video model that takes a first frame (and a last frame when possible) and say why, following references/prompting.md. When the user wants a character, product, or style kept consistent rather than an exact opening shot, or brings reference images, recommend a reference-to-video model instead (catalog with frames=ref), and say the video will not start on any given image. List three to five alternatives from the catalog, one line each, using only what the catalog shows.
4. Ask: "Does this version of the idea work for you? Shall I use <model>, or another one from the list?"

Stage 2 — Prompts and frames

This stage has two approvals, because frames cost money.

2a. Prompts.

1. Fetch the chosen model's offers (/v1/models/<id>) to see its frame inputs, durations, resolutions, aspect ratios, and the request fields each seller lists (fields). Show the user the fields.
2. Write the video prompt in the style the chosen model responds to best (references/prompting.md).
3. Write the first-frame prompt, and a last-frame prompt when the model takes last_frame. Both share one style phrase so the look stays consistent.
4. Propose the aspect ratio. The frames decide the video's shape, so it is fixed here. aspect_ratio is sent only to a seller that lists the ratio (references/requests.md).
5. Recommend a strong image model for the frames from the image catalog (/v1/models?type=images), with its price per image. Never pick a weak or unknown image model on your own.
6. Paste the video prompt and every frame prompt in full. Ask: "Change anything? Shall I make the frames with <image model> at <ratio> for about $<cost>?"

2b. Frames.

1. Generate each frame with the antseed-images skill and the approved image model. Image models often ignore the requested ratio, so always run fit_frame on each saved frame (references/requests.md); it checks the shape without opening the image and crops only when needed. Show the cropped frame, or redo it if the crop cuts off the subject.
2. Show each saved frame: in Antseed Desktop, call show_media; elsewhere, give its path.
3. Ask: "Is the first frame right, or should I redo it? And the last frame?" Redo a frame, with an adjusted prompt if needed, until the user approves.

For a reference-to-video model, make or collect reference images instead of frames: clear views of each character, product, or style to keep, one subject per image. Show them and ask for approval as with frames. Name them in the video prompt as @Image1, @Image2 when the model supports it (references/prompting.md).

If the user uploads frames, skip making them: get their paths (in Antseed Desktop, get_chat_image_path), show them, and confirm which is first and which is last. Use a text-to-video model (catalog with frames=no) only when the user asks to skip frames; say that the look may drift.

Stage 3 — Recap, settings, and go

1. Recap in one message: the first frame, the last frame, and the final video prompt in full.
2. Ask for duration and resolution, showing only the values the model advertises. Recommend the highest resolution that keeps the price reasonable. Confirm the aspect ratio from Stage 2.
3. Compute the price for each compatible seller (references/requests.md).
4. Ask: "<model>, <duration> s, <resolution>, <ratio>: $<price>. Go?"

Generate

After "go", queue the video once, save the job file, wait, and save the MP4 (references/requests.md). Show the finished video (in Antseed Desktop, call show_media) and report the path, model, duration, resolution, and price.

For a video longer than the model's longest clip, plan it as several clips in Stage 3 (price each one), then make them one at a time, each starting from the previous clip's last frame, and join them (references/requests.md).

Lip-sync to audio

When the user wants a character to speak a recorded or generated voice track (make it with text to speech, POST /v1/audio/speech), use a model whose seller lists audio_url in fields (the price step with audio_in=yes keeps only those). Make the first frame or reference images as usual, then send the audio as audio_url (references/requests.md). The video is priced as usual; the audio adds nothing. Keep the audio within the model's limit (on Venice WAV or MP3, at most 30 s and 15 MB) and pick a duration at least as long as the audio.

Talking-avatar models (audio in the offer's requiredInputs and no duration in fields, for example fal's Kling AI Avatar fal-ai/kling-video/ai-avatar/v2/pro) animate one face image (the first frame) to the audio and make a video exactly as long as the audio. There is no duration to choose: duration is the audio's length in whole seconds, rounded down, and that is what the user pays for. Send the audio inline (MP3 or WAV); the seller measures it and refuses a URL or another length with 422 invalid_video_duration before any payment. Quote: "<model>, <seconds> s from your audio: $<price>. Go?" (references/requests.md).

Edit a video

When the user brings a video and wants it changed (put them or another character in it, restyle it, change one thing), use a video-to-video model instead of the three stages above. The source video keeps its motion, timing, and camera; the model changes only what is asked. The steps in references/requests.md do the work.

1. Source. Get the video's path (in Antseed Desktop, get_chat_video_path with the id from the <uploaded-video> tag). Read its length, size, and frame rate with ffprobe, never by opening the file. MOV files are remuxed to MP4 without re-encoding; WebM, VP9, and AV1 sources are re-encoded to H.264. Stop if ffmpeg or ffprobe fails.
2. Model and limits. Fetch the catalog of video edit models (frames=edit). Recommend one for the job and say why (references/prompting.md): Recast or motion control to swap a person or character, Kling edit or Wan edit for a text instruction. Then read the model's source limits and required fields from its provider and check the source against them (references/requests.md): for a fal model, fal's published schema; for a Venice model, Venice's model list, then the same model's fal schema, then Venice's docs. Never quote limits from memory. Show the user each check and where its limit comes from. When the source fails one, offer the fix (trim, scale up, re-encode) and check again. When no limit is published, say so. Recast and motion control need a real person in the source: a source without one (an animal, a cartoon, an empty scene) comes back unchanged or fails.
3. Character and prompt. For a character swap, get or make one clear image per person to replace, in the order they appear left to right. Show the images and the prompt for approval, as with frames.
4. Price and go. The length paid for is duration: the source length in whole seconds, rounded down, on every edit model. For a shorter clip, trim the source first. Show the price per seller. When it is above $5.00, offer a lower resolution, a trimmed source, or a cheaper model. Ask: "<model>, <seconds> s from your video, <resolution>: $<price>. Go?"
5. Generate. After "go", create once with the block in Create the edit (not the new-video create block), then wait and save as for a new video. Recast can take 15 minutes or more. Show the result (in Antseed Desktop, show_media) and report the path, model, length, and price.

The sellers pass the edit to the upstream without checking the source's limits, so the checks in step 2 are the only ones before payment. The seller only checks that the source is inline and duration is its length (422 invalid_video_duration). Fix what the message names and ask before sending again.

Rules