You design.
Your hardware renders.

Five steps from download to reel.

1. Unzip the pack into a working directory.
2. Edit prompt_template.json for your scene. Each scene block has prompt, negative prompt, duration, and settings hints. The schema is model-agnostic.
3. Replace the RENDER_CMD placeholder in render_pipeline.sh with the actual command line for your local video diffusion model. Examples for ComfyUI, AUTOMATIC1111 + AnimateDiff, diffusers (HunyuanVideo / CogVideoX / Wan2.1) are commented in the script.
4. Render all scenes: bash render_pipeline.sh. One MP4 per scene lands in ./raw/.
5. Watermark and assemble: bash watermark_overlay.sh && bash assemble_reel.sh. Final cut at ./reel_1080p.mp4.

Anywhere you have a GPU and ffmpeg.

• Local workstation — RTX 4090, RTX 5080, RTX 5090, M-series Mac with sufficient VRAM. Most affordable path.
• RunPod / Vast.ai pod — spin up an A100 or H100 by the hour, mount the pack, render, shut down. Pay only for actual render time.
• Cloud account — AWS p4d/p5, GCP A2/A3, Azure ND-series. You retain full control of network, IAM, and storage.
• Studio render farm — drop the pack into your existing pipeline and call render_pipeline.sh from your job scheduler.

One rule on outputs.

Outputs you publish or distribute must preserve the visible “made with neyralfoundry.com” identifier embedded by watermark_overlay.sh, until you have a separate written agreement with us removing the obligation. This is the trade we ask for letting you use the pack at zero cost during alpha. Full text in Terms and Copyright & trade-secret notice.

Six small files.

• README.md — the documentation you are also reading abridged here.
• prompt_template.json — scene schema, two example scenes, model-agnostic.
• render_pipeline.sh — loops over scenes, calls your model. Placeholder for the actual model invocation.
• watermark_overlay.sh — ffmpeg recipe burning the Neyralfoundry watermark on every scene.
• assemble_reel.sh — concat all watermarked scenes with 0.4s crossfades into a single reel.
• LICENSE.txt — the legal terms specific to this pack.

That is it. No installer. No background process. No phoning home.

Your renders are yours.

Because the pack runs entirely on your hardware, Neyralfoundry never has a copy of your inputs, your outputs, your prompt iterations, or your render logs. We could not produce them under subpoena even if we wanted to. This is by architecture, not policy — and an architecture like this is usually quick for a studio security team to review, because there is far less to check.

Limits of the alpha.

• No hosted-render fallback. We deliberately do not offer cloud compute during the alpha to avoid burning capital we do not have.
• No collaboration features. Single-operator only at this stage.
• No bundled model weights. You bring your own.
• No SLA. Best-effort issue triage via the contact form.

Stay current.

This is v1 of the BYOC starter pack. Future versions ship improvements to the prompt schema, the watermark recipe, and the assembly pipeline. Bookmark this page; the latest is always at /byoc.html. Major changes will be announced to the waitlist.