auto-startups-vast

ADK multi-agent story-to-video: vision-grounded LTX motion prompts, Grok refs, LTX 2.3 I2V.

Publish AI-generated videos from a Google Sheet queue to YouTube and Instagram. Instagram Login API (graph.instagram.com), token exchange, Drive via gws. Use when publishing final_film.mp4 to social channels from Hermes VPS agents.

Turn story manifests into scene images using agent-composed prompts (prompt.json) and config-driven workflow templates. Supports model swapping (Qwen, HiDream, etc.) without code changes. Covers character sheet generation (Gemini or ComfyUI T2I fallback), prompt composition, batch scene generation, and vision-based evaluation (OpenRouter Gemini 3.1 Flash Lite with thinking tokens, or Gemini API fallback).

Parse ComfyUI workflow JSONs, extract all required models (UNET, CLIP, VAE, LoRA, checkpoints) and custom node packs, research download URLs via HuggingFace CLI (never download!), and generate self-contained provisioning-ready bash scripts for $REPO_ROOT/workflows/setup/. Scripts must be platform-aware (Vast.ai + RunPod) and end with a ComfyUI restart.

Turn stories into highly consistent cinematic videos using a deterministic 9-step pipeline orchestrated by Google ADK and wave execution.

Turn stories into highly consistent cinematic videos using a cloud-based image generation backend (Grok Imagine via fal.ai) and self-hosted LTX-2.3 FLF2V video generation.

Turn story manifests into highly consistent cinematic videos using a 3-stage model chain: Ideogram 4 (T2I) for scene still generation and character sheets, Flux Klein 9B (I2I) for character consistency editing (edit pass), and LTX 2.3 FFLF Seed Hunter for video generation. Utilizes the batch-wave model with automated quality evaluation gates.

Turn stories into highly consistent cinematic videos using Flux Klein 9B for all image generation and LTX-2.3 FLF2V for video — with reflexion-loop prompt engineering for the LF and motion prompters.

ComfyUI infrastructure specialist — API queueing, upload manifest with dedup + retry, FFLF template execution, tail frame extraction, per-story output isolation. Owned by stv-ops profile.

Binary pass/fail QC gate for AI-generated stills and video frames. Uses Gemini 3.1 Flash Lite via OpenRouter. Per-gate image budget enforcement. Strict JSON output. Owned by stv-reviewer profile.

Master orchestrator that creates Kanban boards for story production and dispatches tasks to the 6 specialist STV profiles. Loaded by the default aurora profile. Supports v1.0-native mode (14 decomposed tasks) as default.

Turn story manifests into cinematic videos using agent-composed prompts (filmmaking_prompt.json) and config-driven workflow templates. Uses the LTX 2.3 FFLF (First Frame Last Frame) Seed Hunter multi-stage workflow, supporting smart frame generation, multi-roll seed hunting, spatial upscaling, motion quality evaluation (V2 prompt: FF+LF anchored, 3 frames/video, a/b/c ranked, LF-arrival weighted), seamless continuation chaining, edit-instruction LF prompting, pre-flight FF↔LF audit, and opt-in per-image quality gating.

Provision, monitor, and manage Vast.ai GPU servers autonomously. Uses the vastai Python SDK (pip install vastai) — no CLI binary needed. Stable Docker image + bootstrap script approach. Includes extended diagnostics, tmux background execution, and strict SSH-only remote execution.

Maintain the auto-startups-vast repo — skill symlink management, commit/push workflow, and repo structure conventions.

Generate videos using LTX 2.3 (text-to-video) on a RunPod ComfyUI pod via REST API. Covers workflow submission, monitoring, output retrieval, and audio merging.

Provision, monitor, and manage RunPod Community Cloud RTX 3090 pods with runpodctl. Includes a budget-conscious provisioning script, lifecycle commands, remote workflow execution on existing pods, and troubleshooting notes for daily rent-and-destroy workflows.