The build log for Showspring — the independent, off-hours AI project behind The Doodle Cast. The project itself is at showspring.com. This page is the workshop: the pipelines, what runs underneath, and what’s on the bench next.
Veo, Kling, Higgsfield, Firefly — remarkable at generating individual video clips. But producing a complete episode — with narrative structure, multi-character dialogue, consistent visuals, sound design, and music — still takes dozens of hours of manual stitching and editing. Showspring closes that gap.
Three production pipelines (full episodes, shorts, podcasts), a shared show bible, and a render engine built from scratch for AI-native output. Solo build, deployed nightly, tested in production by an actual show.
The full episode pipeline. Each step has its own workspace, its own AI assist, and persists state across reloads. Skip ahead, jump back, regenerate one scene, ship the whole thing.
Three AI personas independently pitch episode concepts, then debate their merits. A judge model evaluates each pitch with live web search for topical relevance and picks a winner — or you bring your own idea and let the panel pressure-test it.
Generate a full script with structured clips, character dialogue, scene descriptions, and image prompts — on the LLM of your choice. The writer is grounded in the show bible: a living knowledge base that evolves with every episode, keeping characters consistent and plotlines from repeating.
Every character speaks in a distinct synthesized voice. Play through the full episode readout to check pacing, dialogue flow, and structure before committing GPU cycles to visual production. Cheaper to fix a script than a render.
AI extracts every location mentioned in the script and maps them to clips. Each location lives in a reusable library with reference images, descriptions, and default prompts. The studio always looks like the studio — week 9 matches week 1.
Generate images for each clip, informed by character reference sheets, location plates, and scene context. Start and end images per clip enable smooth image-to-video. Full history with undo, AI-assisted editing, and an iPad/PC watcher for hand-drawn or live-camera source frames.
Turn scene images into motion using cloud video models (Veo 3.1, Kling 3.0, Seedance 2.0) for production quality, or open models (WAN 2.2, Z-Image-Turbo) on a local RTX 5090 for free iteration. An NLE-style timeline shows every clip with status badges and a composite preview player that sequences completed clips in real time.
Frame-accurate in/out markers per clip. Adjust durations, preview instantly. The trimmed timeline carries forward to the audio mix and final render — one source of truth from edit through publish.
Four-lane multi-track editor: background video audio, voice dialogue, sound effects, music. Independent volume per track with keyframe automation. Generate SFX and music from text descriptions, drop them on the timeline, fine-tune the mix — all in the browser.
One button, full episode render. The engine trims each clip, applies the audio mix through hand-rolled ffmpeg filter graphs, concatenates everything, and encodes the final MP4. When the local GPU is online it runs on NVENC; otherwise CPU fallback keeps shipping.
AI thumbnails for A/B testing, AI-written metadata, tags and categories, then direct publish to YouTube with real-time upload progress. The show bible learns from each episode that ships — what landed, what didn’t — and feeds it back into the next pitch.
Eight shorts at a time from a single theme — each with unique characters, dialogue, image, animation, voice, and music. Publishes across YouTube, TikTok, Instagram, Facebook, and X with scheduled auto-posting and AI-recommended slots.
Eight ideas at once from the character pool, with optional research-report grounding for topical hooks.
Per-clip scene + action prompts, 8–12 word dialogue, visual-coherence rules baked in for clean I2V.
9:16 starting frames grounded in character references, with full image history and per-clip regen.
Image-to-video on cloud models for production runs or local GPU for free iteration. Same operator surface.
Multi-track audio per short — original, voice, and AI-scored music. Batch-replace or fine-tune.
Batch render all eight, or selectively re-render. Outputs land in Drive and are downloadable on the spot.
Five-platform fan-out, OAuth-authenticated, visual calendar, AI-suggested posting times, alerts on misses.
The whole posting plan at a glance — per platform, per day — with the auto-publisher in the loop.
Conversational episodes where the show’s characters discuss real topics — multi-voice dialogue, AI cover art, and direct publish. Same characters, same continuity, different format. 2 to 120 minute targets.
AI-pitched concepts with topic scoring, bring-your-own ideas, or deep-research mode with live web grounding.
Conversation scripts with character dialogue, word-count tracking, and runtime estimates against your target.
Up to four AI variants per episode, informed by character refs and the episode brief. Pick one or regen.
Full episode preview with character avatars and color-coded dialogue. Review the flow, then ship to the feed.
Text-to-dialogue API generates natural multi-character conversations in one audio stream — turn-taking and all.
Auto intro/outro, configurable silence gaps, act-based batching for 15+ minute episodes — publish-ready audio.
Two systems that turn a flat clip sequence into something that feels like a show. A late-night news-desk segment generator, and a per-clip audience reaction track sourced from a multi-take stem library.
Extensible segment templates encode format rules, comedy mechanics, joke structures, and a small visual style guide — baked into the AI script step itself. Anchor solo shots and two-shots are rendered with OTS news graphics composited per character in the render pass.
Per-clip audience-reaction track from a multi-take stem library — thirteen reaction types, density dial, and a multimodal cue planner that listens to the cut and picks where laughs, gasps, and silence land. The audio bed is generated for the episode, not borrowed from a library anyone else can use.
The same show bible that produces the episodes powers two more things: live voice conversations with the cast, and a pipeline that turns viewers’ own dogs into cast members.
Public cast chat on thedoodlecast.com: pick a character, tap the mic, talk. Each persona is assembled from its Showspring show-bible section (auto-refreshing when the bible changes) and speaks through the same ElevenLabs voice ID used in the episode pipeline. New cast members become talkable the moment they get a bible section — no per-character code.
A viewer submits photos and a few lines (multi-dog households supported). A multimodal model turns the photos into structured character drafts; Showspring generates a per-dog portrait, a “Hello” voice line, and a video preview; the submitter picks their dog’s voice through a shareable voice-pick link. On approval the dog is promoted to a cast character with its own bible section and pinned voice — episode-ready, and instantly talkable in the live chat.
Pipelines come and go. The studio that holds them together — characters, episodes, media — persists across every show, every episode, every week.
Personality, visual description, speech style, reference images, voice profile. The cast persists across all episodes and informs every generation.
Every episode across every stage of production. Filter by status, search by title, jump straight into any production step.
Every image and clip across every episode. Browse by model, date, or episode. Drag-and-drop, crop, rotate — with full undo.
First-class sets and props with their own galleries, references, and history. The same desk in episode 14 as in episode 2.
One screen for the state of the studio — what’s rendering, what’s queued, what’s ready to ship.
What actually landed, per platform, per episode — routed back into the show bible so the next pitch starts smarter.
The interesting design choices live in how these layers talk to each other — not which logos sit on which row.
Multiple top-tier LLMs and image / video models, each pointed at the step where it’s actually best. Routed through a single internal bridge with usage tracking, budgets, and graceful fallbacks when a vendor blips.
A local RTX 5090 handles the high-volume, low-stakes batch work that would otherwise burn cloud quotas — tagging, triage, draft scoring, image and video iteration. A llama-swap router serves Gemma 26B for text; WAN 2.2 and Z-Image-Turbo cover local visuals. Same bridge, different destination.
The part nobody else has: the render engine, the show bible, the continuity system, the publish pass, the analytics feedback loop. Deployed nightly, tested in production by a real show.
Why hybrid: production-grade output runs on the best cloud model for the step. Iteration runs on the local box for free. The operator never picks — the studio routes.
The project page has the story, the proof, and a way to get in touch if you have a show you’d want to put through it.