S-Forge

S-Forge is Erebros’s AI film director, at /s-forge. An idea becomes a written package — scenario, character boards, voice profiles, environments, storyboard and per-clip video prompts — and then the boards and the clips are rendered for real.

What it does

S-Forge writes the whole film package as .md, generates the stills each prompt describes, and renders video clips through your chosen engine. It began as a test polygon cloned from the older Storyboard Forge, outgrew it, and in 6.1.5 replaced it. Everything a project owns lives in one vault folder: s-forge/<seedance|flow>/<project>/ (Scroll World sites under s-forge/scroll-world/).

The manual is in the app: the ? button in the S-Forge header opens the richer S-Forge guide (source: docs/S-FORGE-GUIDE.md). It covers the pieces that are easy to get wrong — what each Veo tier does with your bound stills, why an Image Studio render is a draft until you push it, and what @Voice_Nova actually is.

Video engines

Pick in TARGET (DISPATCH panel):

• Seedance (fal) — 2–12 s per clip, 480p / 720p / 1080p. Billed by video tokens, so resolution changes the price: 1080p costs 2.25× a 720p clip. Renders on Seedance v1 Pro Fast.
• Veo 3.1 Lite / Fast / Standard (Google Vertex) — 4, 6 or 8 s only; 720p and 1080p cost the same, only the tier moves the price ($0.05 / $0.15 / $0.40 per second).
• Auto follows the pipeline: Flow → Veo, Seedance → fal.

The live cost of the current clip is shown under SHOT LENGTH and updates with engine, tier, resolution and duration.

Pipelines

The DISPATCH pipeline picker is a dropdown: Veo Pipe / Seedance Pipe / Scroll World. The two film pipelines are different directors, each built from its own source documents, and each writes its own .md:

• Seedance Pipe (fal) → A. Director's decision · B. Continuity lock · C. Reference map · D. Prompts (Stage 0 environments → 1 character sheets → 2 storyboard grids → 3 video → 4 voices) · E. Run order.
• Veo Pipe (Vertex Veo 3.1) → the full film-director package: A. Director's decision · B. Visual DNA · C. Project Bible (character / style / lighting / environment locks, master palette, camera philosophy) · D. Reference map · E. Production blocks 1–5 (character boards · voices · environment boards · 6-panel storyboard · clip plan + video prompts) · F. Run order · G. Preflight. It detects its input mode (idea only / video reference / style reference / existing Bible), and with a video or style reference it extracts the Visual DNA and applies its logic to your idea.

Running the pipeline

▶ Run pipeline writes the whole package and hands you the .md.

• Mode: Step pauses after each stage for your approval — and waits for the director to actually finish writing before asking.
• Click any pipeline node to read and edit what it produced. Saved edits go into the .md.
• Prose and prompts follow OUTPUT LANG (EN/UA/RU); the structural labels (Stage 1, @image1, CHARACTER 1, NO MUSIC) always stay English — the app routes blocks by them.

STYLE (the genre)

One STYLE choice pins four things the AI-filmmaking guides treat as one decision: the route (universal, or the hyper-kinetic action bible), the STYLE LOCK copied verbatim into every prompt, the storyboard’s third annotation line (MOOD / VOICE / STYLE), and the aspect ratio. There are 15 presets, from Live-action cinematic to AAA game render (photoreal CG on the universal route) to Influencer vlog (9:16).

Scroll World — the third pipeline

Unlike the two film pipelines, Scroll World does not make a film — it opens its own studio and builds a scrolling website: the visitor scrolls and a pre-rendered camera dives continuously down through a sequence of scenes, with the seams between scenes frame-locked so there is no visible cut.

Describe a “dive” idea, press ✧ Draft scenes with AI to have the director design the scenes (each editable), pick your engines, then ▶ Build — it renders live and previews in the window. Finished sites save under s-forge/scroll-world/. Two engine choices:

• Image engine — scene stills, always free on Gemini.
• Video engine — Seedance (fal), Vertex Veo (Lite/Fast/Standard), or Stills only (free), a no-video parallax build that spends nothing.

The full walkthrough is in the in-app ? manual.

Workshop tabs

• Images (including characters)
• Videos
• Characters
• Scene Prompts
• Scenes — the assembled cut. The scene builder ships in a later build.

Click any thumbnail to zoom it in place — ← / →, the mouse wheel or the ‹ › buttons page through the gallery; a click on the picture closes it again.

Scene Prompts

Holds the whole package: the SCENARIO .MD (the pipeline writes it into the project folder — ⧉ Copy, or ✎ Open to edit it in place) plus every generation prompt as its own card — ENVIRONMENT (Stage 0), CHARACTER (Stage 1), STORYBOARD (Stage 2), VIDEO (Stage 3) and VOICE (Stage 4). Each card has ⧉ Copy, ✎ Edit and ↗ Push to chat, so you can review a prompt in the director chat before spending money.

Scene Prompts is also reachable from the header, next to Ingredients — the same cards, with ⧉ Copy and ⇩ Download, readable from anywhere (including with Image Studio open).

Reference modes (Veo)

What the app binds in ⬡ Ingredients is what the renderer gets, and each mode does something different — the director is told which ones are live before it writes a single clip:

• TEXT — no reference.
• FIRST-FRAME — one still becomes the clip’s literal opening frame. Bind a scene still, not a character model board (a board handed to this mode animates as a drifting turnaround sheet).
• INGREDIENTS — several stills condition the character’s identity without pinning frame one. Veo Standard only (Lite and Fast fall back to FIRST-FRAME).
• FIRST + LAST — mark a scene still with ⇥ and Veo generates the motion between the opening frame and that one. Works on every tier, Lite included. Use it for a match cut, or to land a clip exactly on the next one’s opening frame.

Image engines

The Image Studio model picker offers the fal catalog and the Codex CLI, which costs nothing per image — it runs on your ChatGPT subscription (no API key):

• Codex — OpenAI image — its built-in image_gen tool. Photoreal stills, concept art, sketches.
• Codex — code-drawn board (SVG) — Codex writes the board as vector code. Use it when the picture is really an infographic: a model board or environment sheet is callout lines, swatches and a dozen labels, and diffusion models turn text into letter-shaped noise. Vector text always spells correctly.
• fal (FLUX, Seedream, Nano Banana, …) — everything else, billed per image.

Codex takes ~1–2 minutes per render (it is an agent, not an endpoint) and needs the codex CLI logged in.

Image Studio (◈), Ingredients (⬡) and Projects (▤)

• Image Studio (◈) generates the stills the prompts describe: Character mode makes an annotated model board (with a voice from Supertonic — local, free, 31 languages — or Kokoro, or ElevenLabs); Image mode makes any other still — environments, storyboard grids, anything.
• Ingredients (⬡) binds the project’s characters / scenes / voices into the generation, so the director writes for the cast that exists instead of inventing a new one every turn. Rename an asset with ✎ (Aria @image1) and the name rides into the director’s prompt. A bound still is also the render’s opening frame (Seedance and Veo Lite/Fast take exactly one) — bind a scene still for a shot; only Veo Standard does true multi-reference identity.
• Projects (▤) — History and Archive, with ✎ rename, ⇩ archive and 🗑 delete. Every project has a PROJECT ID (SD-0001), and that id — never the title — is the vault folder, so you can rename a project to anything, in any language, and its renders stay put. Deleting removes the dashboard entry only: the files in your vault are never touched.

Deleting bad renders

Use 🗑 on every gallery tile and every row of Ingredients. This matters beyond tidiness: the gallery and Ingredients list every file in the project folder, so a wrong image stays a candidate reference that could be bound and fed to the next clip. Deleted files move to the project’s .trash/ folder in the vault — nothing is destroyed.

Setup

• Vertex (Veo). Open the ⚙ panel in the header and upload your Google service-account JSON. It is stored at secrets/vertex-sa.json (gitignored, 0600) and re-read on the next render, so you can rotate to a fresh GCP account (new $300 credit) from the browser — no env edit, no restart. Full guide with the free credits: docs/VERTEX-AI-SETUP.md.
• Seedance / fal images. Set FAL_KEY (or FAL_API_KEY) in .env.local — powers FLUX images, Seedance video and character model boards.
• Codex images. Requires the codex CLI logged in (codex login, ChatGPT subscription — no API key).
• Character voices. Free local Supertonic (provisioned by scripts/setup-free-engines.sh) or Kokoro, or ElevenLabs (ELEVENLABS_API_KEY).
• Scroll World additionally uses ffmpeg for seam-frame extraction and clip encoding.

Swarm integration

Every Hermes Swarm node has an S-Forge binding — none, Director, Boards or Clips. When bound, that node’s pipeline step runs S-Forge’s own libraries instead of the Hermes CLI while staying a normal bot otherwise. The ❖ Swarm button in S-Forge seeds three such nodes for you.

Tips

• An Image Studio render is a draft until you push it into the project — see the in-app ? manual for why.
• For a shot, bind a scene still (not a character board) unless you are on Veo Standard INGREDIENTS mode.
• Rotate GCP service accounts from the ⚙ panel to keep using fresh Vertex credit without touching env files or restarting.

See also

• Hermes Swarm — bind nodes to S-Forge (Director / Boards / Clips).
• M-Forge — the video app; its Studio pipelines are also ported paid skills.
• Orchestration