Input nodes

The nodes that feed everything else: text and images going into the graph.

#Prompt

A freeform text node. Type a prompt; its output flows down text wires to any node that takes a prompt — Text→Image, Image→Image, the cinema agents, and so on. Prompts are the most-reused currency in a graph.

• Prompt presets — the select under the text box loads a saved prompt from the global library (all projects), 💾 saves the current text as one, and ✕ deletes the loaded preset. A brief you refined once — a Multi-Shot scenario, a house style clause, a recurring negative — becomes one click on any future graph. Hand-editing the text detaches the preset name; the text on the node is always the truth. All saved presets are also listed (and deletable) in File ▾ → Browse library….

#Refine (Prompt Helper)

An AI prompt-rewriter. Wire a Prompt (or any text) into Refine, choose a preset, and it rewrites the text through a text model (Gemini) — sharpening a vague idea into a precise, photorealistic prompt without changing your intent.

Presets are data (fifteen ship by default), including:

• Enhance prompt — the general-purpose sharpener.
• Screenshot → photoreal, Cinematic relight, Camera language, Scene dressing, and more.
• Character turnaround sheet and Character expressions — wire a character image in and get an image-editing prompt for a 4-view orthographic turnaround, or a grid of facial expressions, both identity-anchored to the image with hard don't-change constraints. (The same recipes come pre-wired in the Character Lab library composite.)
• Scenario maker — writes a brief for a Multi-Shot custom scenario (chain Prompt → Refine → Multi-Shot notes).
• Kling video prompt — structures a loose motion idea into Kling's five-part video prompt (scene / subject motion / explicit camera / atmosphere / negative constraints), with the stillness discipline the cinema crew uses. Tell it in the notes which pipeline stage you're at (look test vs final render).
• Seedance video prompt — the same job in Seedance 2.0's grammar: shot-storyboard structure, one camera move per shot, controlled motion, and Seedance's native audio symbols (（…） ambience, <…> sound effects, {…} dialogue).
• Refine shot — plan-aware: feed a shot-plan JSON in and it edits one shot.
• Seedance shot plan — plan-aware: feed the Script Supervisor's whole plan in and every clip prompt is rewritten into Seedance 2.0 grammar with scene-matched audio cues, while ids, still indices, keyframe prompts and frames come back untouched. Splice it between the Supervisor and a Storyboard to run the same film on Seedance — a second Storyboard fed this way keeps its cached keyframes, so a Kling-vs-Seedance A/B never re-bills the frames. Use the notes to direct the soundscape ("birdsong on the terrace, fireplace crackle at dusk").

Some presets are multimodal — wire an image in and the model sees it. Refine's output is editable text: hand-tweak it and your edits flow downstream; change an input or roll again and it regenerates over them.

#Text → Text

A free-form LLM call — a ChatGPT/Claude turn as a graph node. Pick any text model from your providers (Gemini, GPT or Claude — your keys, like everywhere in Brut), optionally give it a role in the System field ("you are a harsh photography critic"), type a message, run.

• Wired text is prepended to your message — chain Prompt → Text→Text, or feed one node's answer into the next. A chain of Text→Text nodes is a conversation materialized on the canvas: every turn editable, cached, and re-runnable.
• Wired images are seen — connect renders or photos and ask "what's wrong with this composition?", "describe this scene for a video prompt", "compare these two options". Images are downscaled for the model's request cap; nothing is uploaded anywhere but to your chosen provider.
• The answer is editable text — hand-tweak it and your edits flow downstream on the next run; 🎲 Re-ask requests a fresh answer.
• On models with a reasoning dial — GPT-5.6 (Sol / Terra / Luna, none → xhigh, details) and Claude Sonnet 5 / Opus 4.8 (low → max, details) — a Reasoning select appears: how deeply the model reasons before answering. Changing it is a new answer (cache miss), like any input change.
• The node is width-resizable — select it and drag the right edge; long answers get room while height follows the content.

Unlike Refine (which rewrites prompts through fixed presets), Text→Text is open-ended: analysis, critique, translation, naming, alt-text, brainstorming — anything a chat assistant does, with the result flowing down a text wire. One-shot by design: for an ongoing conversation with memory, talk to Marcel. Calls are token-priced (typically well under a cent) and land in the ledger like every paid call.

#Text Board

The Shot Board for text: wire one long text in — a Text → Text answer, an agent plan, any pasted script — and it splits into numbered, editable blocks, each with its own output plug (#3 routes straight into a Text→Image while #4 feeds an Image→Video). Free and local: the splitting is deterministic code, never a paid model call.

• Split modes: Auto-detect tries SHOT X fenced blocks (the script format many studios use), then JSON arrays, numbered lists, and finally blank-line paragraphs; or pick a mode explicitly, including a custom separator.
• Blocks are editable in place — tweaks flow out the plugs immediately on the next run. Changing the wired text or the split mode re-splits (edits belong to the old content and reset with it).
• ✕ hides a block from the combined bottom output (its number and plug stay stable); ↺ restores it.
• The board grows to fit its blocks — every tile (and its plug) stays visible, so wires always attach where you see them. Shrink it by hand if you prefer; it only re-grows when the content changes.

The killer combo: a scripted Text → Text (your reel/script prompt in its System field, renders wired to its images inlet) → Text Board → per-shot prompts fanned out to your image and video nodes. One click from "one long ChatGPT answer" to a wired production board. (For the specific case of cinematic reels, the cinema crew's Cinematic Reel (flourish) brief now folds this whole flow into one node chain — the Text → Text front-end stays the flexible, roll-your-own route.)

#Image In

Loads images into the graph. Add them by file picker, drag & drop, or paste — all append, so one Image In can hold several images. Each thumbnail's 🔍 opens the lightbox at screen resolution — ←/→ pages through the node's images, wheel-zoom loads the original pixels only when you actually zoom. Dropping image files anywhere on the canvas creates a ready-filled Image In node right where they land — and so does pressing Ctrl+V over the canvas with an image on the clipboard (copy from Photoshop, a render buffer, or any browser image; a new Image In appears at your cursor). To append a pasted image to an existing Image In instead, click that node first and then paste, or use its Paste button.

Results can be detached the same way: drag any image off a node — a generation preview, a Multi-Shot or Shot Board tile, a Reference Board still, a Storyboard card frame — and drop it on empty canvas to get a ready-filled Image In right there (drop it on an existing Image In to append instead). No re-upload, no copy: the new node points at the same project file, now frozen as a base for further exploration even if the source node later re-rolls or is deleted from the plan.

File vs. paste, for big renders: adding a file (picker or drag & drop) keeps its original bytes — a 12 MB JPEG stays a 12 MB JPEG. Pasting can't: the clipboard carries decoded pixels, not your file, so Brut stores them losslessly as PNG (never a silent re-compression) — a 20-megapixel render lands around 38 MB that way. Same pixels, bigger container. Paste is perfect for screenshots and quick grabs; for production renders, prefer the file — smaller payloads and the exact encoding you delivered.

Image In is the entry point for image-to-image work, composition, inpainting, and the stills that ground the cinema crew. For the crew, your stills are the project's identity anchors, not a shot list — the film can hold more shots than you have stills: invented compositions generate their start frames from these anchors, so every frame stays your building.

#Next

→ Image generation — turn these into pictures.