higgsfield-canvas

# Higgsfield Canvas — Node-Based Production Workspace

Canvas is a node-based editor where prompts, references, and generations from
any Higgsfield model live on a single infinite board. Instead of running each
generation in isolation, you connect nodes — a prompt feeds an image, an
image feeds a video, a style branches into variations — so a whole scene or
campaign becomes one continuous, re-runnable pipeline.

Canvas is its own top-level surface in the Higgsfield nav bar (alongside
Image, Video, Cinema Studio, Marketing Studio). It is model-agnostic: it
*hosts* the other models rather than being a mode inside any one of them, so
this sub-skill routes model-specific craft back to the relevant sibling
sub-skill.

**Why Canvas matters for this skill.** Canvas is another place to *deploy the
prompts this skill builds* inside the Higgsfield workflow — instead of taking
a finished prompt straight to the Image/Video generator or Cinema Studio, you
drop it into a **Prompt node** on the board and wire it into a generator. The
prompt-craft is identical (MCSLA scene prompts, Seedance prompt modes, GPT
Image 2.0 formats — all still apply); Canvas just changes *where* the prompt
is executed and lets one prompt feed a multi-step pipeline. Think of the
Prompt node as the on-canvas home for everything the rest of this skill
teaches.

Translated from Higgsfield's official Canvas guide (`higgsfield.ai/canvas-intro`)
plus the in-product Canvas screenshots and a canvas-workflow walkthrough.

---

## 1. When to use Canvas

Reach for Canvas when the work is a *pipeline*, not a single shot:

- A multi-shot scene or episode where one reference set drives several shot
  prompts.
- An ad campaign that fans one product/brand reference into many variants.
- A storyboard that flows assets → story beats → prompts → image/video
  generation in columns.
- Any repeatable pipeline you want to save as a template and re-run with
  swapped inputs.

For a single prompt-and-generate, the standard Image/Video tabs are simpler.
Canvas earns its keep when steps connect.

## 2. Node categories

A node is one step in the pipeline with typed input sockets (left) and a
typed output socket (right); curved edges carry data from one node's output
into the next node's input. The board exposes these node categories:

- **Prompt** — text starting points.
- **Image generator** — produces or edits images.
- **Video generator** — produces video.
- **Voice / audio generator** — produces voice or audio.
- **LLM assistant** — a wired-in language model you can task with prompt
  drafting or shaping (e.g. a "cinematographer-prompter" role inside a
  storyboard). It is a general LLM node you author, not a fixed named
  assistant. (Separate from **Higgsie**, the in-Canvas chatbot — that's a
  conversational helper, not part of the prompt-to-generator pipeline this
  sub-skill covers.)
- **Style / motion** — style-transfer and motion-control steps that branch a
  look or movement across downstream nodes.
- **Render output** — final deliverables.
- **References** — **Upload** (drop in your own files) and **Assets** (pull
  from your library).

Connections are type-aware: an image output feeds an image input, a prompt
feeds a generator, and so on. Fan-out (one node feeding many) is free-form;
fan-in is more structured.

## 3. Models that run inside Canvas

All seven Higgsfield models run as nodes on the same board:

| Model | What it's for | Detail |
|---|---|---|
| **Soul 2.0** | photoreal images, fashion/editorial, consistent characters | `../higgsfield-soul/SKILL.md` |
| **Seedance 2.0** | flagship video, 1080p | `../higgsfield-seedance/SKILL.md` |
| **Kling 3.0** | cinematic video, multi-shot continuity, native audio | `../higgsfield-motion/SKILL.md` |
| **Wan 2.7** | high-action / dynamic motion | `../../model-guide.md` |
| **Veo 3.1** | Google flagship video, native audio | `../../model-guide.md` |
| **GPT Image 2.0** | 4K image gen, near-perfect text rendering | `../higgsfield-gpt-image-2/SKILL.md` |
| **Nano Banana Pro** | precision image editing and placement | `../../image-models.md` |

Mixing models on one board is the point — generate a character in Soul,
edit a product in Nano Banana Pro, animate the result in Seedance, all in one
graph.

## 4. Named canvas patterns

The in-product canvases ship as recognizable layouts. Use them as starting
shapes:

- **Simple Seedance 2.0** — one image + one prompt → one video. The minimal
  pipeline.
- **Extend Video** — a sequential chain where the last frame of one
  generation feeds the next, for longer continuous sequences.
- **Image Edit** — multi-reference compositing into one image generator node.
- **StoryBoard With Elements** — a three-column flow (Assets → Story →
  Prompts) feeding image/video generators, often with an LLM node shaping the
  shot prompts.
- **Long Video (fan-out)** — one reference set fanned into several shot
  prompts that each generate independently, for multi-shot assembly.

Any board can be saved and reused (see § 6).

## 5. Cost model

Building the graph is free. You pay only for the models you actually run:

- Connecting nodes, editing prompts, staging a board, and arranging a
  50-node storyboard cost **no credits**.
- Credits are deducted only when a node **generates** an image or video, at
  the same rate that model charges elsewhere on Higgsfield.

So you can stage an entire campaign pipeline and pay only for the renders you
choose to execute. (Per-node credit badges shown in the UI reflect the
underlying model's rate; this sub-skill does not restate specific credit
numbers — confirm live in-product.)

## 6. Templates and assets-as-nodes

- **Save any canvas as a reusable template** — for ad variants, character
  sheets, storyboards, or any pipeline you run repeatedly. Duplicate it and
  swap the inputs.
- **Your assets drop in as nodes** — Soul ID characters, uploaded products,
  brand references, and any of your previous generations can be added
  directly to a