13:20
AI image generation
Higgsfield's new AI Canvas is a node-based workflow builder that lets you chain prompt, image-generation, video-generation, and LLM nodes together in one place, giving access to every major AI model at once and letting you build repeatable, shareable content pipelines instead of jumping between separate tool tabs.
Canvas is accessed at higgsfield.ai by clicking 'Canvas' then 'New Canvas'; you build by right-clicking to add nodes (e.g., a 'prompt node') and dragging connections between them.
There is no 'send' button in Canvas; instead you connect a prompt node into a generator node and click 'Run Pipeline' to execute.
Canvas navigation uses Cmd/Ctrl + scroll to zoom, plain scroll to pan, and two-finger trackpad gestures to move around.
The AI Image Generator node exposes every major AI image model on the market (not just GPT Image 2), removing the need for 20 separate subscriptions.
You can fork one prompt into multiple image-generator nodes with different models (e.g., GPT Image 2, Seedream 5.0 Light, Grok Imagine) and run them simultaneously to compare results; in the video's own comparison, GPT Image 2 was judged the most realistic.
A single image or video made this way is functionally identical to just using Higgsfield's native 'Create Image' feature directly, and the video states that for that one-off use case, the node-based workflow builder 'is absolutely useless.'
Chaining is where Canvas becomes valuable: e.g., connecting a generated cat image into a Video Generation node with a motion prompt ('make the cat stand up and jump off'), picking a video model like Stable Diffusion 2.0 Fast, and generating.
Both image and video nodes have an 'Open Settings' panel controlling model, duration, resolution, aspect ratio (e.g., 9:16 for TikTok/Instagram), a 'generate audio' toggle, and number of outputs — flagged as important because otherwise you can end up waiting on an expensive silent video.
Canvas enables repeatable content pipelines: you can swap an input (e.g., 'cat' → 'gray cat') and rerun the whole saved workflow from scratch, which becomes valuable at scale, e.g., a 50-node product catalog workflow, instead of repeating steps manually across tabs.
Workflows can be published: click the share control, go public, and set a permission level such as 'anyone can join' as viewer — viewers can see prompts/content and clone the workflow but cannot edit the original; sharing is done via 'copy link.'
The LLM Assistant node runs Claude, Gemini, or ChatGPT inside a single node and can take other nodes (images, prompts) as inputs to perform an operation on them, e.g., 'describe the characteristics of this person' given an uploaded photo.
A full demonstrated example: an Adidas shoe photo plus a prompt (goal: 'UGC AI video advertisement,' plus a product landing-page URL and a reference photo) is fed into an LLM Assistant node to generate an ideal persona/scene description; that description plus the shoe image feed an image-generator node to create the 'actor' image; that image feeds a second LLM node (using Claude Sonnet) to produce a video script and audio/spoken-line instructions; those feed a video-generator node (Seedance 2.0) to produce the final UGC-style talking ad video.
The creator states there is a separate, more detailed step-by-step guide on building the UGC advertisement workflow, plus a link to clone the exact workflow shown, both in the video description.
Prompt Node — A canvas element added via right-click → 'prompt node' that holds a typed text prompt but does not generate anything by itself. Apply: Right-click the canvas, add a prompt node, type a prompt like 'create an image of a cat,' then drag a connecting line from it into a generator node before running the pipeline.
AI Image Generator Node — A canvas node that turns a connected prompt into an image, with a dropdown giving access to every major AI image model on the market (e.g., GPT Image 2, Seedream 5.0 Light, Grok Imagine). Apply: Connect a prompt node into it, choose a model from the dropdown, open its settings, and click 'Run Pipeline'; duplicate the connection into several image nodes with different models to compare outputs side by side.
AI Video Generator Node — A canvas node that turns a connected image plus a motion instruction into a video, with access to every major video model (e.g., Stable Diffusion 2.0 Fast, Seedance 2.0). Apply: Connect a generated image into the node, write the motion prompt (e.g., 'make the cat stand up and jump off'), choose the model, then open settings to set duration, resolution, aspect ratio, and 'generate audio' before generating.
LLM Assistant Node — A canvas node that runs a large language model (Claude, Gemini, or ChatGPT) and can take other nodes — images, prompts, URLs — as inputs to perform an operation on them. Apply: Connect an image or prompt into the LLM node, write an instruction (e.g., 'shortly describe this person's characteristics'), pick a model, and hit run generate; feed the text output forward into other nodes.
Run Pipeline — The execution command for a canvas workflow, used instead of a chat-style 'send' button since Canvas is a node-based workflow builder. Apply: After wiring nodes together (prompt → generator → generator, etc.), click 'Run Pipeline' to execute the whole connected chain at once.
Open Settings panel (per node) — A per-node settings panel on image and video generator nodes controlling model, duration, resolution, aspect ratio, a 'generate audio' toggle, and number of outputs. Apply: Open a node's settings before running it to, for example, set a 9:16 aspect ratio for TikTok/Instagram or enable audio generation so the resulting video isn't silent.
Repeatable workflow re-run — The practice of saving a wired-up node chain and re-executing it end-to-end with a new input, instead of manually repeating each step across separate tool tabs. Apply: Swap the input on an existing workflow (e.g., change the prompt from 'cat' to 'gray cat'), select the whole workflow, and click 'Run Pipeline' to regenerate every downstream node automatically — useful at scale, e.g., a 50-node product catalog.
Public workflow sharing — A publishing feature that lets a canvas workflow be made public with a configurable permission level (e.g., 'anyone can join' as viewer). Apply: Click the share control in the corner, set the workflow to public with viewer permissions, and click 'copy link' to distribute it; viewers can see prompts/content and clone the workflow but cannot edit the original.
UGC AI video advertisement workflow — A composite multi-node pipeline that turns a product photo into a talking-actor UGC-style ad video by chaining an LLM node, an image generator node, a second LLM node, and a video generator node. Apply: Feed a product image plus a goal prompt (with landing-page URL and reference photo) into an LLM Assistant node to get a persona/scene description; feed that plus the product image into an image generator node for the actor image; feed the actor image into a second LLM node (e.g., Claude Sonnet) to produce a video script and audio instructions; feed image and script into a video generator node (e.g., Seedance 2.0) to produce the final ad video.
The video volunteers its own limitation up front — explicitly calling the node-based approach 'absolutely useless' for single outputs — rather than oversell the feature universally; its entire pitch rests specifically on repeatability and scale, not on any single generation being better.
Because Canvas exposes essentially every major image and video model in one interface, the video effectively uses it as a model-comparison bench: forking one prompt into three parallel generator nodes to A/B test models against each other, not just as a generation shortcut.
Publishing a workflow with 'viewer' permissions converts a personal pipeline into a distributable template/product — others can clone the exact prompt chain and model choices for their own use without being able to alter the source workflow.
The UGC ad example uses the LLM Assistant node twice for two different jobs inside one pipeline: first to turn a product photo + goal prompt into a persona/scene description, then to turn a generated actor image into a video script plus audio/spoken-line instructions — i.e., LLM output is used as structured production instructions fed forward into generation nodes, not as a final answer.
«This could be a game-changer for content creation since it allows you to do everything inside one canvas instead of having to move around in the tool and mess around with all these hundreds of features in separate tabs.»
— 00:05
«So, for that kind of use case, using a node-based workflow builder like this is absolutely useless.»
— 03:51
«Because now this allows you to build repeatable content pipelines.»
— 06:01
«They can even clone this workflow for themselves, but they cannot edit anything.»
— 06:35
«This is LLM node which allows you to run any large language model such as Claude or Gemini or ChatGPT inside this one simple node.»
— 09:00
«Just finished the hardest session of my week and these held up perfectly. Flat sole, locked-in fit, zero slip. If you're serious about lifting, you need a pair.»
— 12:53
Reception
Mostly appreciative audience with genuine thanks and intent to use the tool, tempered by some tool-comparison shopping and skepticism about alternatives.
A hands-on, screen-recorded product walkthrough that is unusually candid about its own feature's narrow value proposition (repeatability and multi-step chaining, not single-output quality), and backs that claim with a concrete end-to-end example (the UGC shoe-ad workflow) rather than only abstract description.
13:20