07:44
AI ad creation
Higgsfield AI's Canvas lets you chain a product photo, an LLM node, an image generator, and a video generator into one repeatable node-based workflow so a single product picture can be turned into multiple ad images and short ad videos without switching between separate tools.
Canvas replaces jumping between separate image-generation and video-generation tabs by letting you lay out a repeatable workflow as connected nodes in one view.
Every node in the canvas has its own 'AI superpower': a prompt node, an LLM assistant node, an image generation node, and a video generation node each perform one step and pass their output to the next.
Workflow build order: drag in the product picture → write instructions in a prompt node telling the AI to analyze the picture and produce six short, detailed ad variation descriptions → feed the picture and prompt into an LLM assistant node (e.g. Claude Sonnet 4.6) and run it to get six descriptions.
One of the generated descriptions is copy-pasted into an image generation node along with the original product picture; the presenter recommends GPT image 2 as the model, sets aspect ratio (e.g. 4:3 to match the source photo) and quality/resolution, then generates.
The resulting ad image is connected into a video generation node with a short text instruction; the presenter recommends Seedance ('Seed ends') 2.0, sets duration (~6-7 seconds), aspect ratio (4:3), enables 'generate audio' (easy to forget and wastes credits/money if the output ends up muted), sets batch size to one, and lowers resolution to 480p to save time and money since final quality looks about the same.
You can also route the prompt into the video generator via a separate connected prompt node rather than typing instructions directly into the video node, purely to keep the canvas organized.
To scale up, connect the same picture/instructions into multiple image or video generator nodes, feed each its own AI-written description, and use 'run pipeline' (selecting all the nodes) to regenerate everything at once instead of one node at a time.
Prompting for these nodes doesn't need to be complex: the presenter frames it as 'talk to the AI as if you were talking to a real person,' and says giving the AI full creative freedom produces random results while more specific descriptions or additional reference images give more controlled results.
A link to clone the presenter's exact workflow is provided in the video description, and the same node logic (prompt → LLM → image generator → video generator) can be replicated in any other workflow-builder tool.
The presenter states the total cost of generating six pictures and six videos with this workflow was $5.75.
The workflow deliberately separates creative ideation from execution: one LLM node is used purely to brainstorm and write six ad-variation descriptions, which are then mechanically fed into separate generator nodes rather than the LLM doing any image/video work itself.
The presenter names a concrete tool friction point — Higgsfield Canvas doesn't auto-pipe text between nodes, so descriptions from the LLM node must be manually copy-pasted into each image node — indicating the workflow-builder isn't yet fully automated end-to-end.
Output randomness (e.g. the shoe floating over the Alps, or floating in the air in the video) is attributed directly to giving the AI 'full creative freedom' with a vague instruction, framing prompt specificity (or supplying multiple reference images) as the main lever for controlling composition rather than model choice.
The 'generate audio' toggle is flagged as a specific, easy-to-miss setting whose omission the presenter calls out as wasteful because it burns credits and money on a muted ad video.
The presenter treats resolution as a cost/speed lever with little perceived quality tradeoff, deliberately choosing 480p because 'at the end of the day, it will still look the exact same,' just at lower quality output.
The stated $5.75 total for six images plus six videos is offered as a concrete cost anchor for treating this as a low-cost, repeatable ad-creative pipeline rather than a one-off demo.
«So, every one of these nodes that you see over here has an AI superpower, basically.»
— 00:59
«something that is a bit annoying over here in Higgsfield AI Canvas is that you actually need to copy-paste these around.»
— 02:55
«I recommend doing GPT image 2. It is the best image generator model right now.»
— 03:23
«I think the Seed ends 2.0 is the best AI video generator available.»
— 05:09
«So, this is not rocket science. Literally, just talk to the AI as you were as if you were talking to like a real person.»
— 06:25
«the total cost of creating these six pictures and these six videos in total is $5.75 right now with this workflow.»
— 07:06
«you don't need to be a prompting genius. Just literally tell the AI what you want to see and do, and it will take care of the rest.»
— 07:24
Reception
Minimal engagement with a tangential self-promotional comment that shows neutral-to-slightly-positive reception without direct video feedback.
This is a promotional tool tutorial rather than a generalizable technique lesson: it hands over a concrete, copyable node recipe (product photo → LLM-written ad descriptions → image generator → video generator) with specific model and settings recommendations, but its usefulness is bound to Higgsfield AI's current Canvas feature, current model lineup, and the presenter's affiliate link.
07:44