10:22
AI video generation
Testing Higgsfield AI's newly launched "Supercomputer" agent platform by having it autonomously build a UGC-style and a professional video ad for a DJI Avata 360 drone from a single prompt, the reviewer finds the fully autonomous output unusable (wrong setting, wrong gender, misspelled product name, inconsistent voice/audio) and worse and more expensive than a manually-built Higgsfield Canvas workflow, concluding the tool needs active human direction rather than a single hands-off prompt.
Higgsfield AI launched an AI agent platform called Supercomputer, and the video tests whether it lives up to the hype
First test: generate a UGC-style ad for the DJI Avata 360 drone from one prompt, a product link, and a product image, using Claude Opus 4.6 as the model
The agent loads a "UGC flow skill" to follow a production pipeline for talking-head ads and reads the product web page, finding a workaround when the page fails to load directly
The agent asks for the desired video duration since none was specified; a 30-second ad is requested
Because the underlying video generator (Seedance 2.0) is capped at 15 seconds, the agent splits the ad into two 15-second clips with separate storyboards (hook/setup, then apply/closer hook)
For a drone product, the agent proposes a location of "home desk or living room," which the reviewer immediately flags as a mismatch
The generated character prompt describes a woman in her early 30s, even though the actual product page's reference images show men using the product
The agent repeatedly pauses to ask for approval before running each expensive step: image generation, script writing, more images, and final video generation
The generated images look like polished, cleaned-up home advertisement photography rather than an outdoor/action UGC setting appropriate for a drone
The reviewer hesitates over spending 270 credits before finally proceeding to see the finished result
The AI automatically stitches the two 15-second clips together without requiring manual video editing
The finished UGC ad is judged totally unusable: it's filmed at someone's home, the actor is "screaming and talking about the product" without demonstrating it in use, unlike a real UGC ad (e.g., a student in a soccer field)
A concrete defect: the narrator's voice changes between clip one and clip two because the video generator has no memory of the prior clip or the conversation
As a comparison, a Higgsfield Canvas workflow built by the reviewer about a week earlier — starting from a product photo, analyzing the product website, and generating a video — produced a "flawless" result in a few minutes
The Supercomputer run took longer and cost more credits than the Canvas workflow while producing a worse result
Second test: a non-UGC, "professional" marketing ad for the same drone from a similar single prompt
That second result is judged even worse: it misspells the product name ("DJI Avita 360"), has inconsistent/robotic audio, out-of-place ambient sound, and a background-image glitch
Overall first impression: Supercomputer is unimpressive because it bundles image generators, video generators, connections, and memory into one chatbot that can only handle very simple tasks well
The reviewer says it isn't fair to fully judge the platform from a couple of bad autonomous examples and plans a follow-up where he actively directs the AI ("I am the boss, I am the captain of the ship") and shares use cases where it performs well
The main failure mode isn't per-clip generation quality but cross-clip continuity: because each 15-second video generation call has no memory of prior clips, a longer ad assembled from multiple stitched segments breaks voice/audio consistency — a structural limitation of chaining independent generator calls rather than a prompting problem
The agent's default creative choices (setting a drone ad in a clean home rather than outdoors, defaulting to a woman narrator when the product page's own reference images show men) suggest its UGC heuristics aren't actually grounded in the product's real usage context, despite having crawled the product page
A fully autonomous single-prompt-to-finished-product agent run appears to cost more credits (270+) and more time than manually assembling a comparable result with a controllable workflow builder (Canvas), while producing a worse outcome — suggesting the orchestration overhead adds cost without adding quality
The reviewer frames the tool's ceiling as dependent on active human steering mid-generation rather than reliable one-shot autonomy, positioning Supercomputer as a raw agent loop that needs a human "captain" rather than a self-sufficient pipeline
«Hicksfield AI just dropped their AI agent platform called Supercomputer.»
— 00:00
«I think there's so much information about drones online and so much footage and video that the AI should actually nail this pretty well.»
— 00:38
«Wow, this actually allows us to do longer videos.»
— 01:50
«Wait a freaking second. No, this is a drone, so at least this doesn't sound too good to me, but let's see what it comes up with.»
— 02:43
«So, a spontaneous iPhone snap of a woman in her early 30s.»
— 02:58
«And I am hesitating to spend 270 credits on this right now.»
— 04:59
«Well, this ad is totally unusable because this is filmed at somebody's home, and it does not look like a UGC advertisement.»
— 06:03
«So, right over here, when the clip changes from the clip number one to the clip number two, it actually changes the voice of this person.»
— 06:34
«Well, well, well, that is even worse than the previous advertisement.»
— 08:48
«So my very first impression of the supercomputer feature is that I am not really impressed at all. Actually, I think it's quite bad.»
— 09:12
«I would need to use this in such a way that I am the boss, I am the captain of the ship, and I will give it the direction when it loses the plot.»
— 09:55
Reception
Audience appreciates the honest critical review, with most viewers agreeing the tool is overhyped but promising, while some defend its current capabilities.
The video is a hands-on first test showing that Higgsfield Supercomputer's single-prompt-to-finished-ad promise breaks down in practice — mismatched settings, wrong casting, misspelled text, and cross-clip voice inconsistency — costing more credits than a manually built Canvas workflow for a worse result, with the reviewer withholding final judgment pending a follow-up where he actively directs the agent.
10:22