20:21
AI video generation
The video argues that Higgsfield AI's advertised pricing wildly understates the real cost of making usable videos, and that most bad AI outputs are never a skill issue but a hard capability ceiling in the model itself — the practical path forward is to hunt for the narrow slice of ideas AI already handles well, use only a handful of core features, and apply two concrete tricks (cheap test renders, image-guided generation) to cut wasted credits.
The Higgsfield pricing page ($47 Plus / $99 Ultra) drastically understates real-world cost — actual spend can run "hundreds of times" higher than the listed price.
Worked math: a 15-second Seedance 2.0 clip costs 135 credits; naively scaled, an hour of raw generation is 1,350 credits, but getting genuinely usable footage (5-10 attempts per clip) can push an hour of real footage to $10,000 or more.
The single biggest way to save both credits and time is to ignore most of Higgsfield's features — many are outdated, redundant, or duplicate what Create Image / Create Video already do with one prompt.
The creator's actual toolkit is just three features: Create Image, Create Video, and AI Canvas (a node-based workflow builder linking image, video, and LLM models).
Of 20+ available video models, only Seedance 2.0 matters in practice — used about 99% of the time; the rest are described as "pretty much absolute garbage" by comparison.
Core thesis: bad AI output is "never a skill issue" — the AI itself is limited, hallucination-prone, and randomized, so most possible video topics/ideas are simply outside what it can currently do well.
Strategy: instead of trying to force any topic or passion into AI, hunt for the narrow spots on the "map" of possible video ideas where AI already produces good results.
Examples of AI's current sweet spot: a reversed earth-zoom-in effect, product/shoe rotation ads, and UGC-style ad clips — all simple, short, and well represented in training data.
AI currently cannot produce a consistent clip longer than roughly 10-15 seconds.
Detailed case study: a prompted "jump into water and catch a fish" clip looks impressive at first glance but breaks down on scrutiny (face changes within 5 seconds, splash looks painterly, fish behaves unnaturally), and a second take shows the same category of flaws — illustrating a capability ceiling that won't move until a new Seedance release.
Concrete credit-saving trick #1: generate a cheap tester video first — drop to 480p, use the lightweight/"fast" model variant, and shorten duration (e.g. 5-7s instead of 15s) — cited as 12 credits vs. 330 at max settings, almost 30x cheaper, before committing to a full-quality render.
Concrete credit-saving trick #2: guide generation with a real starting image (and optionally an end image) rather than generating from scratch, for a more controlled result and to preview the first frame before spending credits.
Despite the critique, the creator still recommends Higgsfield because it aggregates access to best-in-class models as they release, and AI Canvas lets multi-step workflows (Claude for product analysis/script, GPT Image 2 for the image, Seedance 2.0 for the video) run and be shared from one place instead of three separate subscriptions/tabs.
The video ends with an affiliate disclosure/recommendation to use the creator's signup link.
The real cost of a truly usable hour of AI video can reach $10,000+ once the 5-10 regeneration attempts needed per clip are factored in — an order of magnitude beyond naive credit math and far beyond anything the pricing page communicates.
The paid tiers ($47 and $99/month) translate to roughly one to two minutes of actual finished video output, not the impression of broad video-production capacity the marketing page implies.
Reframing "skill issue" as a category error: since the model's raw capability sets a hard ceiling, no combination of prompting, model choice, or time invested moves the needle — demonstrated by two separate generations of the same prompt producing the same structural distortions.
Virality in AI clips (the reversed earth zoom-in trend) is attributed to chance alignment between what's easy for the model and what happens to look good to an audience, not to clever prompting.
Testing at 480p / fast model / short duration before committing to a max-settings render is framed as almost 30x cheaper for materially the same signal about whether an idea works at all.
Feature-sprawl in the Higgsfield UI is presented as mostly redundant: many named features reduce to prompts already achievable via the base Create Image / Create Video tools, meaning the product surface overstates its own functional breadth.
The creator's own workaround for staying inside one platform is the AI Canvas node-based workflow, explicitly pitched as replacing three separate subscriptions with one shareable, cloneable workflow.
«So the plus plan will only give you credits for a very very tiny amount of videos.»
— 01:21
«If 15 seconds costs us 135 credits, 1 hour worth of videos will be 1,350.»
— 02:23
«if that's the case, one hour where the footage is $10,000 or even more.»
— 02:52
«please don't look at the pricing page here at all. This is not what it costs to make videos using Higsfield AI.»
— 03:13
«it is never a skill issue»
— 06:40
«the amount of times that the AI actually produces something usable is very very low»
— 08:25
«It's this easy when you know the right spot.»
— 11:52
«this result will not improve from this until we get the next seedance 2.0 basically»
— 14:39
«AI video generation is ridiculously easy. You literally just tell the AI what you want to see and it takes care of the rest. But the models right now are just not that good.»
— 15:00
«this is quite literally almost 30 times cheaper»
— 16:48
Reception
Strongly appreciative of the video's honest discussion of AI video generation costs and limitations, but with constructive pushback that better optimization techniques could yield better results.
The video functions as two things at once: a pricing-transparency exposé showing the sticker price is a small fraction of real spend, and a broader capability argument that bad results reflect the model's ceiling rather than the user's skill, landing on two concrete, low-effort cost-saving habits rather than a deep technical framework.
20:21