AI 3D generation
Blender Guru replicates three viral AI-3D demos (AI driving Blender, diffusion-based mesh generation ads, and 3D-guided AI video) and grades each as true, plausible or false. He finds the tools real but the demos and ads overstated, and he predicts the future will reward 'chefs' (people with vision who direct tools) over 'cooks' (people paid to do one task in one tool).
The video's frame is that viral clips claim artists will soon be out of work. The author, a 22-year Blender user, teacher and 3D-asset company founder, tests this by redoing three viral demos himself. He grades each as true, plausible or false and promises to answer the job question at the end.
AI driving Blender is grade 'plausible'. ChatGPT (Astra) sends and receives Python scripts to control and evaluate the scene, and the video calls the result 'legitimately insane'. But his scene came out far simpler than the OpenAI demo ('a prison cell'), and the same held for the Palace of Fine Arts animation. A six-hour, 20x-cost multi-agent attempt still could not match the demo's interior arrangements. He suspects the demo was micromanaged room by room, contrary to how the article presented it.
The video states that AI assistance in Blender suits specific, bounded tasks, such as checking normals or writing a short geometry-nodes setup. It is weaker at larger tasks that require judgment, like furnishing a house.
The mesh-generation adverts from Tripo and Meshy are graded 'false'. The video argues the technology is real and the ads are misleading. Tripo's 8K-texture ad image is AI-generated and could not be reproduced, and Meshy's 3D-print ad used a professionally made marketplace model rather than a generated one. Both companies conceded and said they will tighten ad review, and Meshy pulled four adverts.
He separates the grade for an advert's truthfulness from the grade for the underlying technology. Mesh generation, he says, is not false: it already appears in films and TV as a low-cost substitute for photogrammetry, where blur and imperfect meshes don't matter at a distance.
3D-guided AI video (Blender blocks driving a video model) is graded 'true'. He reproduced the result, with caveats: it took many generations, a far more finely tuned prompt, hand-rewriting of the fight choreography, and he doubts Claude produced the tutorial's camera animation. He finds that it removes technical hurdles and leaves you directing, and says it is already useful for short pieces and appears in commercials, though consistency across a long film with many characters would be frustrating.
On whether AI will replace 3D artists, he points to precedent. Cycles automated the manual placement of bounce-light lamps, yet there was no mass displacement, for two reasons: time saved gets refilled with other work (Parkinson's law), and falling barriers bring in new entrants and new industries such as Boeing, IKEA and fashion prototyping.
He argues that AI differs from Cycles in that there is no end to the skills it can automate and it moves at a speed that makes the last 20 years of Blender look slow. His prediction is that the future rewards the 'chef' (idea, story, product, vision, leading a team of agents or humans, iterating faster) over the 'cook' (paid to do one thing with a preferred tool). He says he is not endorsing this worldview.
He ends ambivalent and personal. He worries that some artists will quit, which he calls a huge loss for the industry and humanity. He also says he got into 3D to make a short film, not to do vector math and geometry nodes all day, and AI might finally let him make one alone, though he hasn't thought about filmmaking since he was 20.
The gap between a viral demo and an honest replication is often hidden human effort. The video points to micromanagement room by room (OpenAI house), a report-first-then-implement prompt (Palace of Fine Arts), and per-face colored blocks plus hand-tuned prompts (Higgsfield). The published prompt alone did not reproduce the result.
The video states that texture resolution isn't everything. Sharp results also need good UVs: a good game asset has 70-85% UV coverage, while the Tripo model had about 60% and no stacking of identical left and right sides. He puts a grid texture at 8K on it and finds the grid sharper than the delivered texture. The model delivered the texel density of about a 1K texture at the performance cost of an 8K one.
Tripo suggested isolating generation to the head to raise resolution. The author notes this is true, but a competitor could then claim even sharper textures by cutting the model into even more pieces. This shows that headline resolution numbers can be gamed.
The ChatGPT-generated Blender scene had systematic inefficiencies: level-three bevels on nearly every object, even hidden glass edges, inverted normals causing shading artifacts, and inconsistent instancing that slowed rendering. He says these can be fixed with a follow-up prompt if you know to ask for it, and that the speed of AI development may remove even that need by year-end.
In the Meshy test, the 'generated' model was actually a professional print file called Cyberbee, sold on 3D marketplaces. It prints in a third of the time, needs no supports, and has moving parts. His own generated version needed seven re-splits to get connectors on every part, then failed in PrusaSlicer, with messy prints and supports that could not be removed. Meshy had two other adverts built on existing marketplace products.
Higgsfield's fix for character direction was to color each face of the guide block so the model knows which way the character faces as the camera spins. The prompt also had to state the exact number of ninjas and that bodies appear only behind the protagonist. Even then the model could not handle more than two ninjas at a time.
Using Nano Banana to make a front-facing T-pose of the character improved texture placement in Tripo, and Tripo's auto-rig feature was 'surprisingly decent'. Even so, the blurry result stayed.
In his framing, the labour-market effect of automation is an expansion story. Faster rendering made new uses practical (cabin visualization, furniture visualization, clothing prototyping), and those jobs would not exist if the technology weren't fast enough.
Both Tripo and Meshy responded constructively when he reached out: Tripo conceded it would strengthen ad review, and Meshy thanked him and pulled the adverts. The Higgsfield tutorial author also connected him with a team member who supplied the full prompt and tips.
«Their office looked cozy and lived in, but mine looked like a prison cell.»
— 02:29
«I suspect that they went room by room and micromanaged every detail, which is not how they made it seem in their article.»
— 03:59
«Very good at specific tasks, less good at larger tasks that require judgment, like the house.»
— 04:31
«you would get the texel density of maybe a 1K texture whilst paying for the performance of an 8K one.»
— 07:09
«And that's because of course the whole video is a lie.»
— 09:33
«I repeat, mesh generation is not false.»
— 10:21
«I'm giving it the grade of false because the ads are fake.»
— 10:41
«It's not fully 3D, it's not fully AI, it's 3D AI.»
— 10:53
«I was surprised at how fun this was because there's basically no technical hurdles. You're just directing the entire time.»
— 13:29
«First, when a process gets faster, we fill this with something else. This is known as Parkinson's law.»
— 15:19
«I'd say that the future will likely reward the chefs, not the cooks.»
— 16:18
«The cook is the person that likes to be paid to do one thing using the tool that they prefer.»
— 16:44
«I didn't get into 3D so that I could do vector math and geometry nodes all day.»
— 17:14
Reception
Mixed but mildly positive: a few viewers thanked the creator and agreed with the debunking, one called it slop, and most of the discussion is anxious or divided debate over AI replacing 3D artists, with the 'AI is a tool' view common.
A replication-based audit that separates misleading marketing from real capability. It grades the Blender demo plausible, the Tripo and Meshy ads false and the AI-rendering tutorial true, then frames AI as another round of 3D automation that favors vision-driven 'chefs'. The author states that he is uncertain and not endorsing that outcome.

17:45