How are you actually handling AI-generated mesh cleanup before engine handoff?
Been integrating Meshy and Tripo outputs into UE5 pipelines for a while now and the generation step is no longer the problem — topology and textures come out reasonably usable for props and background assets. The problem is everything between export and the asset actually being in the project correctly.
My current manual steps after generation: health check for non-manifolds, remesh if tri density is too chaotic, LOD generation, rename to convention, export to correct format. On a batch of 30 props that's a half-day of babysitting that shouldn't require a human.
Curious what others are doing. Specifically:
- Do you have any of this scripted or is it still manual per-asset?
- Where does the process break down most often for you?
- For character assets that need to deform — are you still doing full manual retopo or has anything automated that acceptably?
Not looking for tool recommendations necessarily, just want to understand if others are hitting the same wall or if I'm missing something obvious in my setup.
Replies
When using some "result" from some AI, which in fact is some masterpiece of software developement but that's another story .. then why does one think that "some artists" are able to fix the garbage by "some script" or do have to do some manual repairs.
Isn't the problem that some AI just do not know about the whole context of specific workflow which is used in the given use case ? This is even not easy for humans. One can know a lot about any part of 3D but when working in a team then there are often some unwritten rules one has overcome.
Sometime some saying are twisted badly.. it's okay to say: "Make it exist first, you can polish it later." But there is also "concentrate on the right context" (for example "correct" proportions).. but when exchanging "polish" with "better" then one already implies that it can be bad at first. And when the proportions are "shitty" then the whole figure will be..
¯\_(ツ)_/¯
I think AI does have a place in tooling, but in my experience it is currently limited to assisting with filtering large datasets, and it requires a human with subject-matter expertise to verify the output.
The hidden costs are also still all compelling reasons to push back against this tech.
maybe I am just old fashioned, but where is the fun in typing in stuff to produce art?
It's a dimension-translating IQ test. Who doesn't like IQ tests? ;D
When you have to shove 10 or 15 assets onto the same texture page, figure out how to share space and manage the LOD UVs it starts to be a lot of work.
I've seen UV sets for prop sets take talented people a couple of weeks to put together
On left its AI model with 500k tris, horrible topology and textures, Right is 130k tris, optimized topology, proper UVs, additional details and proper textures and LODs if needed for foreground and background usage. Modelling it from blockout might take two to three days with feedback plus full control on topology. Generated via image to 3d AI might save you a day the most but you are now cleaning it up like photogrammetry so I dont see any benefit from production point of view.
Even if I ignore all the added detail of your effort, and iftrue:is4realtimerendering;
It's effort is 4x as memorically (oh yeah) costly as yours, costs more to render, meaning the end-user has to buy more memory.
Meanwhile, your effort has a whoooole lot of shading definition and detail retained that no AI will ever be able to maintain or replicate.
A dev replacing artists with AI, or forcing their artists to use AI is passing the rendering expense onto the end-user, is like saying "our marginal development savings are more important than the gamer's costs. Les MaO: Fuck you."
Yours is better in every single way. I'd argue that the AI effort isn't even worth baking down to a normal map.
You: Ooh, look at the nice shapes and flows in different axises. ououoouuuh.
AI: GHRRHRHEPRBBLBRRRRBHBBLBLLBLRRRBL, BUY $1000 "worth" of more of memory BRRBYRLRLL. THANK YOU FOR YOUR $20 MONTHLY SUBSCRIPTION GRRRAAARRRRRRBLLLBRRRRRR!
Like I wont lie, it does have that charm of bypassing blockout and going through all that edge flow establishing the base mesh and getting to cleanup stage and the result. However this I see only working for non animatable/deformable assets like this statue, rocks or random scifi bits and pieces.
Anywhere you need animation, proper rigging, deformation, vfx or complex object built of many parts, AI fails horribly.
I even asked one of the Image to 3D AI company and they assured me that their 3d models are not built on stolen geometry data but on paid library...
Either way AI is here, using it wisely is way forward.
Here is another example I did and no way I am cleaning this, at most if client insist I will be using this as proxy and remodel from box/plane as it will give me more control + optimized geo given its for playing, animating. If they want as non animatable asset like a toy on table then yeah its quick retopo/fix and call it a day.
I made this post cos I wanted to know what you guys think about workflows such as the one I mentioned, now that AI is in everything.