Photoindra: posts tagged ai

Cleaning up AI slop. A manual retopo workflow in Houdini and Marmoset.

Thu, 21 May 2026 08:15:07 -0500

We’re heading toward an AI collapse, where everyone and their grandmother can generate videos, images, and complex 3D scenes. But for now, the quality of the low-resolution meshes AI generates is surprisingly poor. In the examples that major services show in their paid Instagram ads, the low-res geometry looks more like ZRemesher output from ZBrush. Same story with the UVs.

And while I’m sure this will keep getting better over time, for now the solution is manual retopology of the AI slop. With Modeler inside Houdini and auto-baking in Marmoset, the process isn’t that complicated.

The reference image was pretty small, so I generated 4K images from different view angles, plus some extra textures of acacia wood.

I generated the high-res AI-slop geometry for the main frog with the paid version of Tripo3D. You can use other free services, but in my tests Tripo produced cleaner results and allowed multi-angle references. Then I generated a separate mesh for the stick, just to try a multi-object workflow later on.

Imported both into Houdini, aligned the axes, and scaled to 1 meter. Retopo done with the Modeler plugin. I compared it directly against TopoGun 3 on several projects, and while I love TopoGun, I probably won’t use it anymore. Direct retopo in Houdini feels much more fluid, and on more complex AI slop it lets me hide parts of the geometry faster.

UVs done in RizomUV. I still think nothing comes even close to it.

Then I imported the UV’d low-res geo back into Houdini. Hardened the UV seams with the “Labs Soften Normals” node and merged the geo back with the high-res. I used groups as the name attribute so Marmoset would later understand it needs to bake from GRP_frog_stick_HIGH to GRP_frog_stick_LOW, and GRP_frog_body_HIGH to GRP_frog_body_LOW. This way I can have two AI-slop materials on the high-res geo with bad UVs, all baked down to one clean 2K UV set. When exporting the FBX from Houdini, just remember to tick the “Build Hierarchy from Path Attribute” checkbox and change “Path Attribute” to name. If you just assign the material inside Houdini, Marmoset won’t like it much – it’s easier to create a string attribute named shop_materialpath on the primitives and give it the name you want to see in Marmoset.

Also, if you assign different colors to different parts of the geometry in Houdini, Marmoset can use them as an ID map for baking. I used this to mask the stick later. You can preview it by switching albedo to vertex color in the material settings.

Then in Marmoset, with “auto-reload” and interactive mode in the Bake Project, it automatically picks up every change I make to the retopo in Houdini. I absolutely love this feature for iterating and making corrections – it’s probably the main reason I don’t bake in Substance Painter.

In Marmoset I created a New Texture Project and linked the low-res frog’s material to it, so I can fix all the strange AI texture artifacts by hand. In general, material editing looks very similar to Substance Painter. I think Substance is still missing a rounded-edge baker, but that’s a topic for another post. In a separate folder I exported the low-res geo for Sketchfab, and used the same folder to export the corrected textures from Marmoset.

This was the first time I rendered videos from Marmoset, and I was surprised by how fast it was. One real downside: you can’t link the render resolution to the camera, so if I want both vertical and horizontal shots, I have to change the resolution manually in the render tab.

You can grab the model for free on Sketchfab.

Will any of this matter in a year, when the meshes come out clean and the UVs aren’t garbage? Maybe not. But until then, manual retopo instead of AI is the way to go. The AI does the boring artsy stuff. I get to do the actual modeling. Yay!

Real directors vs AI Slop – Human Pacing Still Wins

Thu, 02 Apr 2026 09:56:03 -0500

There’s “this will change everything” coming from AI news every day. Tons of influencers claiming that “this will end Hollywood”. And sometimes generated videos even look ok on small phone screens. What’s always lacking is pacing, consistency, and a non-AI-generic look on 4K TV screens.

Higgsfield is an AI services aggregator. I don’t use it because of lots of negative reviews online. They sell you the idea of “unlimited” access but then people find out that it becomes “limited” very fast.
But their promo materials are quite fun to watch. Recently they came up with Arena Zero. A 10 minutes video generated with new Chinese model Seedance 2.0 (Yes, the one facing backlash in the USA for allegedly being trained on copyrighted Hollywood movies). The video was generated by 4 directors (from Kazakhstan and Russia) in 4 days. 2 for pre-production and 2 for color-grading, sound and editing. I’m pretty sure there were other designers involved helping generate scenes.

Important part left behind during breakdown:
How many credits did they use? They talk about “5000 decisions”.
Did they have some sort of “unlimited” access without waiting in line?
Is it limited to 720p only? They didn’t talk about upscaling during interviews.
Also from what I see access to Seedance 2.0 right now is only for “business” accounts and “quick business verification required for access” whatever that means. I guess it is to make sure that the company is registered outside of USA to avoid possible future copyright claims.

But it’s the first time I’ve watched 10 minutes of so-called AI slop and didn’t feel bored.

Links to original videos:
Arena Zero from Higgsfield AI

Arena Zero Full Breakdown: Seedance 2.0

If you want to compare and look at influencer generic AI-slop:

Pony Halloween in Star5

Wed, 23 Oct 2024 17:53:53 -0500

While everyone is arguing what AI is best to keep people lying on the grass (Stable 3.5 or Flux), I’m using Pony Diffusion to grab some souls for this Halloween and form a “Cult of Ai”.

Was made with AI in ComfyUI on Windows. But you can build similar graph in Colab. Just need to customize the input and output folder paths.