Preetish Kakkar

Rendering/Graphics engineer at Adobe. Opinions are my own.

[Preetish Kakkar @preeteesh]

@paraschopra @lokidotdev Tell us more about these weird discord channels?

[Richard Geldreich 🇺🇸 @richgel999]

I've begun writing real-time texture encoders using Opus/Fable. ASTC HDR 4x4 was first, now I've got full ETC2. It's faster than etcpak's SIMD encoder, without using any SIMD.
Quality is going up because previously UASTC LDR 4x4->ETC2 transcoding wasn't very strong as it was only exploiting ETC1 modes.
For GPU texture encoding using AI is utterly revolutionary. In hours I can do potentially weeks of experiments and implementation work.

[Preetish Kakkar @preeteesh]

I'm also incredibly grateful to our reviewers, Shannon Woods, Kai Zhang (NVIDIA), and Al Pavlov, for their thoughtful feedback and attention to detail. Their contributions helped us make the book the best it could be.

[Preetish Kakkar @preeteesh]

Excited to announce that my new book, Neural Rendering in Action, is coming next month!

It's the first book dedicated entirely to exploring how neural techniques can be applied across different rendering workflows.

Preorder here -https://t.co/7kJNb7DgZ8

[Preetish Kakkar @preeteesh]

A huge thank you to my co-authors, Mauricio Maurer, Sergey Kosarevsky, and Prashant Shukla, for their knowledge, dedication, and contributions. This book wouldn't have been possible without them!

[Preetish Kakkar @preeteesh]

We also have a dedicated chapter on Diffusion Models, written entirely by Prashant Shukla, covering diffusion-based techniques and working with ControlNet.

[Preetish Kakkar @preeteesh]

The book takes a hands-on approach to implementing and understanding a range of techniques, including Slang, differentiable rendering, texture compression with MLPs, NeRFs, and Gaussian Splatting.

[Jiayin Cao @Jiayin_Cao]

New blog post: Building a Fiber-Based Job System from Scratch.

I walk through how I built the fiber-based job system in my offline renderer, including scheduling, synchronization, and the design choices behind it.

https://t.co/2h5muwLEnx

[Lummox @Lummox_eth]

Claude Code vs Blender is getting f*cking illegal

Two workflows building the same complex 3D rocket assembly, but one costs nearly 19x more in the demo

And the difference reveals something most AI automation demos completely ignore

Claude Code can reason through geometry, generate Python scripts, assemble mechanical components, and automate complex animations

But every unnecessary tool call, failed generation, and repeated rebuild adds overhead that can make the final result ridiculously expensive

Meanwhile, Blender can execute predefined procedural operations without paying for another round of AI reasoning

Here’s where the workflows differ:

Traditional Blender:

MODEL individual rocket components
POSITION mechanical structures and hinges
CONFIGURE materials, lighting, and cameras
ANIMATE deployable panels
REFINE and reuse finished assets

Claude Code + Blender Python:

GENERATE geometry from written instructions
ASSEMBLE mechanical components through scripts
PARAMETERIZE dimensions and deployment angles
AUTOMATE animation and camera movement
REBUILD variations by changing code

The interesting part isn’t which tool wins

It’s knowing when AI reasoning is worth the cost and when a deterministic script can do the job

Want to recreate the experiment?

Open Claude Code with Blender installed and paste:

“Create a procedural Blender Python generator for a detailed rocket assembly with deployable panels, animated hinges, and realistic metallic materials

Separate geometry, materials, animation, and camera setup into reusable functions

Expose all major design parameters, avoid rebuilding unchanged objects, and log execution time and available AI usage costs”

Blender handles the 3D execution

Claude Code makes the workflow programmable

The real advantage isn’t using more AI

It’s getting the same result without paying for unnecessary intelligence

Bookmark this workflow before your next 3D project

[Preetish Kakkar @preeteesh]

@richgel999 Any tips on what should I tell codex/claude, I had it review my past 20 conversation but not much useful it suggested that by default codex wasn't already doing

[Preetish Kakkar @preeteesh]

@Dispatch_Graph @milostosic I suggesting looking at this page to understand bgfx in more details, there is a reason bgfx is so popular even though its shader compiler still can't compile ray tracing shaders

https://t.co/DLEmuWzleB

instead of writing new, why not fork bgfx and make it modern?

[Preetish Kakkar @preeteesh]

A candid note: AI wrote much of this. I reviewed it at a high level and tested it on real GPUs, but not line by line.

The compiler is not open source yet, though I plan to release it.

What would you want to compile, and what would break?

[Preetish Kakkar @preeteesh]

I used AI to help create msl2spirv, a Metal Shading Language → Vulkan SPIR-V compiler.

Basalt is a Vulkan renderer where every shader—including PBR, shadows, reflections, bloom, tone mapping, and ImGui—is written in Metal.
https://t.co/hW39JWSqET

[Preetish Kakkar @preeteesh]

The compiler parses Metal with a Clang fork, lowers it to a typed shader IR, and emits SPIR-V validated for Vulkan 1.3.

It supports the shaders used by Basalt: PBR, shadows, sky, reflections, bloom, tone mapping, and ImGui.

[Preetish Kakkar @preeteesh]

@maanalaolaqy How did you train the network?

[Preetish Kakkar @preeteesh]

@richgel999 Any reasons on why it doesn't helps? Also what is smallest size that can still give decent PSNR?

[Preetish Kakkar @preeteesh]

@Kiiiri7 Waiting eagerly!

[Preetish Kakkar @preeteesh]

@richgel999 Would this require a differentiable BC/ASTC encoder?

[Preetish Kakkar @preeteesh]

@pissang1 Does WebGPU already support full ray pipelines, or only ray queries?