CPU3DAI tools for 3D, video and audio
3D models Video Sound & music
Which AI generators will actually run on your GPU

We install and run the generators ourselves — from 3D models and video to sound effects. We write down how much VRAM each needs, what runs on CUDA and what works on Apple MPS and AMD ROCm, which is open source and which is a cloud subscription.

3D generator on 8 GB text-to-video without CUDA SFX from a description

We read visitors’ search queries and write about them — ask right here.

0articles
0sections
0founded
prompt → resultdrag the slider ↔
A photo of a chair — the generator’s input
The same chair as a 3D mesh — an illustration of the result
$ img2mesh chair.jpg --pbr Hunyuan3D 2.1 · from 10 GB
“a clay teapot
with a faceted lid,
low-poly”
A 3D model of a teapot — an illustration of the result
TripoSR · from 6 GB
A still frame with a paper boat — the input
The same boat in motion — an illustration of the result
$ i2v boat.jpg --frames 120 LTX-Video · from 1 GB
“rain on a tin roof,
distant thunder”
MOSS-SoundEffect · from 8 GB
← input result →

The illustrations are AI-generated — examples of what the catalogued tools do, not their live output. Memory thresholds come from the fact sheets.

What runs locally

The whole tool database →

Pick a type, a chip — CUDA, Apple MPS or AMD ROCm — and a VRAM ceiling. What remains are the models that actually start. Requirements are as described by the authors, with a source and check date for every value.

3D Video Audio Open source Cloud AMD ROCm Apple MPS Up to 8 GB of VRAM
ToolTypeRunsVRAMBackendsFormats
LTX-Videoimage-to-videoVideoopen sourcefrom 1 GB *Apple MPS—
ChatTTStext-to-audioAudioopen sourcefrom 4 GB *——
Zero123++3Dopen sourcefrom 5 GB *——
Stable Fast 3Dimage-to-3d3Dopen sourcefrom 6 GB *CUDAApple MPSglb
threestudio3Dopen sourcefrom 6 GB *CUDAobj
TripoSRimage-to-3d3Dopen sourcefrom 6 GB *CUDA—
MMAudioAudioopen sourcefrom 6 GB *CUDA—
FramePackVideoopen sourcefrom 6 GB *CUDA—

Showing the most accessible by VRAM threshold. Cloud services need no VRAM — they need a subscription: they are in their own list.

AI news — every day

The whole feed →
2026-09-14 llama.cpp v0.4.1: support for Maple 20B-A1B, Tencent Hy 4 and Spark2.5 The llama.cpp v0.4.1 release came out on September 14, 2026. The update adds support for three new architectures: Maple 20B-A1B — a ternary MoE for CPU, Tencent…
2026-09-14 Cheap Chinese Coding Models: What to Choose in September 2026 2026-09-14 BubbleID: Image Segmentation Model for Boiling Analysis Released on HuggingFace 2026-09-11 Agnes-AI releases open preview model Agnes-3.0-Flash with 33 billion parameters 2026-09-11 whisper.cpp v1.9.4: Windows on ARM support and ggml update to 0.23.0

Fresh arXiv papers and model releases — briefly, with a note on what you can already run yourself.

on click · nothing loads in advance

From points to a model

This is how every 3D generator works: first a point cloud, then a surface, then materials. Click and we show it step by step. The page stays light when opened: the frames load only when you ask.

01Prompt or photoOne image or one line of text — the input of almost any generator.
02Point cloudThe model predicts volume: voxels, points, a latent grid.
03Mesh and texturesPolygons, UVs, PBR maps. This is where things break most often.
04Engine or printerWhat you can actually hand to Blender, Unreal or a 3D printer.

The frames load only now, not when the page opens. They are AI-generated illustrations — a real viewer will appear when our “3D from a picture” tool goes live.

Fresh generator reviews

LatestMost read

Site map

AI 3D generatorsupdated Video Generatorsupdated Audio generatorsnew AI Newsdaily Tool picker by hardware