Anime and Games Made With AI, Some of It on a Local GPU — An Anime VAE, Picture-to-3D With Idol, and a 24-Hour Game

Decode the same video with a different VAE and the finish of the paint changes. @yu_ichi_suzuki tried "Minimax H3 Seamless Anime VAE," a VAE specialized for anime-style coloring ¹. The creator writes (in Japanese) that a VAE is the part that decodes the generated latent into a picture ¹. A 1088p video made with MiniMax H3 (run locally, the creator notes) was decoded with 4 VAEs, including the official VAE and Sol-VAE, and compared ².

The result: unevenness in flat-colored areas is clearly reduced ³. Line sharpness stays the same, and jagged edges and anti-aliasing on the line art get cleaner, the creator writes ⁴. But the finish turns uniform, like cel animation, so it may not be the right choice every time; if you want to keep texture, it may not be the best answer ⁵. The creator's read is that it should suit people who want even cel-style fills and clean lines ⁶.

Decode times are in the post too. The conditions are given as RTX 5090, 1088p, and "about 10 seconds"; official took 50 seconds, Sol 7 seconds, Seamless fp16 49 seconds, and Seamless int8 24 seconds ⁷. Of fp16 and int8, the creator says int8 is faster and looks no different to the eye, so int8 seems fine ⁸.

Same spot, swapped for speed or for style

Quoting that post, Phantom (@Phantom_202601) writes (in Japanese): "Just swap the VAE and H3's pictures become anime-style" ⁹. Phantom explains that this VAE replaces only the last step, the one that turns the generated latent into a picture ¹⁰.

The interesting part comes next. Phantom writes that they used LightVAE on that same decode step themselves ¹¹, then calls this the idea of swapping the same spot for style rather than for speed ¹². One decode step, swapped by one person for speed and by another for look. That pairing is the fun of it.

Now the primary sources. According to its download page, Seamless Anime VAE is fine-tuned from MiniMax H3's official fp16 VAE ¹³. One of its aims is better reconstruction across a wide range of 2D illustration and animation styles ¹⁴. It removes blurring and smudging along sharp edges ¹⁵ and reduces subtle color jitter in flat-shaded areas ¹⁶. The license is the MiniMax H3 community license ¹⁷.

LightVAE shrinks a video VAE while preserving its original latent interfaces ¹⁸. There is an independently block-pruned Lite encoder and decoder for MiniMax-H3, with native ComfyUI integration ^{19 20}. It is Apache-2.0 ²¹.

MiniMax H3 itself has open weights, and the ComfyUI docs say that lets you run it locally ²². Using locally generated output commercially requires a MiniMax commercial license ²³.

If you draw your character sheets and storyboards by hand before sending them to video, the pen belongs in the step before all this.

From a picture to 3D

After video, 3D. Kemono-chan (@asdfg21_snivy) reports, with a video attached, trying "Idol," a 3D model generator that runs locally ²⁴. It works even with 8GB of VRAM, and generation is fast at 45 seconds ²⁵. The creator writes (in Japanese) that it is pretty good for a free 3D generator ²⁶.

We covered Idol in last time's ComfyUI roundup too. @NobinLOG writes that it is now usable in ComfyUI, and adds cautions: it is a Gaussian splat rather than a polygon mesh, and the back side is an AI estimate ^{27 28 29}.

Idol's developers say it runs on consumer GPUs, using 7.2GB of VRAM ³⁰. That describes usage, not a minimum requirement ³⁰. From a single image it builds a 3D Gaussian splat and writes it to a standard .ply ³¹. It is Apache-2.0 ³², and a ComfyUI distribution exists ³³.

A full original game in 24 hours

The next step in that flow is a post from Shikhar (@xikhar). The creator writes that they made a full original game in 24 hours using AI ³⁴.

The stack: Opus 5.5 + Flux 3 Max + GPT Image + Tripo + Blender + Three.js ³⁵. "Anyone can make games now" is the creator's own line, and the post says the full workflow and prompts are in the thread ^{36 37}.

This stack is not all run locally. Opus 5.5 is offered through the Claude Platform and other cloud services ³⁸. Tripo needs no install and runs entirely in the browser ³⁹. Blender, on the other hand, runs locally ⁴⁰, and three.js is MIT-licensed ⁴¹. At least part of it is cloud services, and part of it is tools that run locally.

Still, among the jobs that come up in making a game, turning a picture into 3D and writing code are jobs where people have reported running things on their own PCs: Idol above for the first, and the PC Watch example below for the second.

Running it on your own PC

PC Watch's test of the ASUS ProArt P14 involves the same Three.js as Shikhar's stack ⁴². PC Watch writes that the open-weight model Qwen3.8-Flash-Next installed with no fuss and handled difficult coding ⁴³. Given a single line asking for a 3D CG hanging garden in Three.js, it wrote code at more than 20 tokens per second and produced 3D graphics after about 15 minutes, PC Watch says ⁴². That is PC Watch's result on that machine and configuration.

For the ASUS ProArt P16, PC Watch writes that Blender GPU rendering with OptiX holds its own against a mid-range discrete GPU in public benchmarks, as a rough comparison since the versions differ ^{44 45}.

According to NVIDIA, RTX Spark handles 3D scenes up to 90GB in up to 128GB of unified memory, and OptiX and DLSS keep the ray-traced viewport moving ⁴⁶. It also says you can generate images and video on the device with no subscription and no per-prompt fees ⁴⁷. And there are plans: NVIDIA announced that Blender is adding DLSS 4.5 Ray Reconstruction and that RTX Video Frame Generation is coming to ComfyUI, with both arriving this fall with RTX Spark ^{48 49}.

On mixing cloud and local, PC Watch's report from the Japan launch event says NVIDIA explained that you can run AI agents locally and combine them with cloud frontier models as needed ⁵⁰.

If you want to run this week's tools on your own desk, check each model's or node's download page first for the environments it supports. The VAE comparison above was done on an RTX 5090 ⁷.

To look for RTX Spark laptops, start here.

And once you have built a game, a controller that feels right in your hands helps when you play it through to check.

Read next

For the projects that came out of the ComfyUI canvas, see What People Made With ComfyUI This Week. For turning 3D worlds into something playable, see Building Playable 3D Worlds With an Agent. For putting video and music together on your own, see Making a Music Video Solo.