"Local AI Is Becoming a PC Selling Point" — What People Made With ComfyUI This Week: Comic Panels, Long Videos, and Pictures Turned Into 3D

"Local AI has become a selling point for PCs," writes shinmomo (@shinmomo0512) in Japanese ¹. As the post sorts it, Apple's new MacBook Pro comes with LM Studio, while Microsoft's new RTX Spark laptops come with ComfyUI ². The poster adds that, for their own Mac Studio too, the real thing starts from here ³.

The post links to one from the official ComfyUI account, which opens with "The trend continues" and says Microsoft is showcasing ComfyUI for the new RTX Spark laptops ⁴.

ComfyUI runs locally on Windows, Linux and macOS ^{5 6 7}. Its license is GPL-3.0 ⁸.

So what has come out of that node screen this week? A comic page, a long video, a picture turned into 3D. All of it from the same ComfyUI canvas.

Pick a comic layout from a sample

Kamimoto (@sep_is_heim) released a ComfyUI custom node called ComfyUI-MossPanelDataset. The creator writes (in Japanese) that it lets you make comics with MOSS-Image ⁹.

Here is how it works. You pick a panel layout from 196 samples, then just enter characters, scene and dialogue; no hand-written JSON or coordinates, the creator writes (in Japanese) ¹⁰. It handles 2 to 6 panels, and the dialogue can be drawn in Japanese as is ¹¹.

The README says the same. You choose the layout from images, fill in characters, scene and dialogue in dedicated fields, and the node hands it all to MOSS, with no per-layout workflow downloads and no hand-written JSON or coordinates ¹². It supports 2 to 6 panels ¹³, and the bundled layouts total 196 ¹⁴. Character looks and the scene go in as English; dialogue is passed in the language you want drawn. No translation and no extra LLM is involved ¹⁵.

The licensing differs by part. The node's own code, layout data and bundled sample images are MIT ¹⁶. MOSS-Image is Apache-2.0, distributed for ComfyUI by Comfy-Org ^{17 18}.

The README is also candid. It says the node does not guarantee against garbled or broken text, mixed-up characters, or panel counts drifting ¹⁹. If you sketch your comic storyboards by hand, the step before you feed this node is where a pen comes in.

Long video, passed along in a relay

Next is video. @superfang119 posts in Chinese about a new way to make long videos with MiniMax H3. The post says (in Chinese) it does not swap the model; it adds a relay system to H3 so long videos can be chained ²⁰.

The problem it names is concrete. Short clips are easy to generate, but when you try to run for a long stretch, the continuity of characters, motion and picture drifts. The post calls this a real problem with H3 run locally ²¹.

The README of the project they point to, ComfyUI-MiniMaxH3-PrefixStream, describes it like this.

• The clean video and audio latents of the previous segment are written into the start of the next, and protected with ComfyUI's native mask. Video and audio are controlled separately [[F447]]
• A shot bin (Clip Bin) stores each segment's preview, rating and prompt, and lets you pick any past shot from a gallery to continue from [[F448]]
• Each segment is written to disk, and the pieces are joined into an MP4 at the end [[F449]]. FFmpeg is required [[F449]]

It does not change ComfyUI's source. The original model, prompt, sampler and decode chain stay as they are, and only the nodes for continuation are added ²². A 42-node workflow ships with it ²³. The license is MIT ²⁴.

If you are making long videos, you will want room on the drive you write them to.

A video piece from Krea 2 and MiniMax-H3

ou_zhen599 (@ou_zhen599) wrote "Made with Krea 2 +MiniMax-H3 ComfyUI" alongside a video ²⁵. The piece is titled "First Ignition," and the creator describes it as a scene running from suppressing her excitement to truly taking control, as she completed her very first ignition ²⁶.

The creator says the full prompt is in the reply ²⁷. If you collect projects, the words that summoned this one are worth a look.

A look at the tools behind it. Krea 2 is published as a repackage for ComfyUI by Comfy-Org ²⁸, under the Krea 2 community license ²⁹. MiniMax H3 has open weights, so it can run locally ³⁰. Using locally generated output commercially requires a MiniMax commercial license ³¹.

From a picture to a 3D form you can spin

Last is 3D. Nobin (@NobinLOG) reports in Japanese that a model called Idol has appeared for ComfyUI. It makes a 3D object that can be rotated through 360 degrees from a single image ³². The post adds that ComfyUI has converted it into an easier-to-use format and distributes it ³³.

The creator lists cautions too. It is a Gaussian splat, unlike the polygon meshes normally used in Blender, and the back side is an AI estimate, so reproduction is not guaranteed ³⁴. The creator tried it quickly and says the challenge is how to raise quality from here ³⁵. The post closes by saying VRAM is not much of a requirement, so give it a try ³⁶.

Now the official material. Idol's developers say it runs on consumer GPUs, using 7.2GB of VRAM and sometimes running in under 2 seconds ³⁷. That describes usage, not a minimum requirement ³⁷. From a single image it builds a 3D Gaussian splat, including sides that were not visible, and writes up to 1,048,576 Gaussians to a standard .ply ³⁸. The license is Apache-2.0 ³⁹. The ComfyUI version is a fine-tune of TripoSplat and works with existing TripoSplat workflows ⁴⁰.

The official SDK's install requirements are Python 3.10 or later, PyTorch 2.5 or later, and an NVIDIA GPU with CUDA ⁴¹.

What PC Watch measured on an RTX Spark laptop

Line these up and the common thread is plain: comics, long video, a video piece, a picture into 3D, all coming out of the same ComfyUI canvas. shinmomo's post ends by asking (in Japanese) whether readers will look at this when they buy their next PC ⁴². Looking at these projects, it is a fair question.

So how does an RTX Spark laptop do with ComfyUI? PC Watch tested the ASUS ProArt P16 in its 64GB configuration ⁴³. Image and video generation were measured with ComfyUI workflows, with 1 warm-up run followed by 5 measured runs ⁴³. FLUX.2 took about 14 seconds with Klein and about 40 seconds even with Dev ⁴⁴. The Klein run was an NVFP4 workflow at 1,024×1,024 and 20 steps ⁴⁵, and the Dev run also used an NVFP4 workflow ⁴⁶. PC Watch writes that this makes it easy to picture using it to try visual ideas while adjusting prompts ⁴⁴. With LTX2.5, generating 5 seconds of video and audio took about 69 seconds on average ⁴⁷. PC Watch also notes that processing takes longer than the clip itself, so it does not mean real-time generation ⁴⁷. All of these are results for that machine and configuration, with the models PC Watch chose, not the models in this week's projects.

NVIDIA announced RTX Video Frame Generation for ComfyUI and said at the time that it would arrive this fall with RTX Spark ⁴⁸. ComfyUI co-founder Yannik Marek, in a statement on the NVIDIA blog, says in the future tense that ComfyUI users will be able to run very complex multimodal workflows and generate ultra-high-resolution images and video on portable devices ⁴⁹.

The cards below: Microsoft's RTX Spark laptop, the Surface Laptop Ultra, and the ASUS ProArt P16 that PC Watch tested ⁴³.

If you want to try this week's projects on your own desk, check the README and official pages of the nodes and models you want first, to see which environments they support. The MossPanelDataset README, for example, says its original dataset was generated on Windows with an RTX 3060 12GB ⁵⁰, and that real generation on other GPUs and operating systems is unconfirmed ⁵¹.

You can also look for RTX Spark laptops on Amazon.

Read next

For NVIDIA's Japan launch event, see the RTX Spark Japan showcase article. For building video and music on your own, see Images, Video and Music at Home. For spinning and matching a character, see Spin, match and fix a character on your own machine.