Spin, match and fix a character on your own machine

The camera swings all the way around the subject, on a local GPU. That is @AiPhotorealGirl's video. Put the other posts next to it and a small studio takes shape around a single character: lock the face, pull poses out of the orbit, fix the expression, and let an LLM write the motion.

Spin: an orbit video on a local GPU

We start with @AiPhotorealGirl.

The creator writes (in Japanese) that the video was generated with ComfyUI v0.39.0, an RTX 4070 12GB and a MiniMax-H3 360° Orbit LoRA, at 2.0MPixel (1440x1440), and that generation took about 17 minutes ¹.

The memory detail is the interesting part. The creator says they think the previous version ran out of VRAM even at 1.5MPixel, and that the latest version fixes this and is faster too ². That is the creator's own read.

The ComfyUI v0.39.0 release notes list changes such as releasing MiniMax H3 embedding temporaries before the blocks run ³ and a fix for a MiniMax VAE offload issue ⁴. The notes do not say which of these, if any, made the difference on the creator's machine.

Orbit-camera LoRAs are also distributed on Hugging Face. MiniMax-H3 ORB360 is an example. According to its page, you give it a photo and get a camera orbiting clockwise through 360 degrees around the subject ⁵. For multi-view use, the author recommends the step1500 file ⁶. You load it with the standard Load LoRA node onto a MiniMax-H3 Ref2VA ⁷, and the license is the MiniMax H3 Community License Agreement ⁸. @AiPhotorealGirl's post names the LoRA only as "360° Orbit LoRA"; the name ORB360 does not appear in it.

For a comic panel brought to life with ORB360, see Images, Video and Music at Home.

Match: locking a face with RefMod

Next is the opening post of a thread by @ai_hakase_.

The thread introduces RefMod, a way to lock a character's face in MiniMax H3. The creator writes (in Japanese) that a file of about 1MB, made in a few minutes, is enough to "completely fix" the target face ⁹. The steps are in the rest of the thread.

According to the distribution docs, the RefMod node system runs on the ComfyUI custom node ComfyUI-MiniMaxH3Mod ¹⁰. The RefMod models are released as open weights ¹¹.

Match: pulling poses out of an orbit video

@IsekaiBardGuild shows another use for an orbit video.

The creator writes (in Japanese) that they built a ComfyUI custom node that extracts images from several viewpoints out of an orbit video made with MiniMax H3 360 Orbit, estimates the pose, then runs depth estimation and coordinate correction ¹². The resulting pose can be used in workflows such as the Qwen Image 2.1 pose-specification workflow ¹³. The creator adds that the orbit video does not have to come from MiniMax H3: video from other AI tools or live footage works too, as long as it meets the requirements ¹⁴.

Spinning is not the end of the line; the result becomes material for the next step. That is the studio feel of this setup.

Qwen-Image-2.1's README says the developers open-sourced it ¹⁵; the license name is qwen-research ¹⁶.

Fix: dragging handles on the face

The face and clothes look right, but the mouth is open too wide or the gaze is off. @tkvtk has a post for that.

According to the creator, who writes in Japanese, the ComfyUI node "NKD Face Rig" lets you adjust an expression by dragging handles placed on the face ¹⁷. Touching the spot you want to fix directly, they say, is handier than hunting for words or numbers that describe an expression ¹⁸. They add that the node makes LivePortrait's expression control available through this interface ¹⁹. They also say they confirmed it works on both anime and photorealistic faces ²⁰.

NKD Face Rig is under the MIT License ²¹, and so is the LivePortrait code ²².

If you sketch character sheets or expression drafts by hand before the fix, a pen tablet makes that part of the loop easier.

Let an LLM write the motion

The last post is from @haribote0073.

The creator writes (in Japanese) that the prompts are written by an LLM ²³, and that an instruction like "wave, spin around once, wave again, jump and squat" was all it took to produce the attached video ²⁴. The post does not say which tool made the video.

The memory story shows up here as well. The creator writes that adding an RTX4070 to a sub PC and moving to Qwen-3.8 27B at Q4 raised the quality ²⁵, and that on the integrated GPU a 9B model at Q3 was the most they could manage ²⁶. Qwen3.8-27B has 27B parameters ²⁷, is licensed under Apache-2.0 and ships open weights ^{28 29}.

If you plan to sell the results

Commercial use of locally generated MiniMax H3 outputs requires a MiniMax commercial license ³⁰. If you plan to sell your character, read the official terms first.

On your own desk: the memory question

Some of these posts mention memory ^{2 26}: one creator thinks the previous version ran out of VRAM, and another says a 9B model was the most the integrated GPU could manage.

Here is what NVIDIA says about RTX Spark.

• According to NVIDIA's product page, up to 128 GB of unified memory handles 3D scenes up to 90 GB [[F100]] [[F141]].
• On the same page, NVIDIA says you can generate images and video right on your device, as often as you like, with no subscription and no per-prompt fees [[F101]].
• NVIDIA's page also says you can run models up to 120B parameters with 1M context [[F103]].

ComfyUI is a matter of announced plans. NVIDIA says RTX Video Frame Generation will be coming to ComfyUI, with the update arriving this fall with RTX Spark ³¹. A quote from ComfyUI co-founder Yannik Marek appears on NVIDIA's blog (Japanese edition). In the future tense, it says ComfyUI users will be able to run very complex multimodal workflows and generate ultra-high-resolution images and video on portable devices at unprecedented speed ³².

None of the post texts above names an RTX Spark machine as the hardware used. If you want to use MiniMax H3, RefMod, ORB360, NKD Face Rig, Qwen3.8-27B or similar tools on an RTX Spark machine, check each tool's official documentation for supported environments first.

Examples of RTX Spark machines:

Once orbit videos and multi-view images start piling up, a place to keep them helps.

Read next

For mass-producing MiniMax H3 video on your own machine, see Videos Pile Up While You Sleep. The ORB360 comic is in Images, Video and Music at Home.