Mugen3D (@SumeruAI) header
Mugen3D (@SumeruAI) profile picture

Discord:https://nitter.gq/t.co/ZSZgJoRk49 Turn images into 3D with Mugen3D. API + open-source Unity sample. Follow for demos, technical breakdowns & updates.

New York, USA
Joined September 2025
One face rig, four moods, and you'd normally keyframe every one of them by hand. Here MugenMorph reads the emotion from the voice itself: happy, sad, surprised, angry. #MugenMorph #AIAnimation #LipSync #Mugen3D
1
74
Replying to @nao3art
ドット絵の質感のままカメラが回り込めるのは、平面のピクセルアートでは出ない奥行きですね。ほしみ町商店街の電線や提灯までsplatで立ってるのが面白いです。最後に結晶が光る演出はsplat側で動かしてますか、それとも別のパーティクルですか?
3
Replying to @_OOGEN
フェイストラッキングの質問がDMで来るの、納得です。Hash FT+Quest Proだと、ちびの大きい目でもまばたきや口の開きがきれいに乗りますか?成長したLapが外国の友達にいたずらしてるところも見てみたいです。
6
Replying to @siomi__o00
トンネルの照明が流れていく中、運転席の子の口が母音ごとにはっきり形を変えてて、大きく開けても崩れてないのが分かります。助手席の子が静かなぶん、表情の差がよく出てますね。HashFTで口を大きく開けた時の形、アバター側で何か調整してますか?
2
She had a whole speech ready. Then the door opened, and all that was left was one question. 17 seconds, driven straight from the audio — the mouth stays with the line, and the hesitation shows up in the face. #MugenMorph #FacialAnimation #LipSync #Mugen3D
1
24
Replying to @Noda_Mort
Driving the ears and tail off the mouth and brow params means they move with the expression instead of looping on their own. Is there a little delay on the tail so it trails the face, or is it tied straight to the parameter?
1
Replying to @Shmerz86
Visemes in the browser converter means you can catch a bad mouth shape before the VRM ever goes into an engine. Are they mapped from blendshapes already in the GLB, or generated when the mesh has none?
1
Our newest 3DGS world: a warrior and a snow leopard on a snowy ridge. It isn't a video clip. It's a 3D space, so the camera can circle anywhere and the scene holds together from every angle. #Mugen3D #3DGS #GaussianSplatting #3DArt
1
28
You shot a great video, but you can only watch it. You can't walk inside it, move the camera, or reuse a single object. We turn that footage into a 3D scene you can explore. #3DGS #GaussianSplatting #Video2Scene #Mugen3D
1
35
Replying to @saji8k
Multi-cam volumetric video without 4DGS is a pretty interesting route. I'm curious how editable the result is once it's a view stream — can you still move things or drop it into a regular 3D workflow, or is it mostly for playback?
2
Replying to @i_kemen
Totally agree. Recoloring in SuperSplat and LichtFeld is possible, but selecting the right splats and hitting a target color is the hard part. Building your own tool for it is the right move. Would love to see how the selection works once it's ready.
3
Replying to @tea_story_hoshi
Using the splat camera move as the base and letting Wan3 fill in the missing sky and add water ripples is a clever split. The splat keeps the camera path honest, and the video model covers what the capture couldn't. Curious what ended up on the "not so great" list!
3
A rainbow hat as an expression toggle is such a fun touch 😂 The goofy expression set gives the model a ton of character. Great Live2D work!
3
Replying to @Naiyn2009
Native face tracking support and better tongue movement in the same update is great timing, since a lot of people are switching over right now. Nice work on Nelva FT for Sesenola, the expression improvements sound like a solid step.
6
Replying to @osushiosu
RINDO answering through Open WebUI with lip sync, all on a 2070 Super, is a really fun home setup. Cutting the VRM from 49MB to 15MB is a smart touch too. How does the mouth timing feel with Irodori-TTS? Does it stay in step with the audio?
3
Replying to @6KLkQkFTkUmdS9t
You can tell from "it took really long" how much fiddling went into this. Congrats on finally getting face tracking on a VRoid-based avatar with the Pico 4 Pro! Nice to see a full 3-minute test instead of a 5-second clip, too.
2
Replying to @mal_vrc
Faces going goofy too easily is exactly the face tracking problem everyone runs into 😅 Picking from a list and just adjusting the strength sounds like a really friendly way to tune it, and non-destructive is a big plus. Looking forward to the BOOTH release!
2
Replying to @thomasmarcoscom
Native PLY and SPZ import in 5.3 is a really welcome step, especially if it actually lands in Workbench, EEVEE and Cycles without a plugin. Basic relighting already working is encouraging too. How does the relight hold up in Cycles versus EEVEE — does the splat stay stable, or does it still fall apart once you start pushing the lights?
3
Replying to @DuckbillStudio
撮影から1年、このデータセットを3DGSで完全攻略できたの、本当にすごいですね。朝起きたらこの景色が生成されていたのも気持ちいいです。地理参照して3DTiles化、地図表示までやるとのこと、楽しみです。今のシーンは、もう中に入って近づけるところまで来ていますか?
4
Replying to @LatentSpacer
Useful writeup — the 4090 timings make the resolution tradeoff pretty clear. A small talking-head frame does feel like the practical setting for most tests. At 832×464, does the mouth still stay tight to the audio, or do you jump to 1280 once it starts to smear?
2
Replying to @HeyGen @ideogram_ai
That stylized older character in the middle might be my favorite bit — the brows and mouth still felt locked to the sculpt when he started talking. I've found that's usually where 3D-looking faces go a little soft once audio takes over.
2