Drbaph

💀 Illustrator | 3D Designer | ML Researcher

[Drbaph @drbaph]

@BennyDaBall_OG @Kijaidesign @MiniMax_AI great examples 🔥

[Drbaph @drbaph]

@banodoco 🔥🔥🔥

[Drbaph @drbaph]

@Hamza_N_Meo pretty cool , good job

[Drbaph @drbaph]

@RoundtableSpace been using supertonic 3 for hermes on an orangepi, btw if you want to train your own voice and don't feel like paying i have a repo that will help you do it locally https://t.co/Og8ptnwZ7C

[Drbaph @drbaph]

@HectorNonce there is an option for masculine, feminine and blend

[Drbaph @drbaph]

GNM Studio v1.0.1 is out: a free, open-source desktop app and web experience for creating, tracking and animating Google GNM heads.

I reimplemented GNM v3’s released head-deformation core, including identity/expression basis evaluation, pose correctives, forward kinematics and linear-blend skinning in native Rust. Its output were validated against reference results from Google’s original Python implementation.

The Windows app runs natively in Rust with no Python, CUDA, cloud API, virtual environment or separate model downloads. Everything is bundled locally.

Features include real-time webcam tracking, 478 facial landmarks, 52 blendshapes, calibration, smoothing, manual expressions, choice of PBR skintone, backgrounds, lighting and microphone recording.

Export your animations as MP4, JSON or animated GLB for Blender.

I also built a fully client-side web edition for non-Windows users. It includes the base GNM head, tracking, animation, recording and exports; native seeded identity generation remains desktop-only.

#OpenSource #RustLang #MediaPipe #Blender

Give it a spin

Link in comments 👇

[Drbaph @drbaph]

Web app: https://t.co/LcwThvCKdC

[Drbaph @drbaph]

Repo:
https://t.co/fIxCBJp9JU

[Drbaph @drbaph]

@MosiAI_Official @huggingface amazing!

[Drbaph @drbaph]

@cocktailpeanut @kyutai_labs @MireloAI pretty cool, good job!

[Drbaph @drbaph]

@SlipperyGem longcat-2.0 for my Hermes setup atm is sweet, not the cheapest but exceeds deepseek for what i need it to do

[Drbaph @drbaph]

The biggest steady-state issue looks like repeated full-scene work

In the current build, live car reflections are configured with applyLiveUpdates=true, minFrameInterval=1, moveThreshold=0, and facesPerFrame=6, so the reflection path can render a cubemap of the scene every frame. Profiling backs this up: disabling reflections massively reduces draw calls/triangles and improves FPS. Shadows and active traffic physics are the next major costs

My proposed fix would be:
- Keep reflection quality, but cache/invalidate probes instead of updating all cube faces every frame. Update on meaningful car movement, lighting change, or nearby dynamic-object change; spread faces across frames; ideally render a reflection-only proxy scene/layer
- Keep shadow resolution/visual quality, but render shadow-caster proxies and cached clipmaps rather than full beauty geometry for world/building detail.
- Add traffic/physics LOD: player and nearby/interacting cars stay high-frequency, distant traffic becomes kinematic/lower-rate/asleep
- For the 30-60s construction problem, fix PCG streaming priority/cancellation: nearest first pages before details/colliders, real flush/progress, and don’t let far building work clog the queue. The worker already transfers buffers, so the issue is scheduling rather than typed-array copying

[Drbaph @drbaph]

@skbulous sick

[Drbaph @drbaph]

@Machinedelusion @nazar44444444 or egpu / dock

[Ting Chen Liang @ting_]

omg this made my day 🥹

Someone already made a ComfyUI node pack for @MosiAI_Official @Open_MOSS MOSS-TTS Local Transformer v1.5

Clone voices, generate speech in 30+ languages, and export 48 kHz stereo audio right from your workflow!!

thank you for building this🩵 @drbaph

[Drbaph @drbaph]

@ting_ @MosiAI_Official @Open_MOSS thank you for your work, this is an incredible model quite fast and very stable 🔥

[Drbaph @drbaph]

@RoRo_Rhoda @SMOUSE_CG delete it and add it outright from the gumroad shop its $0.00

[Drbaph @drbaph]

@ting_ @huggingface https://t.co/SRCrYgLe5r

[Drbaph @drbaph]

@Gcabrielclark ComfyUI ready, bf16 & fp8 mixed
https://t.co/Tt4WPTTXAK

[Drbaph @drbaph]

@cocktailpeanut @ideogram_ai even with propper json schema if the prompt is fairly simple or short it will trigger it, some people bypass the model's first 2 layers by manual sigmas and it never triggers the safety filter