Look at the armored rover in clay render. You can see the panel seams and bolts in the geometry. They still show up when you orbit the model and the light moves across it.
Direct3D-V3 refines a coarse mesh into detailed geometry using a reference image. The team released the refinement code and weights on November 19, with an interactive gallery where you can compare the result to other methods on the same input.
🧩 How it gets that detail
▪️ The model keeps voxel patches near the surface of the coarse shape, where the detail actually matters.
▪️ One transformer works across those patches to keep the broader shape coherent. Another works at the voxel level to add fine geometry.
▪️ Diffusion runs directly on signed distance values. The authors report working at grid resolutions up to 1536³, with sparse patches around the mesh surface.
The release setup takes an aligned OBJ, GLB or PLY plus the reference image. DreamTech’s own coarse mesh generator and texture model are still on the release roadmap. The repository points to Hunyuan3D-2.1 and UltraShape as options for producing the starting mesh.
If you open one example, make it Serpent Guardian and orbit the clay mesh. That’s where you can judge what the extra geometry buys you.
dreamtechai.github.io/Direct…
github.com/DreamTechAI/Direc…
#3DGeneration #GameArt #3DModeling #DiffusionModels
8
Ten seconds of your voice, and your laptop can start reading a script back to you while it’s still generating the rest.
BlueTTS V3.0.1 from Maxim Melichov runs locally with ONNX Runtime. It can stream speech in chunks and exposes an OpenAI-compatible /v1/audio/speech endpoint, so you can point an existing TTS client at your own machine.
🎙️ What the reference gives you
Use about 8 to 10 seconds of clean speech. In the author’s tests, the clone picked up the speaker’s timbre and typical pitch. Speaking pace and delivery carried over less reliably. There are no dedicated emotion or style controls, so I’d use it for narration or placeholder dialogue where a consistent voice matters more than matching a performance.
⚡ The CPU numbers
On a two-vCPU Intel Xeon container, the author measured 5.4× real time with the INT8 estimator at eight steps, and 9.5× at four steps. That’s audio generated per second after the model loads. Streaming lets you start listening as soon as the first chunk is ready.
The local bundle supports Hebrew, Yiddish, English, German, Spanish, Italian and Russian. The big Windows catch: the bundled pronunciation library has builds for Linux x86-64 and macOS. With the current native Windows setup, Hebrew and Yiddish work, while the other five languages need that missing dependency.
Code is MIT. The bundled audio decoder has separate OpenRAIL-M terms.
github.com/maxmelichov/BlueT…
huggingface.co/notmax123/Blu…
#TextToSpeech #VoiceCloning #ONNX #LocalAI
8
Arrange a few props, change the lighting, pick a style, and see what the scene looks like as a painting.
That’s the loop in Glenmara: Still Life Artist. Bit Loom built it in Godot, and the November 19 interview with the team has a good detail about how the paintings work.
Programmer Henry Pullan built a library of shaders and combines them in different stacks to get different looks. Moving from Unity to Godot meant reworking how the game sends data to and from the GPU at the right time. The painting effects can process an image over a few frames, so the team has room to experiment with those combinations.
The team shows styles ranging from a mosaic to cel shading and pixel art. They all work on a still life you arranged yourself.
I like the trade-off. While you’re arranging the table, the placement and lighting have to feel responsive. When you make the painting, the game can spend a few extra frames getting from the 3D setup to the effect you picked. That’s a useful distinction if you’re building a photo mode or a game where creating the picture is the point.
Bit Loom says the painting styles are hand-built shaders. Glenmara is coming to PC and is available to wishlist on Steam.
store.steampowered.com/app/4…
#GodotEngine #Shaders #IndieGame #GameDev
1
72
The feature demo I’d start with in Silo 2027 is the Gothic Arcade.
It’s one of the scripts @nevercenter ships with the update. You can open it, see how the geometry is built, change its inputs and give those inputs their own little window. Think floor count, spacing or depth on a small tool you can reuse on the next asset.
Silo now has a JavaScript API with over 100 functions, a built-in script editor and keyboard shortcuts for scripts. A script window can expose numbers, checkboxes and menus. Once you find settings you like, you can save them as another script.
🧩 Where the AI part fits
Turn on Allow AI Connection and compatible desktop assistants can connect over MCP. They can inspect the scene, take a viewport capture, and edit or run scripts. You could use that to draft a modeling tool, then keep the script and run it yourself with the same controls next time.
Silo also improved glTF material and texture handling and broadened what its Bridge tool can connect, including loops of different sizes. It’s a paid modeler for Windows and macOS. The sample scripts come with 2027.0.
nevercenter.com/index.php/Si…
#3DModeling #GameArt #JavaScript #MCP
1
39
Paint a bike lane down a procedural street. Now connect that street to the next one without losing the bike lane at the join.
That’s one of the practical changes in ArcGIS CityEngine 2026.1. When two street segments have matching lane counts, CityEngine can generate separate lane shapes through the joint. Those shapes carry attributes like the ones on the street itself, so the same rule can keep working through the connection.
🏙️ Two more details for building a city block
▪️ Building setbacks now work around holes in a footprint. That gives you a way to shape the inside of a courtyard as well as the street-facing walls.
▪️ Esri’s library adds facade components for different floors, including windows, balconies and shading. You can use the same basic facade layout across several buildings and vary the details through its rules.
There are also exporter improvements. Texture transforms can be baked into UVs so the tiling survives export, and the glTF exporter is among the ones Esri updated.
I’d look at the Complete Streets example first. Turn on the lane rules and inspect what happens where two curved street segments meet. It makes the update a lot easier to understand than a screenshot of an entire city.
esri.com/arcgis-blog/product…
#ProceduralGeneration #ArcGIS #3DModeling #UrbanDesign
13
You’ve scanned a prop from dozens of photos. A patch of the texture still looks wrong. Where do you fix it?
3DF Zephyr 9.5 makes it easier to work back from the reconstructed model to the source images. The release also updates tools for editing textures, filling mesh holes and working with terrain. It’s the cleanup part of photogrammetry that makes this update worth a look.
There’s a second release alongside it. @_3dflow_ introduced 3DF Jimbo, a separate tool for training segmentation models on your own labeled photos. Those models can then be used for classification in Zephyr or for masking source photos in the separate 3DF Masquerade tool.
Here’s where I’d try it. Say you photograph the same kind of prop on a turntable every week, and the support rig keeps showing up in the photos. A segmentation model trained on examples from that setup gives you a way to reuse those labels across captures. You still need enough examples for the model to learn what to pick out, but repeated shoots give you a concrete reason to do the labeling work.
The training input is a set of images paired with masks by filename. That gives you a practical split: tag regions in the photos you already understand, then let Jimbo learn the pattern you want to repeat. Inside Zephyr itself, 3Dflow also reports faster camera reconstruction and multi-view stereo, with no single speedup figure in its release notes.
Zephyr has a Free edition and paid editions with different limits. Check the edition before planning a full scanning workflow around a specific tool.
3dflow.net/3df-zephyr-9-5-is…
3dflow.net/3df-jimbo-segment…
#Photogrammetry #3DScanning #VFX #GameArt
1
58
Give two Unity characters the same line. One keeps eye contact. The other glances away, pauses and leans back. Both can share the same dialogue audio.
FaceCue from Firecrest Roost drives mouth shapes, gaze, blinks, head movement and other facial cues from dialogue. Each layer is optional, so a background NPC can use just the mouth animation while a close-up gets the full setup.
🎙️ What happens to the actual clip?
Bake a line, open it in the Clip Composer, and scrub to the syllable that looks wrong. You can tweak the mouth shape and preview the result without entering Play Mode. FaceCue includes preset mappings for families like MetaHuman, VRoid and Ready Player Me. You can map other rigs yourself.
The gaze system is the bit that makes this interesting for cutscenes. There’s a simple look-at mode and a more involved option that models how the eyes and head move toward a target at different rates. That’s useful when a character should glance at something without swinging their entire head with their eyes.
The tool also gives you a way to reuse a performance at runtime. Bake the audio once, add the clip to a registry or Addressables, and play the line from the character’s AudioSource. Timeline integration can listen to ordinary audio tracks through a watcher. That’s particularly handy when you already have a pile of dialogue edits in Timeline.
FaceCue is a paid Unity Asset Store extension. Platform support varies by feature, especially for live recognition. Pre-baked dialogue is a useful place to start when you already have the recorded lines.
assetstore.unity.com/package…
#Unity3D #GameDev #FacialAnimation #LipSync
17
Blender’s tablet roadmap has a funny first milestone: full usability with a keyboard and mouse.
The November 18 update from Dalai Felinto and Pablo Vazquez lays out how @Blender plans to get there. The team is working on making the existing Blender build and run reliably on tablets, with basics like opening and saving files. Android is the current focus. Recent Apple tablets are part of the hardware discussion, with iOS support still under consideration.
🖊️ The touch ideas get more interesting once you see the mockups.
A floating radial menu could stand in for Ctrl, Shift and Alt. The status bar could offer tappable confirm, cancel and axis controls. On small screens, groups of toolbar buttons could collapse into one popover when there’s no room to show them all.
The plan also covers moving projects between tablet and desktop, then exploring compact layouts around specific jobs like sculpting, scene assembly and storyboarding. Those are the parts I’d watch if your ideal setup is sculpting on a tablet and picking up the same file at your desk.
There’s already an experimental Android demo from Blender Conference. The Wacom and iPad images in the November update show interface mockups. The team hasn’t announced a stable tablet release date.
code.blender.org/2026/11/tab…
#Blender3D #3DArt #DigitalSculpting #Android
27
How far can your printed figure bend its knee before the calf hits the thigh?
Click Joints gives you an answer in the Blender viewport. Set the 3D cursor where the joint belongs, choose a ball-and-socket or pin joint, and the free extension cuts the sculpt into printable pieces with the clearance you specify in millimeters.
The part worth watching is the bend preview. The tool checks the ball joint in eight directions. As you rotate the part past the measured limit, the contact area turns red. You can see which bit of the sculpt needs more clearance before you export.
A thin ankle is a good test. The extension checks the wall around the socket and marks a tight fit or a cut that would break through. You can inspect the two pieces in an exploded view, then export individual STL or 3MF files in millimeters.
Blender 4.2 LTS or later, GPL-3.0-or-later. The developer says the fit and bend checks were measured in Blender. I’d start with one test joint to dial in a printer’s tolerances before committing to the whole figure.
extensions.blender.org/add-o…
#Blender3D #3DPrinting #Sculpting
22
The little boat bobs in the lagoon, the waves break around the island, and the shore leaves a line of foam. That’s a Godot scene you can open in your browser.
TinyLanternDev’s Stylized Water pack uses Gerstner waves for the surface and adds depth-based color, shore foam and caustics. The full pack also has floating objects, a waterfall with a foam pool, and an underwater view with fog and moving light rays.
Version 1.1 adds a fun detail: drop an object into the water and you get a splash crown, a jet and rings that fade out. That gives the water a response you can actually use in a game. A rock tossed into the cove and a boat crossing it can both change what the player sees.
If you’re trying it for your own scene, here’s how the downloads differ.
▪️ Free Lite gives you the Tropical preset, a simple example and a shader with two Gerstner waves, depth color and shore foam.
▪️ The $14.99 full pack includes six looks, from a turquoise cove to swamp water and lava. The web build skips screen-space reflections and some of the heavier layers.
One setting I’d borrow for a trailer is the seamless loop. Set its duration and the waves, foam and particles come back around together. That’s handy when you want an idle background or a short clip that doesn’t visibly jump at the cut.
The example scenes use Kenney’s CC0 props. The water pack itself allows commercial game use under its asset license.
tinylanterndev.itch.io/styli…
#GodotEngine #Shaders #GameDev #VFX
1
99
Draw the silhouette of an anime fringe. Split a tip into two. Turn the locks into quads.
That’s the workflow in HairMesh, @koankitsu’s free Blender port of catarackta’s Stylized HairMesh Creator. You draw guide curves over the head, stretch surfaces between them, and convert the locks into a polygon mesh with evenly spaced loops.
✂️ Three places the 1.1 updates help
▪️ Blocking the fringe. Flat locks let you set the width at each anchor. Start wide near the scalp and taper the same ribbon toward the tip.
▪️ Breaking the silhouette. Extrude a curve from either end or split a lock’s tip into separate points. The new volume branches give you another option for layered tufts.
▪️ Revising after conversion. Switch between the editable curves and baked geometry. When you convert again, HairMesh rebuilds the changed pieces. Version 1.1.1 also lets you see the other half of a mirrored group while you work.
I’d use symmetry to get the overall haircut in place, then break it around the face. You can judge that change on the actual mesh and hop back to the curves if the bangs cover too much of the eye.
Blender 5.1 or later. The add-on is GPL-3.0-or-later and the curves live in the .blend file.
github.com/riel-gab/hairmesh…
#Blender3D #CharacterArt #3DModeling #GameArt
20
That giant serpent is made of Gaussian splats. So are the room it fights in and the particles it throws at you.
A developer used Claude and @JungleSilicon’s Ink Studio to build a browser roguelite prototype. In the gameplay clip, the serpent rears up, the player moves out of the way, and the screen fills with particles. The developer says the prototype took a day.
🎨 Where the splats come from
Ink Studio starts with shapes defined in JavaScript using signed distance functions. You can combine simple shapes, bend them and carve pieces out. It samples their surfaces into colored splats, then renders them with a deliberately pixelated look.
Those splats can move. The toolkit includes character controls, particle systems and tools for animating objects. The same pipeline can give you a monster that moves and an effect that breaks apart on contact.
The reason I’d try this for a game jam: you can change a shape in code and see how it reads from the actual game camera. In this demo, the oversized boss silhouettes do more for readability than extra detail would.
There’s room for more than combat effects in the same renderer. Ink Studio’s README includes examples for cloth, hair and procedural shapes. They share the same splat data, even though the rules that move them differ. The toolkit is MIT-licensed.
github.com/siliconjungle/sha…
#GaussianSplatting #GameDev #PixelArt #WebGraphics
1
1
104
A small group of ChatGPT Plus users is getting a Dot that can keep working after they close the chat.
OpenAI’s @thsottiaux says the company is testing Dots with a small Plus cohort. There’s a wider rollout on the work side, where eligible ChatGPT Business Standard members in the US are getting one personal Dot each.
🛠️ What would you hand off?
Give a Dot a Friday routine. Have it review the documents and updates it can access, pull out what changed, and draft a status update. The point is to assign the job once and come back to the result. Each Dot has its own cloud computer, remembers past work, and can use the apps you’ve connected.
📱 The mobile app is changing too
ChatGPT is adding a dedicated Dot tab to the bottom bar. You’ll also be able to pin things like Codex, Projects, and Images there. The redesign is rolling out first to eligible Pro users in supported regions, with a small group of Plus users getting it too.
nitter.gq/thsottiaux/status/2123…
help.openai.com/en/articles/…
#ChatGPT #AIAgents #OpenAI
Heads up, we’re starting to test dots with a small number of ChatGPT Plus users. Looking forward to the continued feedback as we expand and I’m recently seeing a lot more positives around how good the proactivity has been after many improvements.
We’re also bringing dots to all ChatGPT Business Standard users. Rollout will start today.
1
76
“Can you check order B-714? Wait, it’s D-714.”
That’s the kind of interruption a voice agent has to handle while the caller is still talking. OpenAI has brought GPT-Live-1.1 to the API after using it in ChatGPT Plus and Pro voice since October 15. It can listen while speaking, and OpenAI reports better instruction following, more accurate transcription of letters and numbers, and a steadier voice and accent when switching languages.
🎙️ The voice model handles the conversation. A backend you choose handles the order lookup, reasoning, or tool calls. The caller can add a detail while the lookup is running, and the voice model can relay the result when it comes back.
💵 The bill depends on how long inference stays on:
▪️ Active: $0.05/min, including silence before the idle timeout.
▪️ WebRTC Standby: $0.02/min while inference is paused. Set the timeout to 5–60 seconds. The connection stays open and speech wakes the model back up.
Usage is billed per second. Backend models and tools are charged separately. Standby is also available on GPT-Live-1.
The practical test for a support agent is whether it catches the corrected order number and looks up the right order while keeping the caller in the conversation.
Model ID: gpt-live-1.1
developers.openai.com/api/do…
#VoiceAI #AIAgents #OpenAI #WebRTC
1
25
Google built a fake internet for Chrome’s AI agents. You can fill a cart, apply a promo code, or scroll a social feed. Close the test and the next run starts from a clean slate.
It’s called Mirror Dimension. Chrome’s security team uses it to test Auto Browse on thousands of simulated sites. The agent uses Chrome, while the sites and accounts live inside an isolated environment. A checkout can behave like a checkout without placing a real order.
The hard part was making the sites feel ordinary. Give a capable agent a demo page with one button and it may work out that it’s being tested. Google calls this eval awareness. If the model behaves differently when it spots the test, the score tells you less about how it will behave on a real site.
So Google also put agents to work on the fake web itself. Some build the sites. Others click through them, check flows like form validation, and report problems for another round of fixes.
🛒 The test I’d want to see: can an agent finish your shopping task while ignoring a page that tries to give it a different job? The page can tell the agent what’s in stock. You’re still the one who gets to decide what it does.
wired.com/story/google-built…
#GoogleChrome #AIAgents #AISafety #BrowserAgents
30
Harvard’s AI safety seminar has a session on a question every model evaluator runs into: what happens when the model gets better at satisfying the grader, but the answers people get are worse?
That’s one of 13 sessions on the fall 2026 syllabus for CS 2881R. Boaz Barak (@boazbaraktcs), who teaches at Harvard and also works at OpenAI, is posting readings, selected lecture recordings, notes, slides and links to student experiments as the course goes on.
📚 Three places to start:
▪️ Modern LLM Training, September 17. The reading list connects pretraining, instruction tuning and reinforcement learning with the mechanics of inference, including KV caches and quantization.
▪️ Grading and Reward Hacking, October 8. John Schulman’s session has public notes. The student presentation and code explore how to measure the quality of an environment used for RL self-play.
▪️ Open Models, October 22. Nathan Lambert’s talk and the student experiment are on YouTube. Sam Altman’s Q&A was part of the same class, and the posted video starts with the later session.
The syllabus also covers AI economics, model policies, interpretability and frontier regulation. Harvard expects students to be comfortable with proofs, probability, information theory, basic ML and training a simple neural network in Python.
You can follow the public materials without enrolling. Some guest sessions were left unrecorded. The fall 2025 archive has its own videos, notes and projects.
boazbk.github.io/mltheorysem…
boazbk.github.io/mltheorysem…
#AISafety #MachineLearning #LLM #OpenEducation
3
1
99
Watch the edges of the meerkat’s fur in Apple’s depth demo. A good depth map needs to trace those hairs and keep the animal at a steady distance as the video plays. A crisp outline in one frame only gets you halfway there.
Depth Pro 2 estimates depth in meters from a photo, a video, or a live feed from one camera. Its transformer shares information across frames, and a decoder predicts a depth map for each frame at the input resolution. Apple’s largest checkpoint led the methods it compared on boundary sharpness and temporal consistency in the paper’s tests.
📏 Where does the scale come from?
The depth model predicts inverse depth normalized by focal length. If the focal length is known, you can supply it. If the camera metadata is gone, Apple’s separate 23.5M-parameter estimator predicts the focal length in pixels. Scale follows that estimate: a focal length 10% too long puts the predicted depths 10% too far away.
The four depth checkpoints range from 24M to 872M parameters. At 1080p on a single NVIDIA B200, Apple reports 53 fps for the smallest checkpoint in streaming mode and about 8 fps for its largest video model. The CLI can save the depth in meters and the focal length used, so you can use both to build a point cloud.
Apple released the code and weights. The weights are licensed for non-commercial scientific research and academic development.
apple-aiml-research.github.i…
github.com/apple-aiml-resear…
#ComputerVision #DepthEstimation #3DReconstruction #AIResearch
1
87
Watch the lettering on the bakery sign in Wan2.1. It looks fine in a still frame. In motion, the letters bend and flicker. The final video decoder can introduce that kind of distortion even after the model has finished denoising.
Decoding Without a Decoder, from @HanshengCh and colleagues, gives the diffusion transformer a second output head. The original denoising head still steers the sampler through latent space. At the final step, a separate linear head reads the transformer’s tokens and predicts pixels directly.
🧩 Why should that work?
The authors froze pretrained models and fitted simple linear probes to two places: the denoised latents and the transformer’s last-layer tokens. The tokens reconstructed pixels more accurately. For DWD, they fine-tuned the last few blocks, trained the pixel head with its own losses, and kept the denoising head close to its original predictions. That keeps results close for the same seed while changing how the final image is produced.
What changed in the tests:
▪️ Wan2.1 14B: 23% lower inter-frame warping error, with less flicker on signs, wire mesh and dense lights.
▪️ FLUX.2 klein Base 9B: LongText went from 0.909 to 0.908. In that adaptation comparison, DWD took 350 H100 GPU-hours versus 1,100 for AsymFlow.
▪️ AnisoFlow on ImageNet: FID improved from 1.55 to 1.43. AnisoFlow already works in pixel space, so this result tests the value of a separate pixel head on its own.
DWD also powers the experimental direct-pixel output in pi-H3, the two- or four-step MiniMax H3 variant. The DWD project page has same-seed comparisons. Code is still marked as coming soon.
hanshengchen.com/dwd/
arxiv.org/abs/2611.22287
#DiffusionModels #AIVideo #ImageGeneration #AIResearch
1
143
Ask a model what helps a bird fly, and it points to a line about hollow bones. Delete that line and only a few answers are likely to change. In this experiment, you can open the file and see exactly what the model was given to work with.
POSTECH researchers Hyunjong Ok and Jaeho Lee gave Claude Opus 5.5, GPT-6.1 Sol and GPT-6 Astra ten labeled examples from the AI2 Reasoning Challenge, a multiple-choice science benchmark. Their job was to build a separate model without gradient training. Given freedom over the architecture, all three wrote facts in plain text, compiled them into sparse tables, and wrote a fixed procedure that scored the answer choices.
Three numbers from three different runs:
▪️ 54.4% ARC accuracy came from the best fifth-generation Claude-designed table.
▪️ About 1,800× fewer parameters came from an earlier Claude-designed model: 0.94M parameters and 50.8% accuracy, compared with 1.72B and 48.7% for a Transformer architecture whose weights Claude also wrote.
▪️ Up to 17× faster came from the researchers’ own indexed associative-memory design. It scored 52.4% and processed the 3,530-question test set in 18 ms of GPU time on an RTX 3090, versus 302 ms for a size-matched Transformer. The 54.4% table was slower than its size-matched baseline.
The authors’ fine-tuned Qwen3.5-0.8B reached 82.1% on the same ARC evaluation. The appealing part of the written tables is how easy they make it to inspect an answer and patch the fact that drove it.
arxiv.org/abs/2611.22055
github.com/effl-lab/Weight-S…
#AIResearch #MachineLearning #Interpretability #LLM
28
The Classroom already exists as a Blender scene. What if you could send someone the lighting, materials and all those little details as a splat they can walk around in their browser?
SplatPilot turns an object, a collection or a full scene into a Gaussian splat inside Blender 5.3. The useful shortcut is that Blender already knows where the cameras are and where the surfaces sit.
📷 From scene to splat
1. Place cameras around the subject or through the room. The coverage map lets you find badly seen surfaces before spending time on renders.
2. Render the views in Cycles or EEVEE. Camera poses come directly from Blender and the position pass seeds points on the actual surfaces, so there’s no COLMAP solve.
3. Train on the GPU. The model uses color, alpha and depth. Rooms and landscapes also get low-resolution views used just for depth and alpha, which help constrain empty space and surfaces.
4. Keep the native Blender splat or export PLY, SOG or SPZ. The export bakes in the view transform, curves, white balance and per-pixel color effects. SOG is tagged for antialiased rendering, and a viewer with linear tone mapping can show the baked look.
Pilot Studio timed the Classroom with Cycles on an RTX 4070, including the renders:
▪️ Draft · about 2 min for 300K splats
▪️ Standard · about 19 min for 1M splats
▪️ High · about 1 hour for 2.5M splats
That’s one frame of the scene. For an interior walkthrough or a portfolio piece, try the browser demo’s comparison slider and then move the camera. Those two views show how closely it matches the render and what happens when you move away from that angle.
The add-on runs offline on Windows and Linux with a Vulkan GPU. Superhive lists it at $19 for one person or $42 for up to five studio users.
splatpilot.com/demo/
superhivemarket.com/products…
#Blender3D #GaussianSplatting #ArchViz #TechnicalArt
1
3
116









