Now block & catch punches also works for bare knuckles✊ Compared to boxer with thick arm and big gloves, defense with thin arm and bare knuckles will physically more likely get penetrated

Testing some defense moves, now we can catch or block jab, cross, hook, uppercut from either side to body or head by simply move the hand to a desirable position and hold it there.

Was planned to make Uppercut looks cooler for the past few hours, in the end this is how it turned out. Is just one of those days I guess🫠

Working on Block & Parry🛡️🥊 Seems catch the jab & cross is physically working as expected with a very intuitive control design (move hand fast forward is punch pose, other wise default to defense pose at current hand position) With this design, you can even use long guard naturally without any additional setup, as you can see in the second part of the video (well, not quite there yet, but we're getting there)

Testing some simple slip & roll counters, with "mouse movement = head & hand movement" plus current basic NPC is too easy to do

Ladies and Gentlemen, we got our first fight🥊 It took me 8 months of work to reach this point, next would be improving defense & footwork, then work on damage & stamina system, at this pace we'll have the first playable tech demo within next 4 months

Testing the Feel of Punching Power 👊 (with sound🔊) After worked on varies parts of system relates to physics & motion for the past 2 weeks, finally, seems all different type of punches starting to feel right, it's good enough for now, I need to start work on boxer NPC & other motion improvements so we can finally fight some! I also tested it on a big round punching bag, which is physically more friendly for uppercut, check here is you want to see it: youtube.com/watch?v=N1e190uG…

Work on punching bag's surface deformation effect when it get punched as @HyperMadGames suggested, somehow accidentally create this bug/effect, figure it look kinda cool, so I'll just quickly record it before fix it😎

Because the whole motion system is fully procedural, I can add a new boxer within a day, and things will work with bare minimal parameters turning, the end goal would be any player can add any humanoid character they want within a few hours. Physical Motions Fluidity & Impact are not as nearly ass good as the what it can be, I'll keep improving them while working on first iteration of NPC system #gamedev #indiegame #boxing

🥊Physics-based fully procedural boxing motion system that would work with any compatible humanoid character. Video below is a test for one of the character with plenty physical parts.🐺 There are still plenty or room for physical motion & interaction improvements, and more importantly, since now my boxing simulation system can easily adding new boxers into it, I can finally start to work on Boxer NPC system and then we can see some simulated boxing matches🛎️ #gamedev #indiegame #boxing

For some of you who interested in a in depth video about the control methods: youtu.be/ozdQ36IJjaM @VVarrent @JulesMurot @fromdesertdev @V_DropEmOff @Rock_A_Bye_1990 @Den_Kochetov @michaelCuevas27 @LincolnMargison @stevencheng @TechPostsEU @CallMeWhatNot

🥊Same physics-based procedural motion system, two distinct styles simply base on how you move your boxer Figure I should record some structured movements this time instead random one like in last video; As this system keeps getting better, I'll be able to showcase more and more boxing styles more easily #gamedev #indiegame #boxing

🥊6 months into development ⚡️Now boxer's procedurally generated motion works with basic physics control setup, from this point forward things will gradually looking more and more exciting! Took quite some experiments to figure out a two way motion & physics system design, i.e., not only generated motion driven physics, but also physics determine the starting point for the generation of the next frame's target motion. 📜Next I plan to make a YT video to better showcase the core intention driven control method, once that is done, besides keep improving the core motion & physics system, I'll add first few male & female boxers, environments, and NPC after that, so we can finally have a fight! #gamedev #indiegame #boxing

🥊5 months into development Motion fluidity still not quit there yet, but at least head movement, hands movement and footwork are all procedurally generated by the system in real time according to player input, and they are all works together now, as you can see in the video below when control our guy with keyboard & mouse Next, will add physics component, so this guy can finally hit something #gamedev #indiegame #boxing

🎯Player will have Precise & Separate control over character's Head & Hands motion. Head and hands position can be controlled separately, while any pose at any given moment would still looks natural, can't achieve this with pre-made animations or blend a set of pre-made poses by curve. #gamedev #indiegame #boxing

🥊2.5 months into development, finally achieved precise head & hands position trajectory control from a mouse alone 🎮This boxing simulator is designed as the polar opposite of how traditional boxing/fighting game is usually about, i.e., instead of mapping the pre-made animation directly to input buttons, in case using a mouse, you simply hold left/right mouse button and move it to move the hand, and release the button to throw the punch at certain direction, speed & distance (literally throw a punch), head's movement will simply follow the exact mouse movement with a different sensitivity, so you only need think about What you want to move to Where, and my system will figure out How by procedurally generates fluent animation according to your intention! Complete control while no 2 punches will every be the same! (Check the inputs in the video below) 🚧Now that baseline head & hand movements system that enables precise head & hands position trajectory control is done, next will be physics & secondary motions to make the characters motion looks more lively and can interact with other physical objects #gamedev #indiegame

Rubber boxer🥊when you testing different physics parameters😅 #gamedev #indiegame #boxing

To me, life is all about creating something truly novel, something no human has ever created before! 🥊Video below describes a novel control method design for a new boxing simulator, it's the exact opposite of what has been done in any typical fighting game. PS: 2 months into making this boxing simulator now, so far it was more challenging than I originally imagined, but still managed to stay on track somehow, I really should post a bit more often about the development process though

Jab, Cross, Uppercut will also follow the exact player inputs on top of exact head position control, still some work here before working on check/rare Hooks 🥊Punching animations is entirely generated from player inputs on top of entirely generated head movement animations, ZERO pre-made animations is the presistent theme here #gamedev #Boxing

Head Movement System Testing 🥊Head position is controlled by the Mouse input exactly, and whole body animations are entirely generated on the fly by the system #gamedev

👒My One Piece fan game: "One Piece: The Sky Runner" is completed! Let's just say "Nothing Happened" if you know what I meant😎 ▶️See video below for 3 minutes full gameplay 🎮You can also play it here yourself: mrforexample.itch.io/one-pie… @strawhats @onepiece_games @onepiecedaiIys

👒PS: This is a One-Piece fan project that is in development, what you saw here is first 40s of the gameplays (of total about 3 minutes) 💡About a month ago, when I was watching One Piece, I had this idea of designing a game mechanism that allow player to control Sanji to running freely in the sky, since I have some free time now, I figure why not just build it myself. 🚧The main game play is finished; I'm working on a parody 3d animation ending scene and it's almost done, probably going to take another week or two, I'll do my best utilize my free time, man, almost 2026 now, game dev (I call it the Ultimate Art) is still sooooo~ time consuming and hard to get it right. #ONEPIECE #fanart #gamedev @onepiece_games @onepiecedaiIys

#ONEPIECE #fanart #gamedev 🌩️Skywalk!!! 👒PS: This is a One Piece fan project that is in development, what you saw here is just a control mechanism testing video 💡About 2 weeks ago, when I was watching One Piece, I had this idea of designing a game mechanism that allow player to control Sanji to running freely in the sky, since I have some free time now, I figure why not just build it myself. 🚧I'll upload this mini console game for all to play after 2-3 more weeks once it's done, main game mechanism & assets processing is mostly finished; I'll build the actual gameplay & a few 3d animation scenes next. @onepiecedaiIys @strawhats @DailyOPMemes @onepiece_games

Get a feel of the true size of a 1000 X 1000 X 10 matrix, i.e. 10 million Sprites rendered with PointsNodeMaterial Seems this starting to hit the limit of what native three.js plus WebGPU can handle, but is good enough for what I'm trying to build🚀

Experimenting some WebGPU shaders in three.js, seems interactively visualizing matrix with billions of parameters is possible in browser on a laptop with RTX 4090 mobile GPU

🚀I just completed a Diffusion Autoencoder research project after more than a month! 🎯The purpose of this project is to prove this architecture can work well among s-o-t-a VAE models and offers a strong & stable codebase for other VAE researchers to build upon. 🛠️It integrates a deep compression encoder for high-ratio spatial reduction, a single-step diffusion decoder for fast reconstruction, and equilibrium matching for stable generative training, DC-SSDAE achieves compact latent representations while preserving perceptual quality 🎁Entire Codebase: github.com/MrForExample/DC_S… 📜Technical Report: mr-for-example.medium.com/dc… 🏋️‍♀️Experiments Weights & Logs: huggingface.co/MrForExample/… Researching Autoencoder/Representation learning is very difficult and largely underappreciated, hopefully this can at least make some Autoencoder researchers' life easier🍻🤜🤛 @du_yilun @songhan_mit @hancai_hm @EmergentMind @sainingxie @felix_m_krause @Den_Kochetov @camenduru @neutanent @toyxyz3 @DylanTFWang

Tested ShinkaEvolve from @SakanaAILabs ▶️Video below showcased top few programs it found & their generated results when I let it to discover novel ASCII arts that represent the concepts in AI 💸Last run takes about 20 generations and cost about $20 and 2 hours to finish on a powerful laptop I used claude-sonnet-4-5-20250929 for generation & gpt-5-mini for evaluation (It's 2~3x more expensive than using gpt-5-fast models only but this setup gives better results & less API error from OpenAI servers) 🚀There are so many components in ShinkaEvolve system that we can improve to truly unleash its creative power @hardmaru @Den_Kochetov @camenduru @neutanent

☄️Introducing Representation Emerging Mechanism (REM) for Continual Unsupervised Representation Learning REM is a type of Utility-Novelty based Self-Organizing Neural Network (NOT SOMs) that learn well factored representations from input data distribution/stream continuously, locally (Hebbian-like local learning rule without Gradient Backprop), reliably in an unsupervised manner. The reason of this research & researches that will come after this is my realization of Human-level Artificial Fluid Intelligence (or AGI if you may) can't be achieved without a Self-Organizing Neural Network where adaptation decisions are made at the level of neuron and even group of synapses (Dendritic compartments) Metaphorically speaking, the difference between standard gradient decent with backpropagation and REM is like the difference between: - Central Planned Economy versus Free Market Economy - Totalitarianism versus Individualism, where objective/loss function is the tyrant and approximated equity (Homeostasis) among neurons can be achieve by each neuron strive to be more useful in the economy rather than using additional constraint objective (since it's been showed time and time again in the history of socialism/communism forced equity only produce disaster) The main idea is to treat each neuron like an independent agent that are compete with other neurons and trying to be uniquely useful in some way, all the neurons work together to increase the overall utility of the whole network. Two main components are: - Utility Maximization via Kernel Specialization: To dynamically select a few neurons to learn input data according to density of the pattern populations and enable specialization of neurons base on their utilities; We want a neuron to become more special (general) for as long as it’s becoming more (less) useful. - Long Memory Retention via Concentrated Pattern Repetition: The pattern learned by a specialized neuron that are highly useful over a long period of time will stay relatively unchanged for much longer even if data distribution changed drastically, in contrast, the neuron that are at bottom regards utility will pivot to learn other useful patterns that haven't been captured by other neurons yet. The video below is to showcase the first few experiments results using small networks and MNIST dataset. Next, I will scale the system up and compare it with S-O-T-A VAE & DCNN models on large dataset like ImageNet. I want to make a point using REM and future researches that to achieve AGI (whatever your definition may be), we need to focus on find better ways to build better knowledge representations inside model first and NOT trying to chasing the objective score on benchmark (Goodhart's Law come on now), i.e., aiming to create system that can acquire, understand and come up with new knowledge/skills rather than just find ways to instill our existing knowledge/skills into the model and then find better ways to search for them... REM still plenty of rooms for improvements, the research paper and source code will be released within next two months once it's finished @kenneth0stanley @fchollet @LearningLukeD @MiTiBennett @liuzhuang1234 @Den_Kochetov @camenduru @neutanent

Definition of a piece of trash

For those who are interested, here is the paper's main author Kenneth Stanley and Akash Kumar discussing this paper and relevant topics in MLStreetTalk podcast: youtube.com/watch?v=KKUKikuV… @kenneth0stanley @MLStreetTalk

Comfy3D Update: (v0.1.6) - Integrated PartCrafter (Both object & scene mode) - A bunch of UX fixing ML Devs: @Den_Kochetov @camenduru github.com/MrForExample/Comf… @taziku_co @ZHO_ZHO_ZHO @toyxyz3 @RobinJHuang @ComfyUI @neutanent

Comfy3D Update: (v0.1.5.2.beta, dev branch)🗿 - Integrated Hunyuan3D-2.1 (Both Shape & Texture gen pipeline) - Some update on auto-build & install pipeline (WIP, will merge to main branch once install issue in ComfyUI-Manager is solved) Dev: @Den_Kochetov ‼️@GoNeuralAI We are hiring skilled ML researchers & engineers (VAE, diffusion model, 3D/texture gen/editing) to join us on a mission to revolutionize 3D virtual production 🫰If you want to work on open-source projects (Not only Comfy3D), make an impact in the industry all while getting paid handsomely, hit me up in DM! (English | 中文 | 日本語)🆗 github.com/MrForExample/Comf… @taziku_co @ZHO_ZHO_ZHO @toyxyz3 @RobinJHuang @ComfyUI @TencentHunyuan @neutanent

TG2MV is added & Noisy background issue is fixed. @huanngzh Thanks for the awesome models, it's quite big for vram though, hopefully in the future you & your team can create some models with similar capacity but smaller in size for the GPU with < 16GB vram to run🚀

Stable3DGen pipeline

Multiview to Texture projection pipeline from MV-Adaptor

Text + Mesh to MVs with MV-Adaptor

Comfy3D Update: (v0.1.5.1.beta, dev branch) - Integrated MV-Adapter: IG2MV, T2MV & Texturing pipeline - Integrated Stable3DGen, an image to detailed 3d model based on Trellis model ‼️We @GoNeuralAI are STILL hiring skilled ML researchers & engineers (VAE, diffusion model, 3D/texture gen/editing) to join us on a mission to revolutionize 3D virtual production 🫰If you want to work on open-source projects (Not only Comfy3D), make an impact in the industry all while getting paid handsomely, hit me up in DM! (English | 中文 | 日本語)🆗 This update is mostly done by @Den_Kochetov with my instructions, we'll focus on massive update on auto build & install pipeline next to enable smooth install experience on both Windows & Linux github.com/MrForExample/Comf… @taziku_co @ZHO_ZHO_ZHO @toyxyz3 @RobinJHuang @ComfyUI @huanngzh @neutanent

Some more workflows testing: (You can try other Hunyun3D-2 models in Example workflow folder)

Comfy3D Update: (v0.1.5.beta, dev branch) by our newly joined ML engineer: @Den_Kochetov⚡️ - Updated Hunyuan3D-2: * Hunyuan3D-DiT-v2-0-Fast * Hunyuan3D-2mini * Hunyuan3D-2mv * Hunyuan3D-2-Turbo * Hunyuan3D-2mini-Turbo and FlashVDM * Hunyuan3D-Paint-v2-0-Turbo ‼️@GoNeuralAI We are hiring skilled ML researchers & engineers (VAE, diffusion model, 3D/texture gen/editing) to join us on a mission to revolutionize 3D virtual production 🤌If you want to work on open-source projects (Not only Comfy3D), make an impact in the industry all while getting paid handsomely, hit me up in DM! (English | 中文 | 日本語)🆗 github.com/MrForExample/Comf… @taziku_co @ZHO_ZHO_ZHO @toyxyz3 @RobinJHuang @ComfyUI @TencentHunyuan @neutanent

I see, so it's same guys behind those kind of censorship again, just saw they got called into court for monopolistic behaviors: youtube.com/watch?v=I4P7LHcu…

TripoSG Scribble Model also works

Comfy3D Update: (v0.1.5.alpha, dev branch) - Integrated TripoSG (Plus scribble model) - MV-Adapter for high quality texture gen coming next ‼️@GoNeuralAI We are hiring skilled ML researchers & engineers (VAE, diffusion model, 3D/texture gen/editing) to join us on a mission to revolutionize 3D virtual production 🫰If you want to work on open-source projects (Not only Comfy3D), make an impact in the industry all while getting paid handsomely, hit me up in DM! (English | 中文 | 日本語 🆗) github.com/MrForExample/Comf… @taziku_co @ZHO_ZHO_ZHO @toyxyz3 @tripoai @RobinJHuang @ComfyUI @VastAIResearch @yanpei_cao @neutanent

📸FlexiSyncMVD - Generate 2k texture for a 3D object from a text or image prompt in 2 minutes (RTX 4090 GPU) FlexiSyncMVD is a Synchronized Diffusions-based optimization methods for generating RGB texture for given 3D mesh by leveraging pretrained image denoising model like SDXL. It's a research project I did at the end of last year, it works much faster than most optimization methods out there, but not as good as I wanted, still, I decided to release it today so it may help other researchers who is working in the similar direction🍻 github.com/MrForExample/Flex… Let's Accelerate🚀 @neutanent @KimJaihoon @ZeqiangLai @ZHO_ZHO_ZHO @toyxyz3 @taziku_co @yanpei_cao

@magrawala Pretty cool research for Video Generation Next research direction would be: Instead of optimize GPU memory utility using a set of predetermined schedules, we can use one additional pass upfront to let the model determine top-k frames it should focus on, then construct a schedule on the fly accordingly (i.e., in a way like MOEs, or NSA from Deepseek) lllyasviel.github.io/frame_p…

👁️This is a prototype of the LM-based research tool I'm building now; I call it: Super Seer. (Heavily WIP, but I'll starting to show more once it's ready) The "big idea" behind it is: ⚡️Having a tool can enable you to See the big picture and gain foresights for any research goal, WITHOUT the need to reading through hundreds of research materials yourself (Thus the name) ❔"If you want to know the truth, then do your own research." We all heard that, but research is such a time & energy consuming process require a certain level of critical thinking ability and specific expertise which not many people have or willing to spend time to acquire... ⁉️So, what if you have a tool that can take your research goal/problem and looking through thousands of research papers for you, and using First Principle thinking to decompose each solution found into a tree of sub-goals/problems and key components, (i.e., to generate a directed goal-solution graph for you automatically), and then it will group the related solutions together to give us a knowledge map so we can literally see the big picture (If the same graph is crafted manually would take months if not years, just like the simple graph shown in the video, it read 20+ papers and analysed the 3 most relevant papers and decomposed it into the graph within 2 minutes, which if I were to create this graph myself, it would take at least a day) 🚀What I described above will just be my vision for this tool before I release it, more useful feature will be added in time, natural next steps like graph-based novel idea/hypothesis generation, experiment design. I'm building this tool like a valuable product which will keep improving indefinitely instead just another fancy looking research project that will be done, release and move on. 🤗And yeah, of course, I plan to open source it sometime later this year once it become useful enough. @taziku_co @toyxyz3 @ZHO_ZHO_ZHO @yanpei_cao @neutanent

Just went through bunch of GitHub issues and Pull Requests and did a small update for Hunyuan3D-2, you can check the newest release and all changes here: github.com/MrForExample/Comf…

Comfy3D Update: (v0.1.4.alpha) - Integrated Hunyuan3D-2 (I did bunch of work including rewrote its mesh processor from C++ to Python to make sure only one additional custom package is required) - Support newest Comfy UI v0.3.12, py12, cu124, torch 2.5.1, Windows 10/11 (Pre-built wheels are available) * I'll spent my free time in next few days to do some robustness update, check the issues and merge pull requests GitHub Link: github.com/MrForExample/Comf… #ComfyUI #Comfy3D @ComfyUI @TXhunyuan @neutanent @bdsqlsz @ZHO_ZHO_ZHO @toyxyz3

#Hunyuan @TXhunyuan @bdsqlsz @neutanent I just spent a few hours to make Hunyuan3D-2 works in Windows. The time I spent are to make sure I won't adding more messy dependencies into Comfy3D😅 * I rewrote its mesh processor from C++ to Python so it doesn't depend on pybind11 to compile C++ per environment * Bunch of change and fix to make sure it also works in Windows Check pull request here: github.com/Tencent/Hunyuan3D… I'll build wheels for Windows 10/11, Cuda 12.4, Python3.12, then add it into ComfyUI tomorrow👌