[HN Gopher] Hunyuan3D-2-Turbo: fast high-quality shape generatio...
___________________________________________________________________
Hunyuan3D-2-Turbo: fast high-quality shape generation in ~1s on a
4090
Author : dvrp
Score : 161 points
Date : 2025-03-20 01:58 UTC (21 hours ago)
(HTM) web link (github.com)
(TXT) w3m dump (github.com)
| dvrp wrote:
| See also: https://github.com/Tencent/FlashVDM
| Flux159 wrote:
| I think the link should be updated to this since it's currently
| just pointing to a git commit.
| dvrp wrote:
| Reason for that is because of the dedup filter was thinking
| that this release is the same as the one that happened
| yesterday. Besides, the Flash release is only one of many.
| leshokunin wrote:
| Can we see meshes, exports in common apps as examples?
|
| This looks better than the other one on the front page rn
| dvrp wrote:
| Agree. That's why I posted it; I was surprised people were
| sleeping on this. But it's because they posted something
| yesterday and so the link dedup logic ignored this. This is why
| I linked to the commit instead.
|
| There are meshes examples on the Github. I'll toy around with
| it.
| llm_nerd wrote:
| Generate some of your own meshes and drop them in Blender.
|
| https://huggingface.co/spaces/tencent/Hunyuan3D-2
|
| The meshes are very face-rich, and unfortunately do not reduce
| well in any current tool [1]. A skilled Blender user can
| quickly generate better meshes with a small fraction of the
| vertices. However if you don't care about that, or if you're
| just using it for brainstorming starter models it can be super
| useful.
|
| [1] A massive improvement in the space will be AI or
| algorithmic tools which can decimate models better than the
| current crop. Often thousands of vertices can be reduced to a
| fraction with no appreciable impact in quality, but current
| tools can't do this.
| boppo1 wrote:
| Can it run on a 4080 but slower, or is the vram a limitation?
| dvrp wrote:
| They don't mention that and I don't have one -- can you try for
| yourself and let us know? I think you can get it from
| Huggingface or GH @ https://github.com/Tencent/Hunyuan3D-2
| fancyfredbot wrote:
| They mention "It takes 6 GB VRAM for shape generation and
| 24.5 GB for shape and texture generation in total."
|
| So based on this your 4080 can do shape but not texture
| generation.
| boppo1 wrote:
| Nnice, that's all i needed anyway.
| thot_experiment wrote:
| almost certainly, i haven't tried the most recent models but i
| have used hy3d2 and hy3d2-fast a lot and they're quite light to
| inference. You're gonna spend more time decoding the latent
| than you will inferencing. Takes about 6gb vram on my machine,
| I can't imagine these will be heavier.
| llm_nerd wrote:
| It can run on a 4080 if you divide and conquer. I just ran a
| set on my 3060 (12GB), although I have my own script which does
| each step separately as each stage uses from 6 - 12GB of VRAM.
|
| -loads the diffusion model to go from text to an image and then
| generate a varied series of images based upon my text. One of
| the most powerful features of this tool, in my opinion, is text
| to mesh, and to do this it uses a variant of Stable Diffusion
| to create 2D images as a starting point, then returning to the
| image to mesh pipeline. If you already have an image this part
| obviously isn't necessary.
|
| -frees the diffusion model from memory.
|
| Then for each image I-
|
| -load the image to mesh model, which takes approximately 12GB
| of VRAM. Generate a mesh
|
| -free the image to mesh model
|
| -load the mesh + image to textured mesh model. Texture the mesh
|
| -free the mesh + image to textured mesh model
|
| It adds a lot of I/O between each stage, but with super fast
| SSDs it just isn't a big problem.
| llm_nerd wrote:
| Just as one humorous aside, if you use the text to mesh
| pipeline, as mentioned the first stage is simply a call to a
| presumably fine-tuned variant of stable diffusion with your
| text and the following prompts (translated from Simplified
| Chinese)-
|
| Positive: "White background, 3D style, best quality"
|
| Negative: "text, closeup, cropped, out of frame, worst
| quality, low quality, JPEG artifacts, PGLY, duplicate,
| morbid, mutilated, extra fingers, mutated hands, bad hands,
| bad face, mutation, deformed, blurry, dehydrated, bad
| anatomy, bad proportions, extra limbs, cloned face,
| disfigured, gross proportions, malformed limbs, missing arms,
| missing legs, extra arms, extra legs, fused fingers, too many
| fingers, long neck"
|
| Thought that was funny.
| fixprix wrote:
| I recently got into creating avatars for VR and have used AI to
| learn Unity/Blender so ridiculously fast, like just a couple
| weeks I've been at it now. All the major models can answer
| basically any question. I can paste in screenshots of what I'm
| working on and questions and it will tell me step by step what to
| do. I'll ask it what particular settings mean, there are so many
| settings in 3d programs; it'll explain them all and suggest
| defaults. You can literally give Gemini UV maps and it'll
| generate textures for you, or this for 3d models. It feels like
| the jump before/after stack overflow.
|
| The game Myst is all about this magical writing script that
| allowed people to write entire worlds in books. That's where it
| feels like this is all going. Unity/Blender/Photoshop/etc.. is
| ripe for putting a LLM over the entire UI and exposing the APIs
| to it.
| anonzzzies wrote:
| You tried sharing your screen with Gemini intead of
| screenshots? I found it sometimes is really brilliant and
| sometimes terrible. It's mostly a win really.
| baq wrote:
| Look up blender and unity MCP videos. It's working today.
| fixprix wrote:
| Watching a video on it now, thanks!
| ForTheKidz wrote:
| > The game Myst is all about this magical writing script that
| allowed people to write entire worlds in books. That's where it
| feels like this is all going. Unity/Blender/Photoshop/etc.. is
| ripe for putting a LLM over the entire UI and exposing the APIs
| to it.
|
| This is probably the first pitch for using AI as leverage
| that's actually connected with me. I don't want to write my own
| movie (sounds fucking _miserable_ ), but I do want to watch
| yours!
| iaw wrote:
| I have this system 80% done for novels on my machine at home.
|
| It is _terrifyingly_ good at writing. I expected Freshmen
| college level but it 's actually close to professional in
| terms of prose.
|
| The plan is maybe transition into children's books then
| children shows made with AI catered to a particular child at
| a particular phase of development (Bluey talks to your kid
| about making sure to pick up their toys)
| thisisnotauser wrote:
| I think there's a big question in there about AI that
| breaks a lot of my preexisting worldviews about how
| economics works: if anyone can do this at home, who are you
| going to sell it to?
|
| Maybe today only a few people can do this, but five years
| from now? Ten? What sucker would pay for any TV shows or
| books or video games or anything if there's a comfy UI
| workflow or whatever I can download for free to make my
| own?
| CamperBob2 wrote:
| _What sucker would pay for any TV shows or books or video
| games or anything if there 's a comfy UI workflow or
| whatever I can download for free to make my own?_
|
| I think it's about time the industry faced that risk.
| They have it coming in spades.
|
| For example, LOST wouldn't have been such a galactic
| waste of time if I could have asked an AI to rewrite the
| last half of the series. Current-generation AI is almost
| sufficient to do a better job than the actual writers, as
| far as the screenplay itself is concerned, and eventually
| the technology will be able to render what it writes.
|
| Call it... _severance._
| numpad0 wrote:
| Only few are both capable and willing to take on creative
| tasks, with AIs or not. Boring people cannot form strong
| enough cohesive thoughts that can drive an AI, even if AI
| output itself were not as boring as they are.
| fixprix wrote:
| It breaks economics in a good way. Less resources spent
| on all kinds of media and other things is deflationary.
| Prices go down, a single person can provide for a family
| working less hours.
| yfw wrote:
| How is this good in any way for the creative workers? Do
| you think there's a sustainable source of innovative and
| interesting experiences being generated if no people
| wrote anymore?
| fixprix wrote:
| It's good for everybody. We have to work less to survive.
| Everyone is more productive at whatever they do. The
| value is in the ideas not the medium. You might write a
| book and I'll take it and use AI to turn it into audio,
| or a tv series, or a movie, or a video game.
|
| The bottleneck is no longer on labor to turn ideas into
| reality, the bottleneck is imagination itself. It's
| incredible. The cost to produce/consume going down along
| with many other facets of the economy translates into
| deflation.
|
| If you make less money, or work less hours, or only have
| a single person in your family work - that's ok because
| money will go further. That's the whole idea behind Star
| Trek, the first step though was intelligent computers,
| automation and robots. Harnessed in a way that doesn't
| backfire on us.
| YurgenJurgensen wrote:
| The 'professional level' prose to which you refer:
| "ABSOLUTE PRIORITY: TOTAL, COMPLETE, AND ABSOLUTE QUANTUM
| TOTAL ULTIMATE BEYOND INFINITY QUANTUM SUPREME LEGAL AND
| FINANCIAL NUCLEAR ACCOUNTABILITY"
|
| Even if AI prose weren't shockingly dull, these models all
| go completely insane long before they reach novel length.
| Anthropic are doing a good job embarrassing themselves at
| an easy bug-catching game for barely-literate 8-year olds
| as we speak, and the model's grip on reality is basically
| gone at this point, even with a second LLM trying to keep
| it on track. And even before they get to the 'insanity'
| stage, their writing inevitably experiences regression
| towards the average of all writing styles regardless of the
| prompt, so there's not much 'prompt engineering' you can do
| to fix this.
| yfw wrote:
| The value of art is that it's a human creation and a product
| of human expression. The movie you generate from AI is at
| best content.
| mclau156 wrote:
| I have never seen knowledge to be the limiting factor in
| success in the 3D world, its usually lots of dedicated time to
| model, rig, and animate
| iamjackg wrote:
| It's often the limiting factor to getting started, though.
| Idiosyncratic interfaces and control methods make it really
| tedious to start learning from scratch.
| spookie wrote:
| I don't think they are idiosyncratic. They are built for
| purpose, one simply lacks what to look for. Same for
| programming really.
|
| I also think that using AI would only lengthen the learning
| period. It will get some kind of results faster, though.
| spookie wrote:
| If you need time dedicated to it, knowledge is the limiting
| factor.
| tempaccount420 wrote:
| > Unity/Blender/Photoshop/etc.. is ripe for putting a LLM over
| the entire UI and exposing the APIs to it.
|
| This is what Windows Copilot should have been!
| fixprix wrote:
| I'm sure they're working on it. This MCP stuff is early days.
| Even I am just finding out about it's integration into
| Blender and Unity in this thread.
| sinzin91 wrote:
| You should check out Blender MCP, which allows you to connect
| Claude Desktop/Cursor/etc to Blender as a tool. Still early
| days from my experiments but shows where it could go
| https://github.com/ahujasid/blender-mcp
| sruc wrote:
| Nice model, but strange license. You are not allowed to use it in
| EU, UK, and South Korea.
|
| "Territory" shall mean the worldwide territory, excluding the
| territory of the European Union, United Kingdom and South Korea.
|
| You agree not to use Tencent Hunyuan 3D 2.0 or Model Derivatives:
| 1. Outside the Territory;
| Eduard wrote:
| source:
| https://github.com/Tencent/Hunyuan3D-2/blob/main/LICENSE
| johaugum wrote:
| Meta's Llama models (and likely many others') have similar
| restrictions.
|
| Since they don't fully comply with EU AI regulations, Meta
| preemptively disallows their use in those regions to avoid
| legal complications:
|
| "With respect to any multimodal models included in Llama 3.2,
| the rights granted under Section 1(a) of the Llama 3.2
| Community License Agreement are not being granted to you if you
| are an individual domiciled in, or a company with a principal
| place of business in, the European Union. This restriction does
| not apply to end users of a product or service that
| incorporates any such multimodal models"
|
| https://github.com/meta-llama/llama-models/blob/main/models/...
| littlestymaar wrote:
| This is merely a "we don't take responsibility if this somehow
| violates EU rules around AI", it's not something they can
| enforce in any way.
|
| But even as such a strategy, I don't think that would hold if
| the Commission decided to fine Tencent for releasing that in
| case it violated the regulation.
|
| IMHO it's just the lawyers doing _something_ to please the boss
| who asked them to "solve the problem" (which they can 't,
| really).
| ForTheKidz wrote:
| Probably for domestic protection more than face value. Western
| licenses certainly have similar clauses to protect against
| liability for sanction violations. It's not like they can
| actually do much to prevent the EU from gaining from it.
|
| North Korea? Maybe. Uk? Who gives a shit
| lwansbrough wrote:
| How long before we start getting these rigged using AI too? I've
| seen a few of these 3D models so far but none that do rigging.
| halkony wrote:
| This is what I'm looking forward to the most, there's a lot of
| potential for virtual reality with these models.
| bufferoverflow wrote:
| https://anything.world/
| qingcharles wrote:
| This looks great!
| Y_Y wrote:
| How are they extracting value here? Is this just space-
| race-4-turbo propagandising?
|
| I see plenty of GitHub sites that are barely more than
| advertising, where some company tries to foss-wash their
| crapware, or tries to build a little text-colouring library that
| burrows into big projects as a sleeper dependency. But this isn't
| that.
|
| What's the long game for these companies?
| yowlingcat wrote:
| There's an old Joel Spolsky post that's evergreen about this
| strategy -- "commoditize your complement" [1]. I think it's
| done for the same reason Meta has made llama reasonably open --
| making it open ensures that a proprietary monopoly over AI
| doesn't threaten your business model, which is noteworthy when
| your business model might include aggregating tons of UGC and
| monetizing engagement over it. True, you may not be able to run
| the only "walled garden" around it anymore, but at least
| someone else can't raid your walled garden to make a new one
| that you can't resell anymore. That's the simplest strategic
| rationale I could give for it, but I can imagine deeper layers
| going beyond that.
|
| https://www.joelonsoftware.com/2002/06/12/strategy-letter-v/
| quitit wrote:
| Running my usual img2mesh tests on this.
|
| 1. It does a pretty good job, definitely a steady improvement
|
| 2. The demos are quite generous versus my own testing, however
| this type of cherry-picking isn't unusual.
|
| 3. The mesh is reasonably clean. There are still some areas of
| total mayhem (but these are easy to fix in clary modelling
| software.)
| amelius wrote:
| I don't understand why it is necessary to make it this fast.
| Philpax wrote:
| It helps with iteration - you can try out different concepts
| and variations quickly without having to wait, especially as
| you refine what you want and your understanding of what it's
| capable of.
|
| Also, in general, why not?
| amelius wrote:
| > Also, in general, why not?
|
| There are various reasons:
|
| - Premature optimization will take away flexibility, and will
| thus affect your ability to change the code later.
|
| - If you add features later that will affect performance,
| then since the users are used to the high performance, they
| might think your code is slow.
|
| - There are always a thousands things to work on, so why
| spend effort on things that users, at this point, don't care
| much about?
| andybak wrote:
| > users, at this point, don't care much about?
|
| What makes you think this is true?
| TeMPOraL wrote:
| Being _this fast_ is not a "premature optimization", it's
| a qualitatively different product category. ~immediate
| feedback vs. long wait time enables entirely different
| kinds of working.
|
| Also:
|
| > _since the users are used to the high performance, they
| might think your code is slow._
|
| I wouldn't worry about it in general - almost all software
| is ridiculously slow for the little it can do, and for the
| performance of machines it runs on, and it still gets used.
| Users have little choice anyway.
|
| In this specific case, if speed is makes it into a
| different product, then losing that speed makes the new
| thing... a different product.
|
| > _There are always a thousands things to work on, so why
| spend effort on things that users, at this point, don 't
| care much about?_
|
| It's R&D work, and it's not like they're selling it.
| Optimizing for speed and low resource usage is actually a
| good way to stop the big players from building moats around
| the technology, and to me, that seems like a big win for
| humanity.
| llm_nerd wrote:
| They released the original "slow" version several months
| ago. After understanding the problem space better they can
| now release the much, much faster variant. That is the
| complete opposite of premature optimization.
|
| Yes, of course people care about performance. Generating
| the mesh on a 3060 took 110+ seconds before, and now is
| about 1 second. And on early tests the quality is largely
| the same. I'd rather wait 1 second than 110 seconds,
| wouldn't you? And obviously this has an _enormous_ impact
| on the financials of operating this as a service.
| bufferoverflow wrote:
| Fast is always better than slow, if the quality isn't worse.
| awongh wrote:
| What's the best img2mesh model out there right now, regardless of
| processing requirements?
|
| Are any of them better or worse with mesh cleanliness? Thinking
| in terms of 3d printing....
| MITSardine wrote:
| From what I could tell of the Git repo (2min skimming), their
| model is generating a point cloud, and they're then applying
| non-ML meshing methods on that (marching cubes) to generate a
| surface mesh. So you could plug any point-cloud-to-surface-mesh
| software in there.
|
| I wondered initially how they managed to produce valid meshes
| robustly, but the answer is not to produce a mesh, which I
| think is wise!
| coolius wrote:
| has anyone tried to run this on apple silicon yet?
| postalrat wrote:
| That would be revolutionary.
| debbiedowner wrote:
| Has anyone tried it on a 3090?
___________________________________________________________________
(page generated 2025-03-20 23:01 UTC)