单张pro 6000都得触发大量内存搬运

int8版本已出，体积减半

4060都能玩，544*800的分辨率，5秒视频，消耗10分钟，这样算下来都比我wan2.2快了

4090 24GB加128GB RAM, 加載66.3GB BF16模型, 跑480P 10秒 I2V, 用時440秒, 配合加速NODE, 用時200秒, 再加上Sage Attention, 最後用時160秒. 完全沒問題. 這次的H3模型, 只要RAM足夠, 就能順利跑起來.

I disagree with you, but only in ONE slight, minor area.

It's not "less" censored. It is COMPLETELY uncensored in my testing. Completely, TOTALLY uncensored. It doesn't even make an ATTEMPT at censorship. This might be the most fun, but dangerous model ever released open source lmao

It's like the power of sora 2 but it can do anything now

@CapAndABull After more tests and seeing some of the outputs here, I'd definitely agree. Using the model feels like peering into the internet, for better or worse.

@CapAndABull I know. Let's hope the governments step in and neuter all future AI models. Honestly, with the new regulations popping up, I've become rather cynical lately in that regard. I don't need to be protected from this any more I need to be protected from a hammer that someone might misuse. AI has really opened up entire new worlds of fun, you can just crank your hog to things previously thought impossible. It's all rather exciting, but a lot of people are working hard to take that away from us. It seems like the Chinese ability to simply give zero fucks is what's keeping this space interesting right now.

Watch as Flux3 undergoes "stringent safety tests and further safety alignment" before release. I feel like that's what's gonna happen because they missed the release window. EU AI regulations came into effect literally on the day H3 was released. The bullet essentially went right past their noses.

@makiaeveli I will agree with this though. But since you said that, CivitAI has cleaned up a bunch. Like some of the people are really straddling a line in the gallery below, why even do that in a public forum like this one?

@OhNoMyChickens Well there are several reasons, one you already said: "Let's hope the governments step in and neuter all future AI models."

Anyone can use the internet. People who do so blindly look the same to us as people who use it for evil. Any post could be from any one. Just report and move on. Red exists for a reason, to keep those posts from getting to too many people, and getting everyone in trouble. Things have been relaxing out... but people should still be aware of the topics they should avoid posting.

To your other point, I feel like it's why they dropped the weights. Feels a little like Opium Wars, can't put the genie back in the bottle. In some ways, though, it's almost equally unethical to have AI models be gated to only the wealthy and powerful. It would be yet another thing on the list of items in Western countries you simply pay to achieve. The weights are ours, they are us. For better, or worse. I'd much rather see governments run their own CLIPs or even finetunes, using their own guidelines, as open and reliable and free weights -- than have them enforce guidelines randomly on "everyone". Because we all know who those rules will be enforced on.

@makiaeveli I was of course joking about the government part. If anything, we have a bit of a nanny-state epidemic on our hands with governments that insist on "protecting people" and wanting to impose censorship, age and ID controls everywhere for "our safety" plus stringent AI regulations with hefty fines for companies that just make models deemed "too powerful", i.e. they're able to gen sensitive material and such. Not to mention the payment processor monopoly imposing their moral code on companies, forcing them to comply or shutter their business. Then we have the UK with their recent porn laws, Canada with some upcoming one and so on. Thoughtcrime is becoming a very real thing. Personal responsibility and freedoms are being restricted in the name of safety, and also to some degree, in the name of laziness. Because nobody wants to parent their kids anymore. That is of course another discussion entirely. In its current state, the majority of the world is neither permitted to use, nor modify H3 in any way. So finetunes such as Sulphur that were done for LTX, may never happen. As uncensored as H3 currently is, you will hit a ceiling eventually and what then? Flux3? They based in Germany, for them its possibly even worse when it comes to all of this. Why would they be any better license-wise? Content-wise? There are no guarantees anymore.

Hopefully things will clear up at some point soon and the license too. So that these models can be employed in the right spaces with the right guardrails, then developed further by the community. Its safe to say that nobody uses the majority of models in their stock form, the community rallies behind models and transform them into something entirely new.

Its such a sad state of affairs. Im just here to have fun and goon to the impossible, Ill readily admit that.

It's almost completely identical to prohibition mindset. History repeats. Baking weights on a GPU, fermenting booze in a can -- it's a natural phenomena that one can't simply stop.

@makiaeveli Pretty much. Hopefully it'll turn around soon, we should be living in a golden age of personal expression, freedoms and bodily autonomy. But we do not and are regressing as a society on many many fronts. For local models, I fear we may have peaked with H3 until China once again releases something into the wild. But I hope Im wrong and just needlessly defeatist right now. I will say that I do agree with some of the regulations, such as requiring watermarks on videos and images done on AI websites that are deemed "realistic" to stem the flow of deepfakes and misinformation. But many of the other points? No, its just going to lead to much worse products for the average Joe.

@makiaeveli Speaking of, flux3 can now be tested by all through their API. However, the open weights release (which comes out later, "soon" it says) was mentioned to be a "Flux3 dev open-weights VARIANT". Eh, make of that what you will.

@makiaeveli The last time a model impressed me this much was wan 2.2. And this is on a whole nother level. This is probably the most amazed I have been since way back when I got SD 1.5 and had to get over the sheer amazement of being able to generate images LOL

well it has been days, wonder what they are doing...

@RoRue read the licence, won't happen

back up is the safe word.

@SNEHCORE pomegranate*