GaGXZ

@Nobody29195675 @Ratsu_10 AI

Chudai...🥰

@reoparu_PKM @Tenshi_Kata AI

@Tenshi_Kata who draw this?

@cloneofsimo My great idea is to use LPIPS (I know I'm just too smart).

Continuing the journey on f16-c32 AE.... so getting high frequency details is tough... I am trying my best to not use LPIPs/GAN tricks but damn these things are hard to get it right. Help me out with great ideas / literatures.... left is reconstructed, right is input.

@thealtk34012 @jimcrowdeacon @WholesomeOrenji It's never going to happen, the dataset of these models are well curated.

someone made an AI model of my art so I thought I might as well try it out

@madebyollin @askerlee @cloneofsimo But it doesn't seem to bring any negatives, it just feels "hacky". Have you read arxiv.org/abs/2401.00110?

@GaggiXZ @askerlee @cloneofsimo Using LPIPS feels fairly hacky to me - you'd rather your autoencoder learn interesting semantically-meaningful latent representations on its own (without relying on any pretrained models). I love the idea of learning whatever latent encoding lets your AR/DM decoder minimize loss.

Worked on this weekend: open-sourced f16-c32 VAE (will release tommorow or something, but its a quite large model lol) vibe checks out btw: left is ground truth, right is reconstructed. The trick was to use zero-init modulation (like DiT), groupnorm, latent upsamping, and finally again, muP for hparam search🤪🤪

@madebyollin @askerlee @cloneofsimo Why the choice of not using LPIPS?

@askerlee @cloneofsimo Yes, the latent diffusion VAE recipe uses joint encoder->decoder training with L1 + LPIPS + PatchGAN losses. You could easily train a new VAE decoder using pure diffusion or conditional GAN objectives, but training a good from-scratch VAE encoder without LPIPS is tricky

@cloneofsimo A YouTube tutorial can be made by anyone, focus on what's more valuable that you can provide.

I think I only have time for one Should I make youtube tutorial or write paper on training details of AuraFlow?

@Wester_Hare @Trans_Lykeia That meme was fake of course, it wasn't generated with "salmon in river", sorry to ruin it.

>capable of understanding and combining concepts Are you sure about that

@masoner789 @Rizdraws He figured it out because it's written in the bio but most people doesn't seem to read it anyway.

@Rizdraws If you’re wondering how he figured it out, it’s the ear, and part of the finger that’s edited out. There’s a couple of other things too but those are the easiest ones to see.

Every time I get fooled by AI generated "art" I get a liiiiittle bit more depressed

@cloneofsimo Such a model two years ago would have been incredible. Can you share an image of a face? A real one?

An abstract and surreal image of a cascading vase, filled with water that splashes out, creating a dreamy effect. A vibrant bouquet of tulips, in colors ranging from red, purple, blue, white, orange, and yellow, are arranged in a loose, natural cluster.

@kippytan You need to organize your sleep schedule, eating habits and take care of yourself and the environment you live in and your mood will improve, if not it's just a vicious circle.

Lately I don't have any energy, and I want to sleep all the time, I feel tired even if I didn't do anything, If I do a little, I have to lay down for awhile I don't know why. it happens every now and again but It lasts way too long

@dillonrpayton Aren't they already operating in Zaporizhzhia? Dnipro would be much harder to operate with than Kharkiv, it really doesn't make sense. Talk about yourself before accusing others to be insufferable ahah

@GaggiXZ You people will be insufferable at any cost to not admit Ukraine brings bad things upon itself. What about the south? Why do they have to do it to Kharkiv instead of Dnipro or Zaporizhzhia just as an example?

Did everyone memory hole the RVC incursions in Belgorod and the whole “Free Belgorod” movement? This and Ukrainian shelling of the oblast at large, it shouldn’t be controversial to say the current RU offensive and previous Ukrainian incursions have a positive causal relationship.

@dillonrpayton Well then as the Donbas is Russia’s primary objective (Kharkiv operation divert Ukrainian troops from the Donbas) and the fact that they only sent a token force to extend the frontline, it seems realistic to me that in another timeline they would have done the same operation.

@GaggiXZ Russia sent a token force. North Grouping is about 50,000 men while the SMO is reaching 600,000 all together.

@dillonrpayton Are you talking about the Ukrainian army in the last point? If not, I don't see how these could be counterarguments for Russia launching the same operation in a different timeline.

@GaggiXZ I’m not convinced. The Donbas seems to be Russia’s primary objective. Plus they sent a rather token force of bran new reserves for the region.

@dillonrpayton Whether it's Ukraine's fault or not depends mostly on whether you believe that Russia would have started this operation regardless of the existence of the RVC, if so, then there's not much to blame.

@GaggiXZ Can’t help but think you’re deflecting blame from Ukraine. We have the stated reason, it’s a good one politically for Russia, but yes it also helps Russia in the east.

@dillonrpayton I would argue that the secondary objective is actually the primary objective and would have happened anyway if it give Russia an edge in the Donbass.

@GaggiXZ Putin explicitly said the Kharkiv operation is to stop Ukrainian strikes and incursions in Belgorod. Again, people have memory holed this. I already stated the secondary objective of the operation is to pull Ukrainian troops from the main effort in the east

@amayomai @workBDK @MZactojj @j0ch3fvj6nd It's in the bio.

@workBDK @MZactojj @j0ch3fvj6nd the perspective, anatomy and shadows make too much sense for an AI art, you can even see the brush strokes. I believe it's tagged wrong, AI can't do things like that yet

猫イベ、10日目(?) #原神 #GenshinImpact

@multimodalart @model_mechanic There are some speculations about the existence of a LLM that would take your prompt as input and then generate bounding boxes with object labels, text etc

@GaggiXZ I agree with your assessment, but that text is somewhat in the same domain of what I requested, I'm not sure that under the hood it changed the prompt to request that sentence specifically. Maybe @model_mechanic can answer? :-)

Got DALL-E 3 via Bing, and there's a game-changer aspect that no one is talking about prompt: "the line to the first feijoada restaurant in Tokyo" 🫘🗼 but do you see the 2nd line? It almost reads "serving authentic brazil cuisine" That's mind blowing! 🤯 Yup, Imagen, IF, Ideogram can make text, but only what's on the prompt. Emergent, contextual generated text is a first

@multimodalart Also the model generates sometimes completely incoherent text like any other models, most probably when there is no condition by the LLM about that text.

@multimodalart By the images it generates it's clear that the model is told what text to generate as it sometimes misses words, letters or maybe too many letters, not something you would be able to see if the model actually developed deep capability in natural language and world knowledge.