general

[sullen quest]

but you told me the important info, which is that c is more common than g

and thus, people will recognize it more

[keen beacon]

And?

[sullen quest]

and faster

[keen beacon]

No, if a song is written originally in D, people will remember it in D.

[sullen quest]

yeah, and if they hear it in C, it will take a second longer for them to realize what the song is

there is no flaw

[keen beacon]

Reread the last two minutes of our conversation until you realize where is the flaw in your reasoning.

[sullen quest]

otaku, if reddit can't convince you, and I can't, go argue this with a llm

[keen beacon]

I'm done with you.

[sullen quest]

and if you can't give me the flaw, it doesn't exist

[leaden sun]

i got this sometimes on openrouter and claude.ai where 3.5 cosplays sonnet 4 and both opus, interestingly, the real opus if asked whether they are really the real ones, they usually respond philosophically with existence angst, just my observations, I usually end the conversation and open another one to try my luck meeting da real thing

typically, this situation would feel like "ordering champagne and being served champagne flavored sparkling water while the waiter insists it's champagne"...

[stiff glacier]

What is the ki for lmarena anonymous-915?

He doesn't say who trained him, nor does he feel like Gemini.

[wicked mason]

Sir. The new model has been reported to as chat gpt 5.5

[stiff glacier]

Don't fool me, that's not one of those models! 😂

[tardy wagon]

why was seedream4 removed??

[cloud zinc]

@tardy wagon nano banana better

[shy aspen]

A sad cartoon tooth facing the camera with his head broken by tooth decay. The cartoon tooth is sitting in one chair of the two lines of chairs in the reception room of a dental clinic. He is waiting for his turn. Next to him there is a door with the text Dr. Sonrisa on a rectangle on the door. The door knob is on the left of the door. One of the walls of the reception room has the text ZONA DENTAL in blue. Next to him there is another cartoon tooth holding his jaw with one hand as if he had a toothache and he was in pain. There are some other cartoon teeth sitting in the chairs waiting for their appointment. At the scene, there is a pretty and smiling female secretary with black hair behind a desk using a computer. There are two lines of chairs facing each other.

[keen beacon]

lmarena down?

[frank sundial]

Hola, Buenas noches

[robust yoke]

What kind of argument happened here while I was away?

That argument almost reminds me of the civil discussion I had with that one guy whose name I can't remember.

[burnt sinew]

how is 4o better than 5-high??

[robust yoke]

Likely because people liked the sycophancy it had.

And, on some problems, it even outperforms GPT-5.

[burnt sinew]

what's a better way to tell what models are actually better then

[robust yoke]

Hmm…

Personal use.

A personal, curated benchmark.

Try getting either one to code up something for you, or test their writing skills, or math skills.

Whatever test you come up with.

[devout cipher]

3 - Close-up of a yellow spider excavator's claw tearing through desert rock, slow-motion dust cloud, engineers filming from below, cinematic lighting
When machines become legends.
#MachineArt #MechanicalBeauty #EngineeringMagic #FutureIsNow #RoboticPower #ViralTikTok

[robust yoke]

Wrong channel.

You'll wanna go to #video-arena-1 for that.

[devout cipher]

Vídeo de canal

[snow wagon]

Can we generate 5 minutes video

[robust yoke]

Unfortunately, no.

The max is ten seconds, I believe.

[remote arrow]

8

[robust yoke]

Yeah, that.

Sorry to disappoint.

[snow wagon]

Ok which one is good video 1-3,,they are same or not

[robust yoke]

Hmm…

Depends on what you're talking about.

[remote arrow]

You can't choose which one you will use. It's random. If you purely ask for the best, I prefer Veo3 or Seedance..

[robust yoke]

True.

I was gonna say Veo 3.

[snow wagon]

How many videos can we generate in a day

[snow wagon]

Ok 👍

[remote arrow]

5. Each generation you'll get 2 videos. So they are 10 in total.

[remote arrow]

You might want to check my video here to get 3 minutes or longer:
#ai-creations message

[snow wagon]

Ok thanks,

[snow wagon]

Is that your YouTube channel

[remote arrow]

Yes

[snow wagon]

If you create short videos you will gain subscriber very fast

[robust yoke]

They could also create long-form videos and get views from that.

Like how-tos.

[remote arrow]

Never be one of my intention. I just love to dig AI.

[snow wagon]

Yeah but short video is good to get subscribers

[robust yoke]

However, it's also dangerous to be too famous.

There must always be a fair balance between the two.

[remote arrow]

AFAIK, shorts are mostly made in 9:16. I would prefer 16:9 in my works.

[robust yoke]

Which would work well for long-form videos.

Since those are usually in 16:9.

[snow wagon]

Guys we can create 5 times in a day and how can we use same character to get full story

[robust yoke]

So like one of those storytime shorts where someone comes across something disturbing, then requires a full deep-dive into the creator's channel to figure out the rest of the story?

[abstract sinew]

Daily question: Is seedream4 now functional on LMA

[remote arrow]

Don't use "text-to-video" but "image-to-video" instead. Each scene started by an image of your character.

[robust yoke]

Daily answer: no.

[abstract sinew]

Thank you

[robust yoke]

My pleasure.

[remote arrow]

Only in battle mode (random). And not hi-res. I'm using it now.

[robust yoke]

Yeh.

[remote arrow]

So the hunt of Soda Dream continues..

[robust yoke]

And so it seems…

[remote arrow]

Yet, they still don't know which one is really the best on LMArena..

[snow wagon]

Ok let me try😊👍

[robust yoke]

And they don't know that it's obviously Seedream 4.0.

[remote arrow]

No. It's Assistant A

[robust yoke]

And Assistant B too, it seems.

After all, they both look the same.

[remote arrow]

They got characteristic

[robust yoke]

In terms of quality, that is.

[robust yoke]

That's true.

[remote arrow]

Without revealing the names, Google's models can be recognized by better prompt reading. Banana and Imagen show similar behavior. Flux and Qwen seem using the very same basic model, plastic skins.

[robust yoke]

True.

However, I will say that even Imagen tends to have the plastic skin tendency.

[remote arrow]

Yes, but not as much as Flux or Qwen

[robust yoke]

Precisely.

[snow wagon]

I just send my script

[remote arrow]

We can really see the difference in close-up portrait

[robust yoke]

Yeah, that's true.

It seems to really like exaggerating the pores on people's skin.

[remote arrow]

In video arena?

[remote arrow]

Yes. And in that details, it's a real battle between Banani and Soda Dream..

[snow wagon]

Yeah just sent it but we need to wait 10 minutes

[robust yoke]

And, dare I say, GPT's image generator, because that one also likes to add the pores on people's skin.

Although, I will say that it's less of a contender due to the infamous “yellow filter”.

[remote arrow]

If, only if, this one have character consistency. I will choose this. It's detailed. Painterly, yet can do photographic.

[robust yoke]

Indeed.

I agree with that.

However, it doesn't seem to have a perfect understanding of character/environmental consistency, and thus, tends to hallucinate a completely different person/environment.

[remote arrow]

Seems like you just sent "text-to-image" and didn't tell the video generators how are your characters look-alike..

[remote arrow]

Yes. Even it can't read body weight correctly. Lol..

[robust yoke]

That too, heh heh.

[remote arrow]

Before banana, I tried over hundred times to fix old pictures in GPT-Image, ended with dozen new faces I never seen before..

[robust yoke]

True.

However, it is very good at creating original images.

[snow wagon]

Ohh 😅 how to use it

[robust yoke]

You gotta use “/image-to-video”, silly.

[remote arrow]

#1397655624103493813

[robust yoke]

There's two variations: “text-to-video” (the one that gets used by default via the “/video” command), and “/image-to-video”.

[snow wagon]

Oket me try

[robust yoke]

Let us know if you need any help, should it not work.

[remote arrow]

When you are in video-arena channels (1-3), simply type this symbol "/" and you'll see options.

[robust yoke]

So that means you missed a value.

[snow wagon]

Ok

[remote arrow]

Phew.. I never tried to generate from a phone..

[robust yoke]

I certainly have, and it's not actually that hard to do, given you know what you're doing, that is.

[snow wagon]

Ok I got it guys thanks for your help❤️

[remote arrow]

Show me the result later, Buddy..

[snow wagon]

Ok it will take 10 minutes I hope the result will good

[robust yoke]

Pro tip: if you ever want an easier way to use the command, you can click/tap the controller icon, then find the video arena bot and click/tap on it, then find the command you're looking for and click/tap on it. It'll open up a user-friendly UI that you can then use for your video creation needs.

[remote arrow]

That's approx time..

It's strangely peaceful here without prompt-walls..

[snow wagon]

I

How to do it guys

[robust yoke]

Just gotta be careful not to j-j-j-jinx it.

https://tenor.com/view/shaggy-scared-zoinks-surprised-gif-15872371

[robust yoke]

Just gotta wait it out.

It'll notify you in a DM when the video creations are done.

[snow wagon]

Ok let's see👍

[remote arrow]

Looks like it's done:
#video-arena-1 message

[remote arrow]

From what I see, your script is too long for 5-8 second videos. You need to break it down to separate short-scripts, each for 1 character in your script. Then combine them all into 1 video using an editor.

[remote arrow]

That's a nicely written script with timelines. But seems the video generators failed to read it correctly..

[snow wagon]

😭 ok let me do it again like you said

[remote arrow]

Note: Veo3 got 8 seconds length. The other mostly 5-6 seconds.

[snow wagon]

How can I choose veo 3

[remote arrow]

You can't do here. LMArena is for testing purpose. It's random to see which one got higher vote.

[robust yoke]

It's picked randomly.

[remote arrow]

I can send you links if you can access on PC. I don't use a phone.

[snow wagon]

Ok i sent again let's see

[hard quiver]

Did lmarena reduced the rate límit for nanobanana?

[snow wagon]

@remote arrow if the generate failed again I will stop using this ai😔

[remote arrow]

Dude, don't be so easy to be broken.
Check Youtube Apps on your phone, some info says it is now have Veo3 built-in to create shorts.

[snow wagon]

Again 😭

I will try image to video

[snow wagon]

Bro I'm not from supported country 😁

[toxic egret]

guys, theres any way to cancel a prompt? gpt-5 has been generating for the past 30 min

[remote arrow]

Can't absolutely rely on testing platform like LMArena for your creative ideas.

[remote arrow]

What's available in your country? Doubao or Douyin also have good video generator..

[remote arrow]

Refresh the browser or make a new chat.

[snow wagon]

I need free tool..tell me free ai that can generate high quality video

[toxic egret]

already tried refreshing and ill have to resend 10MB in lines of code, so please tell me theres another option

[snow wagon]

I'm from India

[remote arrow]

Wow. Can't help with that..

[remote arrow]

Google have some free services for Veo3..

[toxic egret]

guess im cooked, ill wait a bit to see if it starts answering

[remote arrow]

Hopefully it ends well..

[rancid yarrow]

You can also try out KLINGAI

[snow wagon]

Ok I will subscribe your channel and subscribe my channel too😊

[snow wagon]

Yeah bro I already tried this one

[remote arrow]

Fair. You can DM me for some links. Hopefully can help and available in your country.

[fading rover]

Good morning all. I just logged into lmarena web posted a prompt it gave me error three time " sorry something went wrong" 4th time it generated an image but wasnt really accurate and regeneration 5th time it told me i am out of my limit try fter some 50minutes. Is this common for all of you ?

[hearty ferry]

DO YOU GUYS HAVE ANY PLANS TO BRING HIGGSFIELD SOULD Model?

@echo aurora

[agile cosmos]

yo

[echo aurora]

For plans for new models you should keep an eye our #announcements channel as that's where we'd share new model updates.

[echo aurora]

Sry to hear the problems with the errors, is this happening to a specific model? Was it a new chat?

[fading rover]

That's banana

[hearty ferry]

even gemini is taking 100+ secs to generate and flux 1 kontext pro is giving wrong facial outputs (twisted eyes, no facial match)

[echo aurora]

Did you use the model a lot before logging in? was this the first time trying to generate today?

[fading rover]

Yes it's first time with login

Lately it's been throwing cache generation and repeated mistake or errors to generate .. and trying again does count errors and give the limit earnings so overall 4 failed attempt 1 proper output is the average ratio

[echo aurora]

Is this mobile?

[fading rover]

No I tried this on pc .. not yet went out so when I am out I have same login sync on mobile that I use outside but this is common

[echo aurora]

Thanks for the info, sorry to say I don't have a short-term solution for you. What I think happened is you got some errors (this can happen for different reasons), and those "generations" counted towards your rate limit. We are looking into preventing whena user gets an error as it counting towards rate limit. Also errors in general is something we'd like to prevent as much as possible.

[prime mulch]

Nowadays flux kontent pro give wrong output especially that not give exact face

[prime mulch]

Example

It was whole different person @echo aurora

[hearty ferry]

Yes

flux changing the face too

[prime mulch]

Bro highsfieled is not worth

[hearty ferry]

yeah

[echo aurora]

Was there a previous image added to that same chat before that generation?

[prime mulch]

No this is the whole chat i usually use new chat per image generation

[hearty ferry]

nope
blur images, eye twisted, different face

these are the issues

[echo aurora]

Okay can you make a post in #1343291835845578853 with the details? If you have any before and after examples that’d. be really helpful

[fading rover]

Yes it seems may be a gemini issue I don't blame any one it's just a routine latency increasing with limitation stopping creativity at all 😂

[fading rover]

Yes I seen nano bana model wich I use 99% time for my creations also tend to make soft cached images plenty of time

[prime mulch]

For face seedream 4 highres is best Nano banana usually change face details

[prime mulch]

https://discord.com/channels/1340554757349179412/1419546093578092618

[fading rover]

I had bad experience with face on seedream as it tend toimic wrong shaped eyes and also hdrish toning look really horrible form y tastein natural tone portraits..

[prime mulch]

Nah you need to focus on prompt

This is the reference face

These are the results

@fading rover

This is the result

@fading rover did you like the result

[fading rover]

Well on mobile u can't see really difference much but. Eyes are little off even in ur photos so when I create close portraits it does this a lot prominently not 100% times but on and off invariably it does

[prime mulch]

But it's better than nanobanana

[fading rover]

Yea probably didn't have gone far with it since it's not here on lmarena

[prime mulch]

Can u tell your method of prompt it will help me

[fading rover]

Anything specific??

[prime mulch]

How would to normally prompt not specific one

[fading rover]

Well most time I saw in we need to specify the posing or base scene creation . Then to apply style and then I add some mood based things if it needs .

[rigid portal]

Hello

[fading rover]

With banana I saw mentioned twice it confused model or even if we change scene in way behind prompt it does not give proper outcome

[flat crow]

i have a question about seedream 4's outputs; several days ago i'd been using it through lmarena with reference images that it had generated itself, all of a specific character, which it was able to follow extremely consistently in subsequent generations for a few days.
earlier on the same day the model was pulled from lmarena, all generations suddenly generated the wrong character even with the exact same prompts and references as before. i attached a few examples of different (yet consistent across many generations) output behaviors:

1. example reference image used for generating all of the other images here
2. example of the model depicting the character correctly
3. output with same prompt and references as #2, but facial features were now suddenly off every time
4. example where i then tried with other prompts, now getting this character of the wrong ethnicity every single time
5. tried the model through runcomfy as a sanity check, instead had the new issue where all generations with the same prompts/references as before were plagued with a cartoonishly long nose

sorry this is long but i'm hoping maybe somebody might know why i'd been encountering these discrepancies with seedream 4? i understand that these are all typical mistakes but the consistency in which mistakes happen when & where have been odd

[gleaming lion]

Hi team,
I was able to use Seedream-4 High-Res a few days ago, but now it’s gone from my model list.
Has it been removed or is it only limited to certain accounts/regions?

Thanks!

[fading rover]

Most likely its the anime style conversion model does not retain the look and go mad.. I had same experience when certain style used specific like anime or cartoon they have some ref data and then they try to match with them rather matching given face .. even if it does once twice slowly it goes back in to their own modelling way if u generate sequence of 15-20 story board ur character dimly fade into something else 😂

[proper jetty]

Hello. Ampliando las posibilidades de la IA

[toxic egret]

didnt end well

[steep sandal]

What are the chances of us getting seedream 4k back? 2k's teeth and faces are, pardon the word, disturbing

That being said, 2k is back on the platform, folks

[vestal turret]

A 62-year-old Korean woman, gentle and dignified appearance, slightly thin face due to illness, warm but tired eyes with deep wisdom,
graying hair neatly arranged in a simple style, wearing a modest dark navy cardigan, sitting quietly with hands folded, soft natural lighting,
emotional portrait, realistic Korean drama style, high quality, detailed facial features, maternal warmth despite physical frailty

[fading rover]

Refresh Page don't hit recycle ur limit gonna reach

Even if it don't get u make new

[flat crow]

strangely still having this issue with seedream-4-2k that just got added back? it insists on incorrectly depicting this character as white

in order just now:

1. lmarena seedream-4-2k
2. yupp seedream 4 max
3. runcomfy seedream 4 at 2k res (had to shorten the prompt for this one though)

to reiterate it had no issue with this across a large number of generations for a few days, this began happening suddenly on the 19th only through lmarena

[fading rover]

Liked the gouche painterly style have u tried banana for same ?

[flat crow]

nano banana can't do this style at all, it'll use its own style if prompted to with references, and i've found it unable to match the likeness of characters

personally seedream 4 is the first model ive been impressed by at all, if i can't use it through lmarena that's okay though, it's up to the devs and all

[dense gulch]

/video

[robust yoke]

You're gonna wanna do that in #video-arena-1.

[quaint mango]

hi

[robust yoke]

Greetings.

[keen beacon]

hi

[robust yoke]

Greetings to you, also.

[proud hazel]

Hello

[robust yoke]

Greetings.

[hollow flame]

hello everyone am Godfrey ,am here to experiment with gen ai but most importantly to learn from all you's or even copy some stuff by inspirations

[robust yoke]

Welcome to the server, Mr. Godfrey. We are happy to have you here.

[proud hazel]

Welcome King Godfrey

[robust yoke]

Looks very adorable.

[proud hazel]

Yeah 🥰
Made with Sodadream 4

[robust yoke]

I can tell.

[short igloo]

Hello everyone, just testing and learning here

[robust yoke]

Greetings. It's good to have you here.

[fading rover]

It seems nanobanana broken today gives random results way out of what been given as reference.. chose to provide totally absurd output .

[robust yoke]

Very interesting.

[proud zenith]

Hey there

[robust yoke]

Greetings.

Welcome to the server.

[topaz star]

i test this and seems is not seedream 4 at all :/

[robust yoke]

It is Seedream 4.0, but it's the 2k version of it, which isn't as high quality as the 4k version.

[topaz star]

it did bad result like change the face and those stuff , i am not used to seedream to make those mistakes even in low quality

[robust yoke]

Yeah, that's the lower quality version from the high-res version, and thus it'll tend to make a lot more mistakes when editing images.

I tried to generate an image of Walter White standing in a Japanese classroom wearing a male Japanese professor uniform with the recipe for hydrogen peroxide on the chalkboard. It did generate the recipe, at least I think so. But it made everything look very low-quality and it forgot to give Walter fingernails and even made his face look slightly distorted.

[knotty fable]

Some days ago I talked a bit about different AI's.
And some guy did not seem to know what DDG is.
Well I present a little comparision here - it's based on an image of a fictive billboard - the first one is by WAN

And here is the same generation by DDG. As you can see - the result from DDG is far more interesting. So this old AI is certainly not something to frown at. 😺

[topaz star]

i try to test the same promt in other site provider of seedream and in 2k quality and it was better then what i see in lmarena

[potent snow]

Did you remove this again?

[robust yoke]

Very interesting. I tried the one in RunComfy, and it gave me a pretty bad output quality.

Also, I think I just blanked by accident and deleted my message.

[topaz star]

It was the same site, and honestly it was better. I mean, since I’ve used Seedream before it was removed for a while, I can tell when something is off.

[knotty fable]

Using the same prompt can be a good way of comparing AI generation.
However, not all AI's got the same prompt interpreter - so the syntax should be changed to make a good comparision.

[robust yoke]

Interesting. I suppose my attempt to generate a lot of visuals on a chalkboard must have had something to do with the lower quality then.

[knotty fable]

If the model don't have any examples of that, which I could imagine some do not.
I guess that also could be the reason for the lack of quality.

[robust yoke]

This one seems to certainly understand that the thing in the image is a billboard and that it's not supposed to go off of the billboard, so instead it has the car drive off-screen.

[robust yoke]

Well, by the looks of it, it seemed to have a very good understanding of who Walter White is, and it also seemed to have gotten the formula for creating hydrogen peroxide correct.

The only issue was that the quality was a bit low, likely due to the camera distance away from everything.

[knotty fable]

I actually asked the AI to have the car drive off screen.
Since I intend to add text with greenscreen - something like: 'Even when the world is burning you can trust our car insurance.'

[robust yoke]

And it certainly did a good job of understanding that it wasn't supposed to make the car actually drive off the billboard and float off into the distance. It actually made the car smoothly disappear from the billboard.

[topaz star]

The difference is so big that I feel like I'm comparing a current model to an old one.

[robust yoke]

True.

[unborn lantern]

they also reduce the usage of this image generation model

anyone notice it?

[knotty fable]

Indeed it did! 😺 For those who don't read Portu - anyone? Assegurador which already is there = Insurer

[whole sundial]

it seems like the seedream-4-2k model in lmarena ignores all reference images except the first one, even though the model itself is more than capable of it.

[robust yoke]

That's true. It went from being nearly infinite usage to having a rate-limit usage of around 3-4 uses.

[robust yoke]

Interesting.

[robust yoke]

Funny thing is: I actually happen to be half-Brazilian, and thus can understand.

[whole sundial]

sadly, it seems like there is a trend on lmarena of them secretly making their models worse to save money. first it was gpt-image-1 and now it's happening to seedream 4.

[topaz star]

what happen to gpt-image-1?

[robust yoke]

True, it seems like they're cutting costs to preserve money usage.

[whole sundial]

oh no I tried it again and it used the second reference image, don't know what happened with the first one

[robust yoke]

Apparently, a very noticeable quality drop and issues with things like text getting cut off due to the aspect ratio being funky.

The first one got eaten.

The image generator got hungry.

[topaz star]

that explain why i never like chatgpt generation quality

[robust yoke]

Yeah.

[knotty fable]

The number of GPU cycles they allot certainly do affect the result - mudfaces on background characters is a common symptom of that.

..........or ghostly persons who wave in the wind in the background, that in video gen.

[Which is the reason I sometimes redo stuf on local - and have ppl go whoa asking what the heck I have used to get such a result.]

[topaz star]

why they don't just add ads and from it they can handle the cost of models🤔

[whole sundial]

tbh the text cutoff thing sometimes happens with gpt-image-1 normally through chatgpt or api, but the quality drop is new, only started to happen when nano-banana made lmarena viral. I used to like it when it was still a niche place to get near unlimited use of high end models, but now I don't use it much. it may have helped me get off my gpt-image-1 generation addiction lol

[robust yoke]

Unskippable 15-minute ad.

[robust yoke]

Fair.

[whole sundial]

also for some reason when I put multiple references in battle mode, seedream-4-2k fails to generate

(if it shows up)

[robust yoke]

I wish I was around back when the usage was pretty much infinite.

[soft kernel]

It was around 3 months ago

[knotty fable]

3? We had a lot of free stuf 6-9 months ago, not so much lately.

[topaz star]

No, those are just small ads in the corner or whatever. I saw an AI site that applies this — they improve and develop their platform because they get financial support.

[whole sundial]

it was as recent as a little over 1 month ago you could still get reliable high quality generation, right before nano-banana went viral

[robust yoke]

Ah. Well, I guess I must have been using LM Arena around that time then.

[robust yoke]

Yeah, it was more so a joke, heh heh.

[soft kernel]

I mean kinda true,but 3 months ago we had little to no limits there

[whole sundial]

I remember testing it as soon as I was notified about the model being added to lmarena by a discord bot, I remember looking it up on Twitter that same day and there was only 1 post that mentioned it

[knotty fable]

Ahh owk, no worries. I seem to be gravitating toward local & even animation - as I have gotten quality crazee. 😸

[robust yoke]

I remember back when Seedream 4 High Res had no usage limits, at least none that were too low. Such simpler times we lived in.

[knotty fable]

[Chatting a lot now - as I'm in a telephone queue.]

....but now in spot 3, so will soon go poof

[whole sundial]

i generated this image 25 minutes after it was added to lmarena

[robust yoke]

Looks nice.

[knotty fable]

Good animeish look - go animate it now! 😺

[robust yoke]

I actually remember coming across that in the AI Creations channel, I believe.

[whole sundial]

i don't think i ever posted it on this discord before

[robust yoke]

Unless I am simply having a brain fart and remembering some other server whose image I saw.

[whole sundial]

one of the first of over 8.3 million nano-banana image generations

probably in the first 100

[soft kernel]

Well local's a pretty good choice,but like I got no gpus so pass

[robust yoke]

I see.

[opal wigeon]

are we allowed to like use/keep our generated video for our own?

[knotty fable]

I generated quite a number of a rather crappy graphic card, but it took most of the night.
Tho I've upgraded now. But I heard one guy who actually used a phone - meaning he leeched on the cloud cluster service.

[robust yoke]

Well, so long as you're able to actually download the videos, then yeah, there's no rule that says that you can't download your AI-generated videos.

[knotty fable]

Speaking of anime style...... [Note: The music I use in my clips and films are not AI & Image generation SeaArt + video gen = DDG.]

[robust yoke]

Yooo, no way.

[topaz star]

?

[whole sundial]

i looked through my files and it seems like gpt-image-1 got nerfed on September 3

[knotty fable]

I need to test Claude, never done nuffin with that one.

[robust yoke]

Claude is actually very good at coding.

If you ask it to code up a pretty complex website for you, or even just some JavaScript code or Python code, it'll do it lickety-split in very few steps.

[keen beacon]

internet

[topaz star]

hmm so there no hope we get same quality we used to it from seedream model?

[robust yoke]

https://tenor.com/view/matpat-game-theory-vs-bob-fnf-friday-night-funkin-vs-bob-gif-12176028033304353407

[whole sundial]

i'm surprised they kept llama 4 maverick 0326 in the arena after all this time, i guess everybody forgot the controversy where meta, while testing this model on LMArena, manipulated their system prompt so that more people would vote for it, it was the #2 model on LMArena when it was released, but after people found out about it, they had to pull it from the leaderboard and they had to put a version of the model named llama-4-maverick-17b-128e-instruct on LMArena without their optimizations, that model is currently #76 on LMArena, even behind Llama 3.1 405B, the model it was supposed to replace

[whole sundial]

probably not, i've noticed lmarena being more vague and secretive recently, probably because they changed from being a university-backed research project to being a full-fledged company a few months ago, since they now have commercial products like a real company they are now acting like one as well. really sad to see, LMArena was once great.

[robust yoke]

If only there was some way to bring back the old university-backed LM Arena that was more generous with rate limits and optimization updates.

[whole sundial]

top goals of a company: retain secrets, be dishonest with customers/users, attempt to steal basic constitutional rights away through ToS updates, and, most important of all: make investors happy

[void loom]

hi everyone

[steep sandal]

Never mind, I just wasn't using the correct quality tags. Skill issue. If anyone needs tips, DM me. Seriously, it's night and day difference

[whole sundial]

It's funny how they kept the "Created by researchers from UC Berkeley" part of their About Us page intact even though they broke off of that school and LMSYS months ago. I guess it's still true, but they have no connection to it anymore.

[sturdy yew]

What the hell ?

[knotty fable]

Hunyan is only so-so IMO.

[whole sundial]

you should have been more precise with your prompt wording, try "Make the person in Image 1 hold the baseball bat in Image 2 in his left hand"

[steep sandal]

With a properly well-composed prompt and tags to emphasize facial and body realism, 2K can be a beast. Not on the same level as 4K, but extremely acceptable

[sturdy yew]

before they remove it I still can used this promt.

[knotty fable]

Personally I prefer flow and 'live' feeling over high resolution.

[whole sundial]

i did make a video of this btw, (I used Hailuo-02-Pro on fal, it was only $0.48 and it was 1080p too. Hailuo models have always been good with my prompts, so that's what i used)

[sturdy yew]

see.

[knotty fable]

The prompt interpreter on Hailou is decent - it fail one some complex situations - where one need to use script format. But that Hailou got that is a strength for their model.

[sturdy yew]

the problem is not a prompt.

[whole sundial]

oh I had this issue with the model not recognizing input images earlier, sadly there's nothing you can do about this until LMArena changes Seedream 4 again

[neon idol]

Hi chat

Seedream 4 back

Unfortunately only in 2k 😔

[verbal nimbus]

Lol

How does banana do

[steep sandal]

#general message

[neon idol]

https://tenor.com/qCSrKosgI1f.gif

[void loom]

Is generated videos from here can be used in my YouTube channel without any copyright or legal issue and is it gonna monetize ? can anyone help me with this info please?

[robust yoke]

Greetings.

[neon idol]

At everything question the answer is no

[steep sandal]

Just don't use raw and simple prompts with 2k, because then the end result will be regrettable

Always emphasize the qualities you want to see, from the lips to the feet, and then you will have a clear difference

[void loom]

so, i am gonna face issue?

[neon idol]

Ig you post ckntent no one wil be monetize if you use ai content

Bht yes you can use them for being posted

Without issue

[neon idol]

Hey!

[robust yoke]

Howdy, Greg.

[neon idol]

Hru?

[finite cipher]

great

[robust yoke]

None of the stuff you send to the chatbots will be recorded for YouTube.

[robust yoke]

Doing good, and you?

[neon idol]

Me too thx

[robust yoke]

My pleasure.

I am glad to hear that.

I was chatting on my PC, but then this strange insect crawled across my keyboard and ruined my night.

[odd kindle]

damn so seedream-4k is permanently gone

now its just 2k

not complaining though, happy that its at least back

[robust yoke]

Unfortunately, yes.

[knotty fable]

Cthulhu level strange? 😻

[robust yoke]

Not Cthulu-level, but certainly morbid.

[oak aspen]

is there still 3 image only problem going?

[knotty fable]

Insects can look really weird, specially when infected with something - or swollen up full of eggs.

[knotty fable]

<- Biology researcher in real, so I will not go into details with you guys & guyettes. 😸

[robust yoke]

It looks like a cockroach, except it doesn't seem to have wings and crawls very slowly.

It's also brown in color.

[knotty fable]

Oh not super weird, that look like one Acanthocinus. Completely harmless.

[robust yoke]

It was crawling across my keyboard and I hadn't even noticed until I saw the faint shadow of it moving across, heh heh.

[knotty fable]

Yes Acanthocinus move slowly - I think I did a bulls eye even though the image not very detailed.

[neon idol]

I would have already run away from home

And burned it

[whole sundial]

read the terms of use, specifically this part:
Ownership of Content. The Company does not claim ownership of any Inputs or Outputs (collectively, “Your Content”). Subject to Section 3.3 (License to Your Content), as between the Company and you, you are the owner of all right, title and interest in Your Content. Notwithstanding the foregoing, given the nature of the Service, you acknowledge that: (a) Output may not be unique across users and the Service may generate the same or similar output for another user under similar terms; and (b) the Company does not represent or warrant that the Outputs are protectible by any intellectual property rights under applicable law.
tl;dr: LMArena does not own your outputs: you do. LMArena retains some rights to them (mainly for research purposes), but they don't own them and you can do whatever you want with them without worrying about LMArnea going after you.

[robust yoke]

I don't think it's an acanthocinus, but it might be.

[knotty fable]

There's a load of species in that group, but yes, certain id can only be made with a sharp image.

[neon idol]

So no seedream 4k anymore?

[knotty fable]

Well I've finished my phonecall - and will go back to animation. Cya guys and enjoy your generations.

[whole sundial]

yes, sadly another way for LMArena to deal with the sudden influx of users by ruining the experience

[robust yoke]

True.

[knotty fable]

👋

[neon idol]

yes I have a phobia of insects

every insect

[robust yoke]

As do I.

[robust yoke]

Bye-bye.

[robust yoke]

I suppose not.

[whole sundial]

it's gone now, idk if it will ever come back. api providers still have 4k outputs and they all cost 3 cents per image regardless of resolution

[knotty fable]

Not to worry, even though I'm a biologist - silly as it might sound I got a thing with spiders myself. 😺

[neon idol]

Omg 💀

[robust yoke]

That's not silly.

[neon idol]

please don't invite me to your house 💀

[neon idol]

My nightmare be like

[low charm]

helo

[robust yoke]

Greetings.

Alright, killed it.

[neon idol]

Gg

[robust yoke]

True.

[half trail]

😶

Which one is the best image generator and free ?

[neon idol]

Seedream 4

@vocal token look who is back

[barren prairie]

Seedream 4 free ?

[neon idol]

Yes ofc

[half trail]

Where to use it

[neon idol]

On lmarena

Or on chinese website

[prime mulch]

Is seedream 2k is good?

[neon idol]

https://tenor.com/qVcn3RfZMKB.gif

[half trail]

I mean for official

[neon idol]

No its created with another tool

[half trail]

Like we use imagen on gemini

[neon idol]

Oh yes

You can use it on doubao but now is closed only at chinese new users

[half trail]

Meh chinese always hard to access

These days because of nano banana update gemini becomes dumb

[robust yoke]

I actually managed to create an account on there.

Volcano Engine.

[barren prairie]

There isn t another way ?

[half trail]

It generates the same image whatever prompt i write I had to start a new chat for different image

[half trail]

How

[proud hazel]

There is.

[robust yoke]

I used the phone number log-in option, then I set a password on the account, and from that point on I was able to log in to the account.

[robust yoke]

You can use Seedream 4.0 on LM Arena.

[warm vale]

wish you allo agreat day

[robust yoke]

Or, like I said, you can go onto Volcano Engine and log in using Phone Number Verification. Then you can set a password for your Chinese account, and you'll be able to access it that way.

[prime mulch]

Yes result is ok but need improvement in quality

[robust yoke]

And we all wish you a great day as well.

[versed nova]

Is seedream-4-2k even seedream 4

it gets multiple reference images absolutely wrong, like all the time, the quality is horrendous and seems censored

[robust yoke]

Seedream 4.0 2k is a low-quality version of it compared to 4k, which is high-res and able to generate very high-quality images.

[versed nova]

Like at least it's 2k and not this 1400x1400 or whatever it was before, it also gets the dimensions right and isn't always 1:1 but the thing is once you give the AI more then 2 images it completely sucks down and generates whatever it wants igonring the prompt

[half trail]

I tried but phone number not supported

[prime mulch]

You can control aspect ratio now

[steep sandal]

Absolutely. Trust me. But you need to *feed * the machine if you want proper results, always use terms to emphasize details. I was able to generate images that were almost as good as the 4k version

[prime mulch]

Still need to improve quality

[robust yoke]

Yeah that's understandable.

[versed nova]

Yeah but honestly, it's pretty useless now that the AI can't even work properly with 2 images

[robust yoke]

There is a phone number log-in option.

[prime mulch]

It's good that they atleast bring back this version

[versed nova]

i've been trying prompts for a bit now but i can't seem to get it to generate anything i want with 2 images

[odd kindle]

tried it, its definitely not as bad but needs improvement

[prime mulch]

They need to remove restrictions

[odd kindle]

quality is fine, theres really not much difference from 2k and 4k now that you think about it. 2k still achieves high quality

the most annoying part is just the restrictions

[odd kindle]

seriously just a few runs and then gotta wait an hour

[versed nova]

honestly idc about the quality, it should atleast get 2 images combined correctly otherwise it's just lowk worse then nano banana

[prime mulch]

They need to increase little bit of quality little bit if u compare with highres model you can understand

[versed nova]

the only reason i would say seedream 4 is better then nano banana is that it can generate less censored images

[odd kindle]

i do, ive generated tons with 4k

and i dont see much of a difference

but i still prefer 4k ofc

4k just makes the details look way sharper

[odd kindle]

same

[prime mulch]

Zoom in little bit in generated image

[weary token]

the veo 3 in flow doesn't support all languages?

[odd kindle]

i mean, its bound to happen really
they cant sustain the cost free with these many users

so cant blame them ofc

[robust yoke]

I suppose not.

[odd kindle]

they removed the seedream 4k likely due to high traffic and usage costs

[robust yoke]

It does support a wide range of languages, though.

[odd kindle]

2k may cut some expenses maybe

[prime mulch]

Yes it's understandable

[odd kindle]

still sad though 🙁

[prime mulch]

We can use other tools to increase details

[weary token]

I'm trying to do a video ad using flow scene builder, it messes everything up with video contrast

[proud hazel]

Seedream 4-2k
I want to cry. 😭 What have they done to my boy? @echo aurora

[odd kindle]

😭

seedream's generation is good if u make it simple
once u add a little spice they just end up braindead

[versed nova]

Yeah that's exactly what i mean

look at this crap 💔

[prime mulch]

Can u tag pineapple and show the result he can do anything

[odd kindle]

omg

[timid wyvern]

Hello, I am here to try all this tools

[mellow compass]

fun fact, the model of seedream 4 that appears in battle mode is indeed the 4k one (which can combine 2 images perfectly)

[prime mulch]

Unfortunately tools didn't work well

[versed nova]

dang really?

Lets boycott Seedream-4-2k calling it Nightmare 4.0

[prime mulch]

Bro tag pineapple he can do anything about the issue even there is a possibility of removing restrictions

[versed nova]

@echo aurora my mans, I hope everything in your life is absolutely the best and greatest, seedream-4-2k lowk turned into a nightmare help us out 💔

[proud hazel]

Unfortunately Pineapple himself is unable to do much himself. He merely passes the issue on to a higher instance.

[prime mulch]

😭😭😭

[leaden yoke]

I've experienced the same with style transfer since using the Seedream 4.0 2K version today. A very basic prompt like "transfer the artistic style to the photo" with 2 images (a content image and a style image) worked great previously. Now the result is very poor - no style transfer at all.

[mellow compass]

yea, i feel like the seedream2k only works with 1 image, no matter how many you try to add, might be related to the fact that now we can customize the aspect ratio with that model, idk

[versed nova]

Let's make Pinapple the CEO

[versed nova]

Yeah it only works great with singular images, anything beyond that is just well, bad

[versed nova]

Yeah

https://tenor.com/view/cillianmurphygun-cillianmurphy-jazmincoded-gif-3811002643797340103

[severe mason]

/video genera al gato silvertre con un ramo de flores estilo estilo cinematografico

[robust yoke]

Canal equivocado.

Para usar ese comando, vaya al canal #video-arena-1.

[tacit spear]

"Generate video dari foto [upload foto].

1. Scena:
   • Foto berubah menjadi video dengan latar belakang hujan lebat.
   • Karakter di foto (wanita SMA) mulai berjalan melalui hujan.
2. Gerakan:
   • Karakter berjalan dengan kecepatan normal (match movement).
   • Hujan jatuh dengan efek realistis (partikel hujan).
3. Stylization:
   • Gaya animasi: 2.5D hybrid (perpaduan 2D dan 3D).
   • Warna: Dominan biru dan putih (hujan), dengan neon merah di latar.
4. Audio:
   • Lagu rock heavy (150 BPM) dengan gitar distorted dan drum punchy.
5. Durasi:
   • 10 detik."

[severe mason]

😩

[royal halo]

hi im new here

[robust yoke]

Welcome to the server.

[cedar tide]

@echo aurora ?

And pineapple
hunyuan image 2.1 according to tecent need to have 2k generation
not 1k like now.

[tender canyon]

@echo aurora as been ghosting for over 24hrs now
💀💀

Dude’s not even replying or responding to bug complaints in the #1343291835845578853 forum

[robust yoke]

That sort of stuff takes time to do. Maybe I could do it for him.

[final field]

greetings everyone. Prompt success

[robust yoke]

Well done.

[knotty fable]

Looks like Indonesian up there.

[robust yoke]

Looks to be.

[knotty fable]

Yeah I sent it to deepl and it was - some languages got a certain 'melody' or style that I figure without understanding a word. 😺

[robust yoke]

Interesting.

Translation:
"Generate video from photos [upload photos].

1. ** Scena:**
   • Photo turns into video with heavy rain background.
   • The character in the photo (High School Lady) starts walking through the rain.
2. ** Movement:**
   • The character runs at normal speed (match movement).
   • Rain falling with realistic effects (rain particles).
3. ** Stylization:**
   • Animation style: 2.5 D hybrid (blend of 2D and 3D).
   • Colors: predominantly blue and white (rain), with neon red in the background.
4. ** Audio:**
   • Heavy rock songs (150 BPM) with distorted guitars and punchy drums.
5. ** Duration:**
   • 10 seconds."

[indigo anvil]

Hi I'm New here, from Indonesia

[coral gate]

video

[robust yoke]

To create those, you're gonna wanna go to #video-arena-1.

[coral gate]

/video

[unkempt spoke]

@tiny palm 你好

[crude storm]

hello, i want to use nano banana

[robust yoke]

You can do so on LM Arena's website then.

[crude storm]

ok, thx

[robust yoke]

My pleasure.

[brittle blade]

@brittle blade একটা ভিডিও

[proud hazel]

https://tenor.com/view/paul-mccartney-beatles-paul-mc-cartney-the-beatles-gif-5662981

[robust yoke]

I've got this, ladies.

[robust yoke]

আপনি বিশেষভাবে এই চ্যানেলে ভিডিও তৈরি করতে পারবেন না. এটি করার জন্য আপনাকে #video-arena-1 এ যেতে হবে

[neon idol]

Hi chat

News aboit 4 high res?

[robust yoke]

Greetings, friend.

Still no new news.

[neon idol]

✌️

[neon idol]

Rip

[robust yoke]

True.

[native bloom]

/video prompt: A cinematic, slow motion shot of https://pathwayguides.myshopify.com/ coming to lfe with the company logo

[robust yoke]

https://tenor.com/view/pffttt-cantholdlaugh-father-laugh-dad-laugh-gif-13166789396741539132

[native bloom]

/video prompt: a cinematic slow motion shot of a lion roaring over a sahara lands in Africa

[knotty fable]

May I kindly direct you toward the video creation channels https://discord.com/channels/1340554757349179412/1400148597768720384

[native bloom]

sorry

[robust yoke]

It's alright.

[wide briar]

...hello?

[robust yoke]

Greetings.

[wide briar]

Alright. Thank you.

[robust yoke]

No problem.

[flat crow]

agree, i've had the same experience

[wide briar]

Okay. So uhh, oh man, this LMArena web thing is saving my life. Y'know what i mean? Since ChatGPT make the "GPT-4o" model only available for paid users..., i thought i couldn't find any alternative websites for that one. But, here i am...

[robust yoke]

Yeah, I definitely get where you're coming from. At first, I thought I would never be able to find a website where I can use different models like Grok 4, Claude 3.5, or even the 4 models. I definitely understand being able to actually choose what model you want and talk to them completely for free.

[raven pike]

I came here to learn, thank you.

[robust yoke]

I'm glad to see that you get to use the 4o model and have a good time.

[robust yoke]

Welcome.

[knotty fable]

우리는 모두 배우고 있습니다. 오랜 시간 해왔더라도 말입니다. / We are all learning, even when we've been doing this for a while.

[wide briar]

Yups. But, ehh, still lacking on the web search though. Like, you should start a new chat again to use that feature (with "web-search-based" only model tho, chatgpt4o-latest can't use that on this web). Soo, i hope there's an update for this issue in the future... (Or..., idk.)

[robust yoke]

Well, about that - there is actually a toggle to enable search mode. However, upon doing that, it will only allow you to select the web search models that get utilized whenever chatbots do a web search to pull information.

Whenever chatbots perform web searches, they actually use a separate model to pull data from the web, which they can then use for whatever task you are putting them to. So in order to use those models, all you have to do is select the little globe icon.

[dusk stag]

just joined and saying hello

[robust yoke]

Greetings.

[dusk stag]

Thank you. I am actually excited to test using your platform

[robust yoke]

Ah, it's not my platform, silly.

It's the LM Arena development team's.

[forest sleet]

Well, an opportunity has come up for us to test all the services. Let's go!

[elfin flax]

lm arena by lm arena

😎

[robust yoke]

Eggs act Lee.

[vocal token]

idc

[wide briar]

I see. But i mean like, it doesn't work like what ChatGPT does. Where you actually can turn on/off the web search, anytime, on one-time model. Well, if LMArena does not designed to be like that..., i wouldn't judge then. 🤷🏻‍♂️

[robust yoke]

Well, think of it this way: you'll have one conversation where you use a model that is capable of searching the web, and in another, you'll have a model that is capable of providing you with fast answers without searching the web. So in turn, you have two AIs that are both equally useful for whatever needs you require.

[robust yoke]

https://tenor.com/view/daniel-damn-ar-one-enel-gif-26279142

[wide briar]

Mhm. Y'know it, the A.I. sometimes are not answering correctly when you ask specific thing, that the A.I. doesn't know much about it (or don't even know it maybe!). That's when you should use the "web search" feature thing.

[robust yoke]

Exactly. In case you have an inquiry that requires a web search model, you'll already have a conversation where web search is engaged, and you can just go to that chat room and ask that question that you have. If you are in a situation where web search is not required, then you can just go to the other conversation.

[wide briar]

Hmm-hm. Ah, i see i see...

[wanton grotto]

heyyyyyyy

[neon idol]

Lol

[robust yoke]

Greetings.

[wanton grotto]

guys i am building an company related to ai software solutions i have 2-3 ideas related to it, is any one interested then lets have a talk onn it i have worked on agentic ai and automaton stuff.

i don;t have company yet but i am lookin gto build a team..

i want cloud and full stack person to collaborate if anyof u is interested

[robust yoke]

Hmm…

Very interesting…

[tardy wagon]

its back

[earnest rover]

I mean these isn't allowed in lmarena as far as I know @echo aurora remove it.

[honest atlas]

Hello fams I came to check out tools

[finite oak]

great portal

[true swallow]

hello

[nocturne sail]

mandatory message to check of discord boxes

[peak owl]

hello

[cloud granite]

I came here to learn how to generate videos.

Someone please help me.

[remote arrow]

How could the site become reliable in positive-negative feedback on AI models, if they are not put the models name correctly? Even worst: use their own names that confusing people..

[granite heron]

please check in here first: #1397655624103493813

[static sorrel]

Can it add sound in the video?

[sonic hinge]

Word up LMArena

[ornate agate]

Fast is similar elo to a bunch of open source models, and actually below a few of them if you remove style control.

I don’t see the use case for it personally

Gemini flash is also in that position performance wise but it’s integrated with GCP so it has a lot of natural users and also probably internal traffic from Google too. What’s the equivalent for grok4 fast? Running the X algo internal traffic or something?!

[sullen quest]

What, 2.5 pro is better than 2.5 flash

[worn tundra]

is seedream 4-2k the same as seedream 4-high-res?

[topaz star]

nope

[worn tundra]

thanks for the quick response, one more question -> is the high-res version only through anon battle available?

[topaz star]

i didn't see tbh

someone say it is there , give it a try maybe u will find it

[ornate agate]

I hope that the grok 4 fast pricing encourages others to offer competitive pricing. Flash used to be way cheaper then they jacked up the prices lol

Yeah I agree with you. Inference even of open source models is way too expensive at the moment compared to what it could cost

I’m not sure Google was undercutting before though, just I think that now someone made some bad increase the profit margin type decision.

Google should have the lowest net cost by far of anyone, because they use their own chips. They should be able to beat everyone else by a factor of 4 or more on price.

[minor kernel]

#general

i am here to create and test more videos

[knotty fable]

It's quite funny to accidentally find out what's under hood of some models.
I did some posts on 2 other groups in the past, pointing out that Hailou had a sheepload of BBC footage - where one othe chinese one had Italian 1960-70's footage pooping up whenever one did mention scifi. 😸

[minor kernel]

how can i generate more videos

[knotty fable]

Go to any of those who offer free generation - depending on model and AI you can do 2, 3 or 8 per day - though some with watermark.

[versed nova]

Pineapple is back

https://tenor.com/view/squid-game-squid-game-2-cheers-happy-excited-gif-13406404385856133045

[echo aurora]

It's going to be random if you video gets sound or not. All the models available don't have sound support, and since it's random what model you get, it's going to be random if you get sound or not.

[echo aurora]

[cedar tide]

@echo aurora
Thanks for fixing Seedream, now need to fix Hunyuan Image and everything is fine.
#announcements message

[echo aurora]

sup

[cedar tide]

hunyuan image 2.1 according to tecent need to have 2k generation
not 1k like now.

[civic spindle]

Do the laugh they say? Pfft

[echo aurora]

Would you mind putting together a post in #1343291835845578853 and put the details there so I can review in one place?

[civic spindle]

MWHOHOHOHOHO

[prime mulch]

Hey pineapple thanks for releasing seedream 4 2k but that need some updates can u remove some restrictions

[glossy umbra]

Does this mean in other words "Can you make it NSFW"

[prime mulch]

Noo

No nsfw

[knotty fable]

I knew I'd see that question soonish or lataz. 😹

[versed nova]

Honesty Seedream 4 was generating hella nsfw

without even asking

[prime mulch]

I mean if u compare the result previous seedream highres and seedream 2k you can see some difference in seedream 2k not i quality i mean in output

[prime mulch]

Bruh explain him what i mean

[versed nova]

lowk no idea what u meant either

[knotty fable]

Yeah it happens, DDG and SeaArt do so at times - which can be both amusing, as well as frustrating - when a perfectly good one end up under the censor screen and points lost.

Some models got quite some pron under the hood so to speak.

[cedar tide]

Thanks for replying and it's done ✅ https://discord.com/channels/1340554757349179412/1419716746054729798

and besides that do you know what the grok 4 fast settings are? (high reasoning?)

[echo aurora]

Ty for putting that together

[echo aurora]

grok 4 fast
iirc this was high reasoning, but will double check and keep you updated if I'm wrong.

[cedar tide]

Thx

[echo aurora]

that need some updates can u remove some restrictions
would you mind elaborating? Sry to say I'm not following.

[echo aurora]

on our platform it was?

[versed nova]

Yep

[prime mulch]

it gets multiple reference images absolutely wrong, like all the time, and sometimes it give wrong output even in single image reference

[prime mulch]

@echo aurora this is one of the example

[sage hinge]

don't remove this my boy

[flat crow]

i can corroborate this as well, it can't follow my reference images the vast majority of the time anymore. either with 1 or multiple

[sage hinge]

honestly, seedream 4 2k is much better than the 4k version, especially in the time to generate, it is much faster and the results are very good

[echo aurora]

Okay thank you for the examples. I'll flag this.

[versed nova]

Until you add 2 images, after that it just fails miserable compared to the 2 version before

[lapis imp]

hi jsut wondering how do we use video arena 😅 ? i want to generate a video for a trailer im making currently or?

[proud hazel]

#1397655624103493813

[sage hinge]

oh, not everything is stable, but it's a good version

[flat crow]

yeah, the difference in resolution has been totally fine at least

[echo aurora]

Are you noticing this more when it's multiple image uploads instead of 1? Seems like @flat crow you're seeing it with just 1 reference image it's not working?

[versed nova]

Yeah resolution wise it's pretty good, but it pretty much lost the most important feature

it also gets the aspect ratios right now

[prime mulch]

This is the issue i mentioned

[flat crow]

hm, i thought i had tried one with only 1 ref but apparently not. will check right now

[prime mulch]

It suppose to be give this result @echo aurora

[calm grotto]

Hello, I am here to learn new ways of making Ai videos and checkout new tools.

[ember iris]

Seedream4-2k (face without sunglasses) doesn't create face properly... Seedream4-high-res (face with sunglasses)was much better

Faces getting 🍌 ed up in this version of seedream4-2k

[knotty fable]

You might be in the right place then - go check the video arena channels, and see how different people use prompts. Some add way too much terms - which in a few cases have confused the AI so much they got no result at all.
While a few use actual cinematic scripts, defining what action should be seen at any given second. [I've used that format myself with Hailou.]

[north pike]

Seedream4 is messing with contrast way too much...

[knotty fable]

Commoon problem, most AI's show oversaturated colours and contrast - one reason I prefer image to vid generation.

[sage hinge]

now is it possible to choose the aspect ratio of the generations?

[green tusk]

Hello

[desert mesa]

hi

[desert mesa]

finally! got a unique tag

not this channel

💔

[sage hinge]

Sure, I'm creating it right now

[gleaming iris]

What image generation model do you guys think would be the best for trying to create a ‘sequence’ of sorts? Like, a five panel comic-style sequence.

[flat crow]

with one reference it got the character's ethnicity right (image on the left) but its still subtly off in similar ways, as in the character's skintone and lips are off. interestingly i went to compare with seedream 4 max over on yupp using only 1 reference, and for the first time it made the same mistake with the character's ethnicity (image on the right). i usually use multiple references so i imagine maybe only one isn't usually enough to get its features accurate. maybe the bug on lmarena comes to it only using one reference at a time regardless?

[versed nova]

Image ination

[flat crow]

it could be that the issue being seen with more than one reference is if the prompt depends on multiple at a time, rather than just reinforcing the same thing

[prime mulch]

@echo aurora you can see this examples what i try to tell

[gleaming iris]

I… don’t see that on the model selections for LM arena’s direct chat…

[versed nova]

I know, it was a joke

honestly prob nano banana

[gleaming iris]

Ah.

[versed nova]

https://tenor.com/view/jgmm-cat-meme-ragebait-cat-think-cat-ponder-gif-7034802297606826390

[glossy umbra]

Grok 4 fast doesnt have the 2m context window on LMarena , does it? @echo aurora

[ember iris]

Atleast you have a face recognizable... Check out the eyes of the images.. 😁

[flat crow]

tried once more with a single different reference that has similar lighting to the desired image and the character's face is up closer, this time it captured him correctly. so that may be it (sees only one reference at a time even with multiple provided)

[prime mulch]

Can u put this as bug and tag pineapple

[ember iris]

Have spoken to @echo aurora since they took down seedream4-high-res

[hearty ferry]

FLUX 1 KONTEXT PRO isn't working

are u guys not fixing?

[proud hazel]

https://tenor.com/view/jeremy-jeremy-clarkson-bmw-top-gear-hit-gif-17392956

[hearty ferry]

?

[lament thicket]

hi

[echo aurora]

On our end we're not seeing this model error out more than usual. That being the case it sounds like you're catching these errors because of rate limit.

[echo aurora]

thank you for the example.

[flat crow]

tysm!

[weak meteor]

[radiant cradle]

hi everyone

[cyan harbor]

did lmarena made changes on their side or google for nano banana. I lit can only edit 3-5 images only like wtf

even generate that many and then wait 40-50 mins

[echo aurora]

We did make a change to rate limits, about ~2 weeks ago.

then wait 40-50 mins
You're having to wait 40-50 mins for each generation? How consistent is that?

[cyan harbor]

i just said above didnt i? like after 4-5 generations/edits

thats mad disappointing

[echo aurora]

even generate that many and then wait 40-50 mins
Are you saying you then have to wait for the rate limit to drop, or for the generation to happen? Sorry I'm not following and could use some clarification.

[cyan harbor]

yes i need to wait

[echo aurora]

for the rate limit or for the generation to happen?

[cyan harbor]

limit

[echo aurora]

Gotcha, yeah that's intended behavior. We do have rate limits in place.

[cyan harbor]

kinda sucks but ok i got u

myb lower waiting time?

[echo aurora]

It's possible, if we do make a change like that we'd be sure to share.

[echo aurora]

I wasn't able to repro this btw, looks like it was able to get the face right.

[ember iris]

I should wear sunglasses for all the image generation then...

This was the result I expected.

[ember iris]

This was generated using seedream4-high-res

[ember iris]

This picture is generated by seedream4-2k

[loud crag]

Seedream 4 is weird. It's amazing at almost everything, but eyes aren't... great.

[frosty adder]

Looking for someone to collaborate with! If you’re up for teaming up, let’s connect

[proud hazel]

For what?

[frosty adder]

For big project

[verbal nimbus]

I thought that was Russell Crowe, lol

[lunar glade]

I concor

[verbal nimbus]

(because they're dead inside) /jk

[haughty smelt]

hey guys does the 10 dollar subscription can i use all the models without limit or is there some sort of credit

[lunar glade]

Banana at least it can keep things look genuine

[paper smelt]

How confident are you in the future of medical AI models? Do you think they will play a dominant role in healthcare going forward?

[proud hazel]

What subscription are you reffering to?

[verbal nimbus]

Not sure, a pharmacokinetics student shared a prompt and all models got it wrong

All models failed to identify a fatal combination of drugs

I tested it myself

[verbal nimbus]

Because they did not consider basic mechanisms or something (too complex for me)

GPT-5 was the only model to get it with the hint

Oceanstone got it wrong

[paper smelt]

In my company we have multiple projects which related to drug discovery, brain stroke etc. also we have phd ppl in that project. It works fine with fine tuned model with real data.

[verbal nimbus]

It seemed to fail at less common combinations

[paper smelt]

I agree its not 100% accurate. But it can ease doctors burden mostly

[verbal nimbus]

Ones not on the internet, but which you can just work out the mechanism yourself

[lunar glade]

I'm not scientist or medical certified nobody but if AI can and understand the logic "inspection, listening and smelling, inquiry, and palpation " the very basic orders to infer the condition, course, and location of the disease, then we can talk

[verbal nimbus]

It says not fatal but it was

[paper smelt]

Yes we have robots which already working in hospitals. Where it can operate surgeries .

[verbal nimbus]

If you're a pharmacokinetics student ig. It looks like alien language to me.

[verbal nimbus]

Isn't that telerobotics?

i.e. human controlled

I remember there was a meme about surgery on a grape using one of those systems a few years ago

[paper smelt]

No, advanced robotic system for orthopaedics, especially for joint replacement surgeries

[verbal nimbus]

What does it do?

[paper smelt]

Even i remember it
When i came to know about our own product 😐

[paper smelt]

It helps with precision, planning, guidance, and execution of parts of the surgery.

[verbal nimbus]

https://www.auckland.ac.nz/en/news/2025/03/31/AI-knee-surgery-breakthrough.html

Is it this?

The computer model goes through and analyses the 20,000-25,000 potential positions and ranks them according to what would be optimal for that patient.

Very smart

[paper smelt]

Using the 3D model, the surgeon can run through the surgery virtually

Not this

[verbal nimbus]

AI surgeon?

How so?

[paper smelt]

Its Misso

[verbal nimbus]

Yeah that's what I'm thinking

Oceanreef is too inaccurate to be Flash

Probably flash lite

[paper smelt]

Cuvis is kinda good but misso in early stage

[verbal nimbus]

I googled it, very interesting application of AI

[paper smelt]

And thank you
Its good to have a conversation here 😇

[paper smelt]

Here am who architect core projects in AI

[verbal nimbus]

With all the Gemini integrations this month, I feel like 3.0 is around the corner...

I feel like it's going to have video built in or smth

Because they just added Gemini to Android TV

[paper smelt]

Is anyone working in fine tuning the base model with real data?

[verbal nimbus]

That's when they start their trial of Gemini in Home Assistant too

Maybe they're very confident in Flash/Flash Lite to deploy it on so many devices

They're starting a limited trial in October

Yeah exciting stuff

I actually have 3 Google Home Assistants in my house lol

[verbal nimbus]

I just hope they don't forget to add the basic stuff

But it could use a bit of accuracy boost

I heard Apple is going to use Gemini too

They really missed out on a huge opportunity

[restive scarab]

Hy can someone help me

A large code was generating by claud and gpt 5, but it auto stop at the edge of end

And it shows still response is generating

But no chat loding

[verbal nimbus]

Especially LED ones that can change color and are dimmable

[restive scarab]

After refreshing no chat spears of that

[verbal nimbus]

It's fiddly in the app

[restive scarab]

What's effor

[verbal nimbus]

Some were okay, others were a pain

[sweet tinsel]

Craig, you also have been in here for a bit longer, would you agree on me that the general quality of a conversation here has degraded? It surely feels different from the LMSYS times for me.

[verbal nimbus]

I think something to do with automatic 5 ghz switching

But once added it's pretty reliable

The problem with Gemini is that they force you to disable Google Assistant

(the app)

But Assistant can do a lot of things Gemini can't

Lol

I need an AI to spy on me and tell me where I left the keys

[chrome cosmos]

Hello everyone from the Canary Islands. Delighted to explore AI advances in this place and share with the community

[soft kernel]

@verbal nimbus hey you tested oceanstone

??

[leaden egret]

Is the new Qwen Image edit even close to gemini and seedream editing

[proud hazel]

No

[leaden egret]

Are there any benchmarks or is it just inferior by testing

[proud hazel]

by testing

[plucky whale]

Deepseek R1 its not available

[verbal nimbus]

Yup, but I hardly encounter it. I did manage to test it on the question all models get wrong except GPT-5-High, and it got it wrong

It's better than nightride though

But limited sample size because I encounter it too infrequently

[fiery gull]

What is oceanstone? A new secret model?

[verbal nimbus]

Oceanstone and oceanreef are rumored to be 3.0 models

Because their style is very different from 2.5

[fiery gull]

Flash 3.0 finnaly? 🥹

[verbal nimbus]

Hopefully Oceanstone is flash

If it's Pro, it'll be a bit disappointing :p

Since it didn't beat GPT-5-High on hard questions

[steel sorrel]

I wonder Why there are so many native image model release recently? Is there any reason like new research paper? or just it's so good to get attention from public?

[fiery gull]

I really hope it's gemma 4 and even more likely to be him

Gemma 4 27b

Or gemma 4 24b a6b? Idk

[verbal nimbus]

Qwen just released a 30B Omni model

[fiery gull]

I saw this

[verbal nimbus]

Like 2 hrs ago

[fiery gull]

Just waiting for the iq2 kxxs to run on my phone

[verbal nimbus]

That'll need about 8 GB of RAM?

[fiery gull]

Nope, 12gb for iq 2kxxs and 16gb iq 3kxxs

My phone have 16gb ram

[verbal nimbus]

Qwen 3 is free on OpenRouter (with logging) hopefully this will be too

[verbal nimbus]

Wow, mini computer almost

What phone is that

[fiery gull]

Redmi note 12 turbo

[verbal nimbus]

Are you sure it's 16 GB? Google says 8-12

[fiery gull]

This is because android consumes by itself 4gb, I have already tried to run it on an 8gb phone, even in airplane mode the phone crashed

For 8GB phones I would use the Qwen 4B 2507 IQ 4NL

[verbal nimbus]

I usually just use DeepSeek or Qwen via the API on OpenRouter since it's cheaper than Flash

Also better

[fiery gull]

I don't use api, I just use the standard chatbot from aistudio, and I use qwen local to serve as a particle teacher while I'm doing walking and there is no internet

From what I tested the oceanstone, it seems to be more like the gemma 4, and the gemma 4 is at launch time (6 months after the 3)

[verbal nimbus]

Interesting

Oh shoot Nvidia is up again

Should have bought some stocks the other day

Better

[fiery gull]

The oceanstone is to close to tie the 2.5 pro?

[verbal nimbus]

It got some things right that 2.5 didn't, but I encountered it too few times to know for sure

Also seems less sycophantic, which is good

But formatting seems less cohesive than Pro

[fiery gull]

Hmmm so this increases the chances of him really being the gemma 4

[verbal nimbus]

I doubt Gemma 4 would be on the level of Pro though

Just because Gemma 3 wasn't even at the level of Flash 2.5

[fiery gull]

Gemma 3 = flash 2.0

[verbal nimbus]

Yeah, and if Gemma 4 is on par with Flash 2.5, I don't think it would be beating 2.5 Pro

I suspect Flash or Lite must be very good

That's why Google has been deploying Gemini everywhere this week

[fiery gull]

I think the Gemma 4 will be better than the Flash 2.5 because the 2.5 is an experimental model for MoEs.

[verbal nimbus]

Android TV, Chrome, Google Home in October...

[fiery gull]

But just looking at the date, I'm 90% sure it's gemma 4

[verbal nimbus]

I feel like they'd be working on Gemini first

I don't think there was an anonymous model for Gemma 3, was there?

[fiery gull]

By the pace of the releases it will be the gemma 4 or both

[fiery gull]

As far as I remember there was no secret model of the gemma 3

[thorn valley]

i believe the main thing would be the processor, is it a snapdragon?

[fiery gull]

7+ gen 2

[verbal nimbus]

Hmm they did release Gemma 3 like 2 weeks before 2.5 Pro

[faint osprey]

This hand comes flying to the ground with wings

[fiery gull]

They'll hang out together or very close

[fiery gull]

Is the new qwen image?

[faint osprey]

Yes

@fiery gull yes

[balmy mist]

anybody test qwen new model?

[thorn valley]

suffering trying to run deepseekR1-qwen3 locally lol

it gets very hot

[fiery gull]

Why r1 qwen 3 8b? And much better the qwen 3 4b 2507

[fiery gull]

Mine gets cold, I get angry it seems that he is not giving his 100%

[loud crag]

I don't think it's a bug but just me being stupid, but I can't get Seedream 4 to improve a face from one image, with a more detailed face of the same person from a 2nd image. Don't know what prompt to use

[thorn valley]

I can test later

[thorn valley]

I downloaded the Edge gallery and installed Gemma 3n4b to at least have some offline

deepseek got very hot

[echo aurora]

I've seen others report something similar earlier. Are you able to provide an example?

@empty mauve you should check out #1397655624103493813 for more information on how to use the bot.

[crisp sequoia]

hola

[thorn valley]

hi

[verbal nimbus]

Impressive if true

It's supposed to support real time audio/video

[crystal dock]

I am not a spammer

I want to be a voice actor

[ornate agate]

I believe them, can probably try it on their website if you want

will take a bit of time for it to work with other stuff

[crystal dock]

What do you mean?

[broken quiver]

Hello, I am here to test some of my ideas

[wispy ember]

hello to everyone. i am new here and still learning how to generate videos, i tried once but didn;t succeed. tutorial seemed fast. if anyone has more simple explanation, i would appreciate the help. thanks

[remote arrow]

Simply go to video-arena (1-3) and use the "/" symbol on your keyboard. Then you'll have options to create "text-to-video" or "image-to-video" or "text-to-image"

[echo aurora]

Be sure to check out #1397655624103493813

It has a lot of helpful information in there.

[granite holly]

hello

[granite igloo]

gemini is stuck on generating, what do I do? Can't just open a new chat as the chat history is pretty important

[hollow ivy]

have you already encountered route66? (it seems to be a GPT5-derivative)

i wonder if it's better than gpt5-high

and how good is that gemini-model called "sorting-hat"?

-# (i also encountered "Bailing" by Ant Group of Alibaba)

oceanstone seems to be decent for python-coding, if not asking too complex things

"Bailing" = Ling-flash-2.0

strangely a Claude model was identified as "GLM-4.5" (although it talked like Claude)

so that old bug still seems to exist :/

[golden ocean]

-# (i also encountered "Bailing" by Ant Group of Alibaba)

[misty vault]

-# (i also encountered "Bailing" by Ant Group of Alibaba)

[languid ridge]

saw a video speaking about and was like "Hmmm what's catch?" No catch works as advertised.

[stiff glacier]

New one

[sullen quest]

we really need a channel just for cloaked ai encounters

@echo aurora what do you think of that?

[echo aurora]

Ooo interesting idea. Yeah it's possible. I'd like to see a bit more interest from others before we do so. I'll create a feedback post so others can weigh in there.

[modest nova]

Hi all I've just joined and loving the idea of expanding my knowledge

[echo aurora]

#1419844575064756385 message

[wide hatch]

Hey, I'm C-jay, I just got intoduced to LMArena, and am checking out all the wonderful services.

[echo aurora]

welcome welcome!!

[golden ocean]

stop

[hearty fog]

HELLO

[sage briar]

Hey, I'm prime sparrow, I just got intoduced to LMArena, and am checking out all the wonderful services

[proper field]

nice! what style and model did you use?

[echo sinew]

Hello! Please read #1397655624103493813 to learn how to properly prompt the bot.

[fallow holly]

Hey, good night peoples!

What's the best AI for coding?

I've done several tests, but I want to get an opinion from someone else

[whole sundial]

scroll up, it was nano-banana and it was supposed to be ghibli-style but nano-banana was never trained on that style it seems so it went with a generic anime style instead, which i still think looks nice

[azure notch]

Hello! How and why does LM Arena offers all premium models for free?

How is that even possible?

[still temple]

Just me?

[mellow compass]

is down

nvm is back

[whole sundial]

they pay for it themselves. for stealth models the cost is shared between LMArena and the model provider

the money they are currently using is VC investor money, but they are making commercial products to gain income

[azure notch]

Why would they pay for it in the first place?

I don't see any subscription or any business model

[lunar glade]

just use it while you still can lol

[hard quiver]

I was testing the SeaDream 4 model, specifically by sending it a photo of a person and a photo of an outfit I wanted that person to be dressed in. But when I executed the prompt, the output gave me a completely different person—it generated a random image, with a different pose, a different skin tone… Instead of editing the image, it created a new one. Meanwhile, other models, like nanobanana with the same prompt, worked just fine. Why do you think this happens?

[prime falcon]

create a video of a Indian farmer fighting with a tiger in Indian forest and many animals are coming to see this fight

[flat crow]

there were a few earlier conversations about this, i think it's a bug where it can only see one reference at a time with seedream 4 (seems to be the last image in order of how they're displayed). in the meantime maybe an ad-hoc solution could be to merge your refs into a single image if possible

[sage thorn]

hello I am here sent by Galactica crypto project

[honest olive]

Howdy

I an looking to test Text to Video Features

[prime mulch]

It because of restrictions make a post about that in bugs

[whole sundial]

they pay for it to give feedback to the AI companies for free (but they are trying to monetize this part)

[abstract sinew]

Hello, I've noticed that Seedream4-2k is not performing as well as it used to when it comes to handling face consistency.

[lusty tinsel]

is it me or seeddream 4-2k making everyone look kind of korean?😆

[flat oak]

you can edit your face then in editing then

[echo aurora]

Hey everyone - been seeing reports of this. Let's start to consolidate examples here if you don't mind #1419929810469126226

cc @lusty tinsel @abstract sinew @flat crow @hard quiver

[echo aurora]

Not sure what that is so may be in the wrong place

[echo aurora]

Howdy be sure to check ouit #1397655624103493813 for more information on how to use the bot!

[mighty thicket]

hello

[fiery moss]

hello

[rancid wadi]

Can somebody tell me how do i exactly turn my images into the video, cause i did tried to do that in the video arena 1 but no response came, and i also noticed that the people who turned their image into video has a tab of image-video conversion.
Can somebody be kind enough to help me out!

[echo aurora]

You need to prompt the bot, use slash commands (type /image-to-video) the info in #1397655624103493813 should be helpful too

@near mantle ^

@fiery moss @mighty thicket

[remote arrow]

🤣

[near mantle]

/image-to-video A Japanese samurai standing on the street on a rainy night.

[hushed ingot]

Hello

[lusty tinsel]

@echo aurora any chance seededit 3.0 come back to direct chat? i used to love how it make 3d models like actual game render

[prime mulch]

https://discord.com/channels/1340554757349179412/1419929810469126226 can u post the seedream 2k issue in here

[formal cedar]

Navratri का त्योहार आया तो गाँव में भक्ति और उमंग छा गई। घर-घर दीपक जले, माँ दुर्गा की प्रतिमा सजाई गई। महिलाएँ गरबा नाचने लगीं, बच्चे खेलते रहे और पंडाल में भजन गूंजे।

[strange drum]

/image-to-video Cloud computing having a datacenter in cloud

[verbal nimbus]

Yup, once. It won the match, but didn't seem outstanding

[thorny kelp]

Hello! I wanted to check out ai video generation

[blissful silo]

Hello

[hexed tree]

Can anyone help
Typing / but no video commands appearing lol

Any 1?

[proud hazel]

#1397655624103493813

[knotty fable]

@echo aurora While it might be interesting, I don't think this post belong here #general message

[weak meteor]

It load like 30 min already

Refreshin doesn't work

And Ive finished my breakfast

Ts pmo

[unkempt onyx]

Generate a slow-motion cinematic video of a giant chocolate fountain overflowing, but instead of chocolate, it pours rainbow-colored molten cheese that hardens into mesmerizing patterns. Background: upbeat viral TikTok music. Camera angles: close-up dripping shots + wide reveal with people’s shocked reactions.

[outer dove]

Hi, so please does someone know what is the AI used by ChatGPT actually to create and editing Pic please?

I'm actually trying to see if
Seedream 4 / flux-1-kontext / Nano banana or GPT what is the best in different kind of prompts (creating / restoring / editing)

Thanks.
P.S please don't blame my English I'm French. Thanks

And does AI used by ChatGPT for Pic is on lmarena

[prime mulch]

There is no pic editing ai in chat gpt use nano banana or seedream those 2 only support editing