>The token bill comes due: Inside the industry scramble to manage AI’s runaway costs https://techcrunch.com/2026/06/05/the-token-bill-comes-due-inside-the-industry-scramble-to-manage-ais-runaway-costs
06/05/2026
>RhymeFlow: Training-Free Acceleration for Video Generation with Asynchronous Denoising Flow Scheduling https://simon-dcs.github.io/Website-of-RhymeFlow
>Can We Predict The Human Preference For Text-to-Image Content Prior To Generation And Is It Even Useful To Do So? https://github.com/LSU-ATHENA/HPM-Predict
>When AI builds itself: Our progress toward recursive self-improvement, and its implications https://www.anthropic.com/institute/recursive-self-improvement
>>109001866 I prefer 40 steps because it gives me a free hamburger. Also, what is Ayakon in your filename? Is it some kind of sampler or scheduler? An Ayakon scheduler? I've never heard of it, could you tell us? ;)
>>109001976 I guess one could make a case for slow training as it kinda makes sense that slow training could improve quality. I don't know if I have the patience for it though. I generally want to be done training after 30 epochs
>>109002175 telling me to "hone my skills" while defending an image with no background, simple colors, and slop colors despite this being fixable with also a prompt doesnt work retard
>>109002190 >defending an image I'm not defending the image I'm defending the model because you blamed the model and not the genner >>109002092 >despite this being fixable Then why did you claim it's the models fault?
>>109001724 >zimg the average zit gen got stale about a week after its release and only one or two anons use loras with it zib is more flexible in that regard
Is there anyone here who has had sex? Does anyone know how horizontal nipples are created? Is it something genetic, or does it happen because of external factors like having babies or hormonal changes that affect the body for baby making?
>>109002202 >I'm not defending the image I'm defending the model if the model needs to be told to do the most basic shit like generate a background, any background fitting an image, otherwise it leans heavily to generating an empty background despite it costing the same compute, then it means its a shit model that doesnt cater to what basically any end-user wants while using it.
>>109002230 >>109002250 >then it means it a shit model not madel that doesnt cater to what any end user wants while using it. Why would you want a model to generate things you dont prompt for? I don't need the model to handhold my prompts.
Once a finetune of Anima based on e621 "like Noob" (not Noob/Illust as was incorrectly claimed) does release you will encounter the same problems you have with Anima. You don't want actual Noob, you want a slopmix that makes up for you lack of prompting abilities.
>>109002257 >Why would you want a model to generate things you dont prompt for? I don't need the model to handhold my prompts. do you also want to have to tell the model that a character should be alive instead of dead, to always prompt a pose otherwise the character is in T pose, to prompt where the character is looking otherwise it has dead eyes staring blankly, to prompt art style line thickness otherwise the lines are actually invisible and theres nothing in the image?
imagine being so retarded you have to defend a model wasting the same amount of compute on generating a white background instead of literally anything else for free.
>>109002323 its not rhetorical. do you want a model to hold your hand or do you want to prompt art style line thickness otherwise the lines are actually invisible and theres nothing in the image?
>>109002337 I already told you I don't want a model that inserts things into the output that I didn't ask for. If the artist I'm using is known for both thick and thin lines, for example, and if I don't specify line thickness then it should be an amalgamation of both extremes. That's why if you don't prompt for a background you're most likely to get a simple white background.
This is just how raw finetunes work anon. It was the same with base Illust and base Noob, you were just using mixes and merges.
>>109002356 and if I don't prompt a background it should be an amalgamation of all backgrounds that fit the image.
>It was the same with base Illust and base Noob, you were just using mixes and merges noob most definitely does not give an empty background and simple as shit colors by default
For the anons who say that Anima needs more steps at higher resolutions If 30 steps are necessary for 1024px, how many would be needed for 1536px? No, I'm not doing a rule of three calculation.
>>109002380 >it should be an amalgamation of all backgrounds that fit the image. It should be an amalgamation of the average Danbooru background which is white, simple. >noob most definitely does not give an empty background and simple as shit colors by default Base Noob requires you to be incredibly autistic with the prompt, just like Anima. You'd know this already if you used it a lot.
So is Ideogram 4 edit-capable or not? Locally that is. I can't find a solid answer anywhere. The Ideogram website has editing but it's not clear what model is actually used for it.
>>109002502 >I can't find a solid answer anywhere. go to the official ideogram 4 github repository and comfy pages, the answer will be there (as in, if they don't mention it it can't)
>>109002571 /ldg/ is split into two groups, for 2 extremes of the iq bell curve low iq retards are too dumb to set it up yet and/or cant run it as fast as the other vramlet shit they are running. high iq kinnoisseurs know its just a slopped censored model with regional prompting and some more world knowledge thats not worth setting up.
LTX is very good at synthesising real people's voices based on tiny samples. Unfortunately their jew magic also turns you into a woman hating incel sexist misogynist
For lora training in Anima, is this the famous text encoder? I have it set to 0 because the model tends to forget things, right?. I also have Train UNet Only enabled.
>>109001866 this isnt really useful if you dont mention the native generation resolution as well as the sampler (some wont meaningfully converge even at 1,000,000 steps) anon
i get really disappointed when my posts dont make it into the collage. judge me all you like but thats the only reach i can hope to achieve with my gens
>>109002571 I'm not using because... I can't use it, I need to feed a online chatbot something for it to make a json for me to give to the model, and its slow af, its just... annoying, if it at least produced ultra high quality images
>>109002666 theres not that many images anyway and its more interesting when you can actually just quickly glance over everything instead of more than half of the gens being removed. there is enough pixels of space, no reason to not include everything.
>>109002746 see? cant engage -> derails by picking one: 1. why do you care 2. you are mad 3. random insult bonus if its a vaguepost in the thread without directly quoting too. every time lol
Anima seems very sensitive (negatively) to "slop prompts" where you spam duplicate shit. I think because it's better at following the prompt, this means you get worse images because it's trying to follow everything. Every word matters now.
>>109002788 you can't, thats the point. npcs have to pick between the 3 points every time, same reason you did so 4 times in a row and will again to this comment
>>109002796 >>109002684 >not that mant images 92 in the last thread >when you can actually just quickly glance over everything 4chanXT >no reason to not include everything. see >>109002666 and hdg highlights
I see 98% of the rest of humanity is still fucking retarded, arguing over fucking nothing. All day every day pure seethe, should just fucking kys desu.
>>109002803 isnt the 4chan limit 10k by 10k pixels or something? 92 images is nothing
also arguing that >that would make the collage less special meanwhile like 1/4 of the fagollage is currently a loss graph screenshot despite only 10% of the images from the previous thread being handpicked to be in it, lmao
>>109002571 Advertising doesn't really work on 4chan because nobody likes anything here... and as bad as the Ideogram spam is, it's nothing compared to the Nanobanana launch. Google was paying everybody!
>>109002656 >doesn't post a gen to help make his case I think OP made the right choice.
>>109002902 i don't mind that desu since the video models i use alter the skin regardless. I just need something that works and gives decent enough results that the video model can latch onto. But klein is good for image to image edit mode for make it more realistic sure and i often use it for that.
>>109002902 >>109002920 in fact my workflow uses klein 9b edit on last frame for each clip I gen using wan, but it changes the skin to much and does not always behave. >>109002932 yeah its always going to make their skin better and everything overall. But anima is a good little model that even has a controlnet model for it now, it would make a nice little last WAN frame restore model imo
>>109002902 it can get pretty good with the right samplers. and pid for qwen came out just this week too >>109002932 what the fuck is that tanline though
>>109002924 they will be waiting for the next commercial western hype model i think, then they will crush them. This is the Chinese's plan all along, they will destroy competition before they make any money from their own models.
>>109002985 did you use some sort of tagger? It very accurate, i find prompt klein a bit too hard or i can't be bothered because it takes an essay of just the subject to get right.
I know Wan video prompting perfectly, i can create entire minutes long videos with that thing because it easy once you know 81 frame 5 seconds action limit. But klein is awful to prompt for i hate it.
diffusion-pipe now supports Ideogram4. The model is very good for training actually. I gave early access to the code to a trusted acquaintance and he informs me that it's trivially uncensored just by training it. "it's basically less censored than z-image or klein when you train it" - him
>>109003014 thanks russ >he informs me that it's trivially uncensored just by training it. "it's basically less censored than z-image or klein when you train it" - him big if true
>>109003034 from my own experience with euler you tend to have to go up to 70 to 90 for convergence also try my yarat lora if you havent merged it in already, it was partially trained on 2048px
>>109003038 probably the text encoder, its great because it has tags but when they don't work you can literally just tell it in your own words and it usually gets the idea. They done an amazing job here.
>>109003029 train lora on 500 NSFW images and it gives much better results than z-image or klein 9b can achieve on the same dataset. even after just a few hundred steps you will never get "image blocked by safety filter" anymore >>109003042 With the latest commit, diffusion-pipe can train directly on the native comfy quantized weights, like fp8_scaled. I doubt any other training script can do this currently. For example, I believe AI-Toolkit requires the original checkpoint format released by Ideogram. It works very well with basic NL. JSON might be better idk, I can't be arsed to format a dataset like that.
>>109003073 sounds pretty fucking neat. thanks a lot >JSON might be better idk, I can't be arsed to format a dataset like that. i happen to have a larger data set with people based bounding boxes floating around anyway, just need to find the motivation to burn money on training in spite of the shitty license kek
>>109003094 >in spite of the shitty license kek what's up with all these shitty licences anyways? anima and this i4? What the hell is up with these companies?
>>109002972 they will wait for the next google video model and then BASED Bytedance will mog them with Seedream 6.0 and Seedance 3.0 API while Qwen releases a shitty happyhorse 2.0 that by all accounts should be local but they too put it behind API. Then Hunyuan comes along 6 months later with an LTX-tier video model with 4 text encoders that require 140GB VRAM to run train on seedream outputs. meanwhile kekstone discovered a revolutionary new way to train at 128x128 pixels to maximize flux.1 schnell training speed
>>109003083 >>109002867 >>109002865 this is what i mean when i asked yesterday: >>108998101 what's with the shitty ideogram shilling? it's the same thing again and again: "this model is NUTS if you know how to bypass the censor, just a little training and it gets real wild! with the right workflow you can generate ANYTHING!" but when asked to show NSFW results, nobody can deliver.
>>109003029 Klein is unironically better than Z Image out of the box at booba. Neither can do PP at all out of the box though. Hunyuan 2.1 and 3.0 can / could, though.
>>109003151 what do you mean? private lora commissions is literally one of the examples of what you cant do by tdrussell himself on the HF thread where he explains the license yes you obviously can circumvent it, doesnt change the fact its objectively worse than z-image legally speaking
>>109003139 >"this model is NUTS if you know how to bypass the censor, just a little training and it gets real wild! with the right workflow you can generate ANYTHING!" but when asked to show NSFW results, nobody can deliver. i know youre more describing the average plebbitor since no one here really cares _that_ much about i4, but this is just the way in which all free download models are spoken about when theyre new
>>109003162 >private lora commissions is literally one of the examples of what you cant do by tdrussell himself on the HF thread where he explains the license i see he mentions providers like civitai and fal not individual bakers getting commissioned
>>109003177 the difference is that tdrussell could probably enforce this. he wont but you have to rely on what comes down to a pinkie promise. not really a good thing if you want to actually make money from this also dont get me wrong i fully understand why it is this way, its still objectively worse than MIT or whatever and theres no reason to pretend otherwise >>109003188 https://huggingface.co/circlestone-labs/Anima/discussions/37#69af0cad85dd88d442416f99 >If you are locking down the lora and selling access to it, that is commercial use of a derivative model and isn't allowed. If someone commissions a lora that you then post publicly, I would say that it simply someone donating to you to help with training costs, and isn't commercial use of the model itself.
>>109003224 >If you are locking down the lora and selling access to it, that is commercial use of a derivative model and isn't allowed. If someone commissions a lora that you then post publicly, I would say that it simply someone donating to you to help with training costs, and isn't commercial use of the model itself. I'm not gonna release every lora I train. People who get them can upload them, I couldn't care less.
Hello, returning from a 4-month break here. Can we get an update on AnimAnon's OpenAnima project he said he was starting alongside LAX from Laxhar Labs, the creator of noob? I heard they were working on an Apache licensed cosmos-based model just like Anima, but it was going to be more "open for game developers to use". I looked around and couldn't find any info on it. It seems like LAX has joined ComfyOrg to work on Noob 2, and AnimAnon's UI hasn't been updated yet either. Can someone update me on the status of AnimAnon and what he is currently working on? I am looking forward to seeing his 4 months of progress towards an open diffusion model for all.
>>109003320 they pivoted again to some shit called sensenova. noob is dead. like most models, it was a one-hit wonder. the next big thing after anima will be developed and released by someone nobody has ever heard of, as was the case for every relevant finetune.
>>109003354 >the next big thing after anima will be developed and released by someone nobody has ever heard of, as was the case for every relevant finetune. this anon speaks true
i downloaded this one and another today, the above link imo is best. no lora used in that image eular_a simple 2048 squared 50 steps. 1 cfg because its a turbo model but it can easily do 4 cfg with 30 steps.
author states 1 cfg 16 steps max, but then you never really listen you just gonna test settings.
using some other detail or realistic lora's might give better results but i've not testing it fully.
>>109003401 trust me on that model, it is insane for what that little model can do. If lodestone was not such a kek he would just do a fine tune of this thing it would be mental.
i love ai surrealism. it really activates the creative part of your brain. also its important not to gen 1girls over and over. you risk falling into coomer mode, and it's hard to get out once you're in
>>109003502 its probably not even in its dataset as realistic photo that will be why, however if i did in anime style first and then forward to controlnet (it exists google it) and second stage to realistic i think it will work just fine.
Russell should have developed AniRealism first, then finetuned it for anime with Anima, and later created a UI called AniStudio. He would have made much more money, and the anime model would have had much better anatomical consistency.
>>109003569 could have been cool during the Victorian cat girl age but became a tranny schizo that has to hijak the OP to spite people she doesn't like. that started like 4 years ago and she still does it
stupid old cherry faced drunk man angry in his kitchen, threating the camera, propper nutcase. very short hair, scruffy beard. flushed face, angry red face, blood shot eyes, deranged, phycopath
4 cfg is interesting but too extreme, i will post that next, here is 2.5 cfg 16 steps
so easy to prompt.
but realism it can't do unless you prompt it right, so i need red skin not flushed face
>stupid old drunk man angry in his kitchen, threating the camera, propper nutcase. very short hair, scruffy beard, angry face, blood shot eyes, deranged, phycopath, (pale red skin:3)
>>109003664 >stupid old drunk man angry in his kitchen, threating the camera, propper nutcase. very short hair, scruffy beard, angry face, blood shot eyes, deranged, phycopath, (red skin:3)
did it, removed the pale which i used to anticipate the overly read face, so now i just need to lower the cfg. yeah i spelt threatening wrong so fuck.
>stupid old drunk man angry in his kitchen, threating the camera, propper nutcase. very short hair, scruffy beard, angry face, blood shot eyes, deranged, phycopath, (red skin:3), mouth wide open, irate, shouting loudly, teeth showing
>stupid old drunk man angry in his kitchen, threating the camera, propper nutcase. very short hair, scruffy beard, angry face, blood shot eyes, deranged, phycopath, (red skin:2), mouth wide open, irate, shouting loudly, teeth showing
lowered strength from :3 to :2 on (red skin) done it, this is how most real British men feel about their politicians right now.
now do that with baby klein the safety model lol fuck off. I will be surprised if anon can do that same angry expression in the time i did with klein 9b. I think it could do it but it would look fake and cringe, like a pathetic cherry picked stock photo cringe you can find using google images. But what ever keep using that cucked model if you want, we want the real deal though.
>>109003425 I got curious about how close this anima merge could get to realism. So I hit it with my choice of anime settings and lora enhancers and upscaled it. It came out better than I expected.
you don't like us because our memes are linguistic and compact way to compress ideas, concepts etc. all you jerk offs create are cartoons because your fucking heads are full of cartoons, you = turd do not fucking (you) me faggot.
i can teach so much about wan prompting for example because that is what i mostly worked with over the last year or so. 81 frames no more because it will loop back on it self and mess shit up, that is given and should already be know, the official prompt guide explains the prompting but it does not explain the 5 second limit. and one action, you can only usually get away with 1 action per clip it looks natural but its the best we got without it producing body horror.
people take little time to understand limitations of models, they all have limitations.
>>109003827 that is a really clean gen btw, the detail is superior to pdxl, close to klein. I noticed anima does well with detailed prompts, but has a lot more flexibility. Shorter prompts produce less detailed results.
>>109003929 it is impressive actually, i'm trying to at least get that level of detail on the first pass as you have done. My images look a bit blurry when you zoom in, you're has skin detail still when zooming in. The upscale looks bang on, what upscale method did you use? I suspect some upscale model on tiled or? It does not like anima was used on each step of the image gen.
>>109003929 >>109003954 and do note i try to avoid massive workflows but if you have one i can look at it and probably strip out 10 or more nodes because mostly its not needed.
>>109003954 For the upscaler I used nvidia-vfx. A damn near perfect upscaler. You can also use SeedVR2 but it way more slow. To get around Anima's limits, I also use MultiDiffusion in img2img. I don't have a workflow to share since I use Forge Neo.
You can find many on Hugging face using sha256... And that's really shitty lazy merge. SubtleShader is probably one of the worst piece of shit in the creator community. This scammer delete almost any negative comment and black list users saying how bad his merge are bad.