TLDR:
This is a trolling model. int8, 16-step, light, fast, high quality, ns4w AI slop master.
I spent no more than 30 minutes creating this model, including collecting data and training it.
This model is to demonstrate how much time and effort you need to build this type of slop model. Do you need H100? No, if someone says they even used H100, you know they're talking nonsense.
Why is this here:
Got rage baited by random dude claiming that training AI slop and merging models also requires "LOTS" of time and efforts, even hours of H100. Therefore his ckpts are much better than other finetunes. Apparently he is lying so he can sell his ai slop model, and has no ideas what 1 hour H100 means.
So, how long do you need to make a AI slop ckpt?
This ckpt is a speed-run demonstration.
Grab 89 AI slop n**w images from Pixiv with downloader extension, ~5min. Took some time handpicking good looking images instead of just auto grabbing.
Auto caption with LLM. ~1min.
Train a LoRA on the cheapest 4090. 1600 steps. ~15min. Probably overtrained.
Merge above LoRA, and RDBT, and acc LoRA into ckpt in ComfyUI. <1min.
Total ~20min. How does the ckpt look? (hmmm, tbh looks not bad)
EDIT: Disclaimer:
I am not against AI style. I am against lying and deceiving people, pretending to be a "great" model trainer, with ai slop.
"great" trainer == gigachad == who has millions of images and use GPU clusters.
My 10k dataset only needs ~1 h100 hour for a full run. I'm not a "great" trainer.
FYI:
How to determine if a model has undergone extensive training:
Don't trust the creator. Don't trust anyone. It's internet.
Don't rely on model/LoRA size.
Don't rely on how strong its default style is. On the contrary, a strong default style indicates less training data and it was overfitted.
Instead, look at the breadth of knowledge the model possesses. For example, model knows thousands of new styles; model has unique effects to various completely different styles, etc.
If someone claims their model has been extensively trained, even using H100, and that the model has a strong default AI slop style, no additional knowledge, they are simply lying.
[this_is_the_taste_of_a_liar.webp]
Sharing merges using this model is not allowed. This "restriction" won't affect anyone. It's only aimed at those who steal others' models to sell.
Special trigger words and latent watermark included.
Description
FAQ
Comments (28)
you seem mad sir
Got rage baited by random dude claiming he used hours of H100 training AI styles so his ckpts are much better than other finetunes...Apparently he is lying so he can sell his ai slop model, and has no ideas what 1 hour H100 means...
Is this model remotely decent compared to model from big creators? I know you want to make a point but I only care whether it works
still use rdbt base, it's just a experimental model
This model was created as a method to troll other trainers and creators who post their paid models and claim they take a lot of time and effort to train/create. It does function (and reasonably well at that, the way I understand it), but is more intended to be the "shots fired" at some of the other pretentious "big" creators on here that simply produce a model that looks like every other Anima model and claim it to be something special.
A couple notable examples might be Alexds9 or ElectricDreams, who consistently post their "flagship" models as paid-only, but those models consistently produce inferior results to other models from lesser creators.
what a shame, tried it and it just looks the same as 99% of anima models out there
Bro... the title literally has "Al slop" in it.
The author uses this model to prove 99% of anima models out there are AI slop, styles in ckpt, and can be made under 20mins.
edit: https://civitai.red/models/2987749/rdbt-or-random-ai-slop?dialog=commentThread&commentId=1356565
I've rewritten the description, it's more clear now.
That's the entire point 🤣
I was surprised that you actually trained a model specifically just to fight a debate. To me, that’s quite fun.
Based on my testing, under prompts that explicitly include “@artist name,” it does in fact still complete the task relatively acceptably. However, with pencil sketches, it clearly has a tendency to come out overly smooth. Even so, I still think the AIslop you trained is actually a passable ANIMA model—at least it holds the lower bound. Some ANIMA models on CIVIT can’t produce satisfactory images even when “@artist name” is used.
In any case, in my humble opinion, although good training material can secure ANIMA’s lower bound (RDBT managed that), the upper bound is still limited by ANIMA’s own architecture. ANIMA can meet my current local deployment needs, but its increasingly outdated knowledge base and not-very-strong semantic understanding still make me wish for a better base model.
Just throwing this out there. Also, thank you for your work.
_ better base model
Qwen image 2.1
knowledge base is infinite, it can transfer multiple characters from reference sheets to any scene
if you don't have reference sheets, It can transfer most characters from a clean image to reference sheets
Now I mainly use anima for raw materials and styles refinement.
Tried it, aesthetic looks surprisingly great. I'm actually gonna keep it on my disk for hentai gentleman styles.
I hate the fact that the AI slop aesthetic has grown on me.
Maybe that is the next step in evolution? Human art are much cruder than AI
I am amused. Honestly though i didnt realise making a slop lora was so quick and easy, maybe i should try it
It's really not hard, but it's also really not as easy as he's saying to produce truly decent results. You do need to have an eye for good training material, and while auto LLM captions work they're not amazing either... Most LLM vision models don't really understand anime nsfw too well so asking them to caption hentai produces mixed results. That said, most popular models tend to be merges of other people's work anyway--very little training required to make merges...
RDBT AI Slop Mix vs Nova Anima vs MiaoMiao Harem who wins the slop war
Then I need a properly prepared version.
Lighting and env: RDBT
Character: MiaoMiao
Clear loser: Nova
To be honest, I don't understand why so many people like the Nova model. It's the worst slop model I've ever seen. It suffers from severe overfitting, extremely poor LoRa compatibility, and prompt adherence. All characters' faces and expressions are fixed. Also, I don't think the creator of the Nova model did any training. He never released a LoRA. It seems like a pure merge to me.
I wish that slopless models were this stable. Good job, but I like your base rdbt more. Waiting for 2.4 base (allergic to turbo distils).
lmao, nice
I know this is a troll model. Seriously, I think this model is actually extremely solid. Better than 95% slop models out here. If you find this model, you should download and try it.
Here is my humble feedback:
The lighting quality of this model is far superior to all other slop models. Basically, all slop models have unnatural lighting and colors. They always add meaningless white shiny reflections. This model gives soft and natural lighting that doesn't feel like it comes from slop dataset at all.
The prompt adherence is as excellent as the RDBT model.
The only downside is that this model likes voluptuous women by default. Might change character body shape a little bit. Kind of expected, nsfw dataset.
Slop quality can still be boosted by other LoRA, like https://civitai.red/models/2564225/style-greater-gpt-image-2. 0.5 strength is the sweet spot.
And it's surprising how high the quality of a 16-step distilled model can be.
If you have time please can you do a w4a8 and or nvfp4 Quant of Kroma v0.3.1 OPD: https://huggingface.co/lodestones/Kroma?not-for-all-audiences=true
Thanks
https://huggingface.co/jakethessmake/Krea-Kroma-v0.3.1-Turbo-OPD-ComfyUI-Quants/tree/main
@reakaakasky The W4A8 version is 9 GB, while pure W4A8 is around 7.5 GB. I guess the repo author wanted to maximize quality by using mixed INT8 blocks, but now it doesn't fit on an 8 GB GPU anymore
I'd really appreciate it if you could still try making a standard W4A8 quant without mixed INT8 blocks
Does offloading in ComfyUI still have huge performance drawback? I saw they added some crazy optimizations this month. I think nowadays offloading is basically free for most gpu cards. Not tested though. Even if have a 8gb gpu you can still use the 14gb int8 ckpt.
The "dynamic vram" feature.
@reakaakasky I tested the speeds of different W4A8 quants, and the one that weighs around 7 GB runs 2–3x faster than the 8–9 GB one. I tested this on the original KREA 2 right after it came out.
No matter how hard the ComfyUI team tries, there's only so much they can do about the PCIe 4.0 x8 bandwidth on an RTX 4060. Dynamic VRAM doesn't really help in this case, and can sometimes even make things worse...
just tested, simulated 8gb vram + 14gb int8, it works.
If you found offloading is somehow still 2-3x slower. I think dynamic vram isn't working on you end. If it is working, you can expect 10% slower, but definitely not 200%.
PCIE bandwidth is not the bottom neck.
You just need enough ram (no swap.









