Independent ML researcher. The First step in knowing is admitting you don't
- Tweets 6,819
- Following 642
- Followers 2,120
- Likes 8,651
Pinned Tweet
AI visionaries tend to be. A dreamer who can not dream. They are utterly engulfed within their own doctrine that their daring stabs at the truth amount to moving numbers on a plot.
2
2
9
3,424
Have Fun <3! I merged 5 models together into something close to dalle3 level by accident! SDXL DPO is the main helper in achieving this! huggingface.co/dataautogpt3/…
20
43
346
70,933
𝙒𝙝𝙤 𝙣𝙚𝙚𝙙𝙨 𝙪𝙥𝙨𝙘𝙖𝙡𝙚𝙧𝙨?
The TempestV0.1 Initiative is a powerhouse in image generation, leveraging an unparalleled dataset of over 6 million images. The collection's vast scale, with resolutions from 1400x2100 to 4800x7200, encompasses 200GB of high-quality content.
With a groundbreaking 3 million iterations in its training cycle, TempestV0.1 underscores the rigorous effort input by its creator. This training intensity notably eclipses that of all other contemporarie models. TempestV0.1 shatters the conventional limits of image generation, particularly in delivering unparalleled detail and texture.
huggingface.co/dataautogpt3/…
33
26
173
17,134
You'd think with a forehead that size, he'd have room for more self-awareness than just 4Head chess moves
7
6
159
18,169
Literally the worst possible thing you could do with it lmao.
Automated Brain-rot while still completely gating the tool from artists with a forever stagnant waitlist.
If any other company on the planet was run like google they would go bankrupt overnight.
4
3
140
3,320
i just finished what i think was the last finetune.... I am not one to say this by my god this might be dangerous to open source! Hands down MJV6 level. this is not a joke or exaggeration i feel at this point.
The new proteus mobius model has:
1. The most diverse outputs
2. The strongest prompt control
3. The best aesthetic
Of any model I've used to date. I mean just look at the perspective on these!
ALT fish-eye wide angle view of an anime girl holding a baseball bat, highly exaggerated perspective, warped wide angle perspective, best quality, beautiful masterpiece artwork
ALT fish-eye wide angle view of an anime girl riding a skateboard, highly exaggerated perspective, warped wide angle perspective, best quality, beautiful masterpiece artwork
13
7
137
19,643
civitai.com/articles/5732/te… Civitai bans all use of SD3 related resources and content on their website
25
23
107
39,826
I have joined @Corcel_X as Head of Research! I will be working full time as a Applied ML Researcher/Head of their diffusion Research department. working on open sourcing state of the art Diffusion based models.
18
11
100
8,849
I implemented me own "prompt injection",
Now I can... prompt SPECIFIC BLOCKS of the UNet in different ways, And I implemented three different kinds of nodes for different workflow usages Each of these on the screenshot does the same thing, inject a specific prompt in OUTPUT0 and OUTPUT1, but the inputs/nodes is different
7
14
87
8,444
I am happy to present and announce huggingface.co/dataautogpt3/… ! a fine-tune of my OpenDalleV1.1 using 220k captioned gptV dataset to align and then dpo tuned on 10k dalle3 captioned image pairs. results are another step above OpenDalle in terms of prompt following and style.
6
12
84
20,907
huggingface.co/dataautogpt3/…… after some serious consideration and thought. I just open sourced both ProteusV0.2 and ProteusV0.1 for Commercial use under GLP 3.0.
6
19
81
9,782
We reverse engineered the ELLA training for 1.5 and successfully made a finetune of it. We are working on adapting the script to work with SDXL. major disappointment in them for not releasing it. So when in doubt do it yourself!
9
6
79
5,327
I'm shocked my first Lora even works, lmao. I just used the default settings and threw a 20k image dataset at it, assuming it would blow up.
SD3.5L is shockingly Stable! SDXL would have blown up.
also the image results are amazing considering I did even attempt to dial it in.
fixes the saturation issues for photorealism.
for more realistic results in SD3.5LargeTurbo try this config. First img standard, second with the attached config
12
3
71
7,324
The PixArt Sigma 900M really helps fix the issue of grainy faces that the 600M base had! still undertrained right now. going to talk and cover information about it and Prometheus with @HelloCivitai tomorrow during their podcast/interview with me!
2
5
71
3,728
This cycle has repeated every release.
Step 1. Angry Anticipation: "WHERE THE F*CK IS THE MODEL!"
Step 2. Release Rage: "THIS MODEL F*CKING SUCKS"
Step 3. Reversion Resignation "I'm going back to 1.4, it make t*ddies good"
Step 4. Furries F*ck is released "This model is so f*cking good, but why can't my GTX970 do 8k?"
13
3
66
5,933
Found way to handle the Loras onto the base checkpoint without causing a leaking on styles/concepts across each other.
Fluxteus is born!
6
4
67
3,666
Trained SDXL arch from scratch. Still very much undertrained atm.
This is NOT associated with Corcel, This is my own personal pet project.
7
4
66
3,725
Okay, I am now convinced a large amount of the OpenAI fanboys have actually lost their marbles now
2
60
4,493
In other news, I just finished Retraining SD1.5 for complete scratch! examples are not upscaled, more prompt adherent than SDXL for sure.
9
6
64
6,651
PrometheusV2 is SOTA. I am frankly convinced of that based on my testing over the last hour.
5
4
67
4,444
The 900M Pixart Sigma Model should be done cooking by this upcoming Wednesday. shout out to huggingface.co/jimmycarter for making his adaptation to the base model possible. (it produces noised images since its not trained)
1
3
61
2,036
Just achieved full rank finetuning of Stable Diffusion 3 Medium 2B on a SINGLE consumer GPU!
Only months ago, this was thought possible only on 80GB VRAM cards.
Now:
✅ Model: SD3 2B
✅ Hardware: 24GB VRAM RTX 4090
✅ Dataset: First layer of Proteus (prompt-alignment subset)
4
7
64
2,441
Found the issue that was causing the lack of detail. retrained the FLUX-syntheticanime lora overnight and its a lot better. going to redo megalith next. the quality is so so so so much better now.
5
3
58
1,892
sorry for the delay regarding the 900M Pixart Sigma model. Been extremely sick, never got hit with covid all the way up until now. went to the docs yesterday because it got so bad. they have me on quick special dial because i also have a autoimmune condition.
20
60
2,818
yo WTF,
even Edward Snowden is even shitting on SD3?!?!
Appalling that @StabilityAI so ruthlessly censored SD3 (for fear that someone might draw—gasp!—a nipple) that it nearly lost the ability to draw women entirely. Behold, "A woman lying in the grass:"
Fire the entire "AI Safety" team and replace them with an anatomy textbook.
10
3
53
3,652
WIP implementation of NovelAI V3's SDXL paper (arXiv:2409.15997v2)
To make things faster is is the Public repo. feel free to try and help me improve it while things are in the works
github.com/DataCTE/SDXL-Trai…
Status: Initial implementation of Sigma(∞) and ZTSNR training pipeline, testing with 10k dataset
#MachineLearning
4
10
57
6,453
this is a MASSVIVE deal
I trained a new VAE with 16x depth and 42 channels (kl-f16-d42). I am now training SD1.5 to work with it, which will double the output size of SD1.5 without much additional compute overhead.
Every time I train a new latent space, it always starts out inverted. It's so odd.
4
5
53
3,118
lady's and gents we are about to work on the models and bot full time with 3 employees under my company DataPusle AI. this is all thanks to investment by a undisclosed party (will be revealed in a bit)
12
1
53
2,403
Novel flux-dev conditioning method in ComfyUI:
Independently condition CLIP and T5-XXL with architecture-specific prompts to influence style and content.
Example:
T5-XXL (content): "a cat sitting on a bench drinking a latte in anime style."
CLIP (style): "cat, anime, anime style, 4k, stylistic, expressive"
this should be very easy to automate.
Key parameter: Flux guidance scale set to 1.8. Deviating from this value leads to output instability and degradation.
Leverages model-specific conditioning to disentangle style and content, enabling targeted control over each component.
Caveat: Potential for text distortion due to differential conditioning. Warrants further investigation and quantitative analysis.
found a cool way to condition flux-dev in a way that you can squeeze actual styles out of it! it messes with text tho
6
8
54
5,725
Okay this inference glitch is insanely cool,
Tried to do some tiled generation along with a vae scaling thing and somehow got this.
8
7
54
1,372
900M Pixart Sigma is almost done cooking. again it fixes to a decent degree the issue the previous model had with faces. (its not quite done yet)
2
1
52
1,686
Oh, also, btw I got Flux training working
First Lora will be a synthetic anime lora because I absolutely love the style! Btw, I found a way to even get a 32 rank lora to train on 24vram. instead of just 16 rank. I'll make a pull request
3
2
49
3,087
new arch just finished training on over 1.5 million images. results are looking promising. any questions are welcome!
7
2
52
3,523
🚀 Double release day! 🎉
1️⃣ PrometheusV1: huggingface.co/dataautogpt3/…
• First full rank finetune of Playground v2.5
• Optimized for open-source accessibility
• Custom CLIP integration (use clip skip 2!)
2️⃣ ProteusV0.5 is out too!
Pushing boundaries in #AIArt. Try them out and let me know what you think! 🖼️✨
#MachineLearning #OpenSource
2
9
51
1,843
Here is our working implementation of Training ELLA 1.5!
still working on getting it working for SDXL!
github.com/DataCTE/ELLA_Trai…
2
7
49
5,278
Does anyone want to collaborate on making a FLUX lora? I have some very, very well curated datasets. I have datasets ranging in size from 400K to just 120. The Flux training has been very finicky however. Please let me know!
10
48
2,185
📣 ProteusV0.5 Release:
🧠 Custom CLIP
🖼️ 400k+ dataset: Enhanced photorealism & style diversity
⚙️ Optimized LORA integration for targeted model improvements
HF: huggingface.co/dataautogpt3/…
Civitai: civitai.com/models/267242
2
10
51
4,962
We gave the @FAL early access to the upcoming Mobius model and its only been up on imgsys.org for 3 hours. its already the best stable diffusion based image model in the world based on human preference data
It's insane, my dude. Clear winner most of the times. If this thing is open sourced - you wll clearly be the best open source model provider on the market.
8
8
47
6,038
my friend Mobi who helped created the prompt injection node just made a clip switch node for SDXL clip to SD3! github.com/G-370/sd3-clip-sw… github.com/G-370/sd3-clip-sw… no weird requirements.txt, super light, extra simple, but make sure to read the README of the repo,
5
8
48
3,656
Our upcoming paper outlines and enables making entirely new base diffusion models without the need to extensively pretrain a new model from scratch. We can in a controlled way, break all the quality and composition associations without damaging the baseline understanding from the preexisting training. Making it so we can completely train those associations without having to start from scratch.
Mobius proved that basically, we just lowered the cost to entry for making new base diffusion models 50 fold
5
2
47
3,180
MobiusV1.3 has been achieved internally. still improving the model before release. improved anime and detailing overall without any compromise
6
2
43
3,027
working on getting a implementation of #SD3 working with a sponsor. We are fairly confident that we can get one trained. wish us luck! more announcements coming soon.
8
3
43
3,201
Pro tip for using the Mobius model on @huggingface:
For realistic photos, set the CFG to 3.5.
For artistic images, crank it up to 7.0.
Experiment with these settings to get the best results!
cfg 3.5 image below!
4
4
43
2,533
PrometheusV2.7 is live on my Discord bot!
please help me test it out!
(its arguable sota in terms of aesthetics and comp)
discord.gg/rZDB5kHFmx
3
5
43
2,537
samples from the current continued training run are looking great. 1B params, it should run on a potato. also MJ data was not used period. throw some prompts at me!
14
4
43
3,212
huggingface.co/spaces/dataau… thank you to @multimodalart for setting up this space for me!
5
43
2,237
I put a lot of effort into making the playground 2.5 arch/model backwards compatible with as many existing SDXL tools, loras, samplers, and etc. as possible! that's one of the key aspects of the model!
Prometheus from the good friend @DataPlusEngine seems to be working with IPAdapter which is interesting considering it's a Playground finetune. huggingface.co/dataautogpt3/…
2
7
44
1,548
We Will be Training SDXL 3B model from scratch aswell as a MMDIT (SD3) 15B model from scratch (sample from our sdxl finetune)
6
4
43
2,453
sorry for not being reachable or active over the last two months. i was in and out of the hospital due to medical complications/ a health scare. they found out the core cause and i am getting treatment for it now but i was bed ridden in the ER for about two weeks there
12
41
1,802
playing around with a custom sampler that has very fine control and i am loving it! choose the noise map and everything custom selection of dpm or eular per step
two phase sampling
etc
4
4
41
1,510
Moved the repo to here and i will be updating it with the 200k and than 12M when done
huggingface.co/dataautogpt3/…
5
4
38
7,206
and here is what i assume people actually want,
PrometheusV2-aesthetic. I would recommend using the refiner with it. also it has a trained trigger word of "aesthetic".
civitai.com/models/596963?mo…
6
5
42
2,179
ProteusV0.4: The Style Update huggingface.co/dataautogpt3/…
This update enhances stylistic capabilities, similar to Midjourney's approach, rather than advancing prompt comprehension. Methods used do not infringe on any copyrighted material.
6
9
39
3,905
Pretty sure this is the FINAL version of Mobius internal before release. look at the details!
2
2
37
1,862
finally after months of work i found a way to apply my techniques to 1.5! all of these images have less than 5 word prompts!
3
1
40
1,907
i am rather agnostic regarding timelines and tend to lean on the quote "When a distinguished but elderly scientist states that something is possible, he is almost certainly right. When he states that something is impossible, he is very probably wrong" bigthink.com/pessimists-arch…
1
1
37
6,310
