Article
Differential Transformer V2
•
34
Matricardi Fabio
FM-1976
AI & ML interests
control system engineering, AI, LLM with python. ThePoorGPUguy on substack
Recent Activity
upvoted
an
article
about 6 hours ago
Differential Transformer V2
liked
a model
about 6 hours ago
OlegSkutte/FLUX.2-klein-4B-GGUF
liked
a model
2 days ago
AbstractPhil/tiny-flux
Organizations
None yet
upvoted
an
article
about 6 hours ago
liked
a
model
about 6 hours ago
liked
3
models
2 days ago
reacted to
AbstractPhil's
post with 🔥
2 days ago
Post
937
Meet FluxLailah;
AbstractPhil/tiny-flux-deep; 220m Flux variant currently pretraining at BF16. She is experimental, does not produce solid images yet - and yet she is producing. There is both an EMA and a raw weights pair producing different images. The EMA is particularly interesting at times.
Lailah uses flan-t5-base, clip-vit-l-14, and BlackForestLabs Flux1s VAE.
SEQ limit 128, images 512x512 for now. Lailah's early form is based on three variants. TinyFlux's weights were carefully planted into a deeper structure and trained yet again - dubbed TinyFlux-Deep. This variant has 15 dual-stream blocks and 25 single-stream blocks, nearly identical weight code as Flux with a similar attention mechanism - but intentionally deviant and compacted with careful consideration to scaling and purpose of mechanisms.
She went through quite a few growing pains with her earlier attention mechanism which required a reimagining today and careful consideration of the consequences, and now I present to you the preliminary look into Lailah.
The preliminary training is still heavily under way, the mechanisms are still being augmented, and her stability is currently being measured. The potential for fidelity, depth, and quality are still in measure - so I will be shifting attention and pivoting utility based on the needs over time.
Lailah uses flan-t5-base, clip-vit-l-14, and BlackForestLabs Flux1s VAE.
SEQ limit 128, images 512x512 for now. Lailah's early form is based on three variants. TinyFlux's weights were carefully planted into a deeper structure and trained yet again - dubbed TinyFlux-Deep. This variant has 15 dual-stream blocks and 25 single-stream blocks, nearly identical weight code as Flux with a similar attention mechanism - but intentionally deviant and compacted with careful consideration to scaling and purpose of mechanisms.
She went through quite a few growing pains with her earlier attention mechanism which required a reimagining today and careful consideration of the consequences, and now I present to you the preliminary look into Lailah.
The preliminary training is still heavily under way, the mechanisms are still being augmented, and her stability is currently being measured. The potential for fidelity, depth, and quality are still in measure - so I will be shifting attention and pivoting utility based on the needs over time.
- 2 replies
liked
a
model
2 days ago
reacted to
wangbuer999's
post with 🔥
3 days ago
Post
2609
HunyuanImage 3.0-Instruct just dropped
fresh -sourceImage 3.0model! Spent 20 mins testing it on a Messi + retro scrambler fusion case
Ran on diffusers v0.26.3 + CUDA 12.1 | 8B MoE params (1.3B activated) | zero VRAM issues
strength=0.9 Messi #10 kit/tattoo sharp, moto’s rusted metal texture blurred (classic open-source pain)
strength=0.7 Moto/cobblestone background crisp, Messi’s jersey details faded completely
strength=0.75 + prompt "Blend seamlessly, keep all original details": both subject & background sharp
No ControlNet, no manual masking the model’s chain-of-thought reasoning parses image+prompt first
Already outperforms Qwen-Image-Edit 2511 (GSB eval +25.7% on single-image edits) | 100% open-source
👉 Repo: https://hunyuan.tencent.com/chat/HunyuanDefault?from=modelSquare&modelId=Hunyuan-Image-3.0-Instruct
technical report：https://arxiv.org/abs/2509.23951
Anyone else struggled with strength tweaks for fusion? This fixed it for my Messi+moto case did it work as well for yours?
fresh -sourceImage 3.0model! Spent 20 mins testing it on a Messi + retro scrambler fusion case
Ran on diffusers v0.26.3 + CUDA 12.1 | 8B MoE params (1.3B activated) | zero VRAM issues
strength=0.9 Messi #10 kit/tattoo sharp, moto’s rusted metal texture blurred (classic open-source pain)
strength=0.7 Moto/cobblestone background crisp, Messi’s jersey details faded completely
strength=0.75 + prompt "Blend seamlessly, keep all original details": both subject & background sharp
No ControlNet, no manual masking the model’s chain-of-thought reasoning parses image+prompt first
Already outperforms Qwen-Image-Edit 2511 (GSB eval +25.7% on single-image edits) | 100% open-source
👉 Repo: https://hunyuan.tencent.com/chat/HunyuanDefault?from=modelSquare&modelId=Hunyuan-Image-3.0-Instruct
technical report：https://arxiv.org/abs/2509.23951
Anyone else struggled with strength tweaks for fusion? This fixed it for my Messi+moto case did it work as well for yours?
- 6 replies
liked
a
model
4 days ago
liked
a
model
5 days ago
Can you also make zai-org/GLM-4.7-Flash same as you made this model?
🚀
9
2
#6 opened 9 days ago
by
dibu28
strange characters during deneration
1
#1 opened 5 days ago
by
FM-1976
liked
a
model
6 days ago
upvoted
a
collection
7 days ago
liked
a
model
7 days ago
liked
3
models
8 days ago
reacted to
MikeDoes's
post with 🔥
9 days ago
Post
3017
How do you prove your new, specialized AI model is a better solution? You test it against the best.
That's why we were excited to see the new AdminBERT paper from researchers at Nantes Université and others. To show the strength of their new model for French administrative texts, they compared it to the state-of-the-art generalist model, NERmemBERT.
The direct connection to our work is clear: NERmemBERT was trained on a combination of datasets, including the Pii-masking-200k dataset by Ai4Privacy.
This is a perfect win-win for the open-source community. Our foundational dataset helps create a strong, general-purpose benchmark, which in turn helps researchers prove the value of their specialized work. This is how we all get better.
🔗 Great work by Thomas Sebbag, Solen Quiniou, Nicolas Stucky, and Emmanuel Morin on tackling a challenging domain! Check out their paper: https://aclanthology.org/2025.coling-main.27.pdf
🚀 Stay updated on the latest in privacy-preserving AI—follow us on LinkedIn: https://www.linkedin.com/company/ai4privacy/posts/
#OpenSource
#DataPrivacy
#LLM
#Anonymization
#AIsecurity
#HuggingFace
#Ai4Privacy
#Worldslargestopensourceprivacymaskingdataset
That's why we were excited to see the new AdminBERT paper from researchers at Nantes Université and others. To show the strength of their new model for French administrative texts, they compared it to the state-of-the-art generalist model, NERmemBERT.
The direct connection to our work is clear: NERmemBERT was trained on a combination of datasets, including the Pii-masking-200k dataset by Ai4Privacy.
This is a perfect win-win for the open-source community. Our foundational dataset helps create a strong, general-purpose benchmark, which in turn helps researchers prove the value of their specialized work. This is how we all get better.
🔗 Great work by Thomas Sebbag, Solen Quiniou, Nicolas Stucky, and Emmanuel Morin on tackling a challenging domain! Check out their paper: https://aclanthology.org/2025.coling-main.27.pdf
🚀 Stay updated on the latest in privacy-preserving AI—follow us on LinkedIn: https://www.linkedin.com/company/ai4privacy/posts/
#OpenSource
#DataPrivacy
#LLM
#Anonymization
#AIsecurity
#HuggingFace
#Ai4Privacy
#Worldslargestopensourceprivacymaskingdataset
replied to
davidmezzetti's
post
9 days ago
is there a way to use txtai training pipeline also only using CPU?