Article
Training Design for Text-to-Image Models: Lessons from Ablations
•
49
Pierre Colle
piercus
AI & ML interests
HD AI
Recent Activity
upvoted
an
article
3 days ago
Training Design for Text-to-Image Models: Lessons from Ablations
replied to
their
post
24 days ago
The Loss is a field with holes 🕳️⛳️
🛞 Batch Size is your wheel size.
🏎️ Learning Rate is your speed.
Organizations
upvoted
an
article
3 days ago
replied to
their
post
24 days ago
But higher LR makes the path more volatile/unstable ? @nroggendorff
posted
an
update
24 days ago
Post
256
The Loss is a field with holes 🕳️⛳️
🛞 Batch Size is your wheel size.
🏎️ Learning Rate is your speed.
🛞 Batch Size is your wheel size.
🏎️ Learning Rate is your speed.
-
- 3 replies
posted
an
update
3 months ago
Post
3971
Starts erasing! 🎉 🎉 🎉
This is made with a one-step SD1.5 LBM [1] eraser !
Data is open. Data pipeline is open. Training code is open.
On our LBM fork : https://github.com/finegrain-ai/LBM
[1] LBM: Latent Bridge Matching for Fast Image-to-Image Translation (2503.07535)
This is made with a one-step SD1.5 LBM [1] eraser !
Data is open. Data pipeline is open. Training code is open.
On our LBM fork : https://github.com/finegrain-ai/LBM
[1] LBM: Latent Bridge Matching for Fast Image-to-Image Translation (2503.07535)
- 1 reply
posted
an
update
3 months ago
Post
408
🚧 Reproducing LBM-Eraser… in the open [1] !
A major recent paper on erasing is OmniEraser [2].
They open-sourced an evaluation dataset [3] (and I'm using it for the evaluation of our LBM-Eraser 😉).
It's not a big dataset (70 samples), but it's good quality pairs, and that's what matters !
cc @BaiLing
[1] Finegrain LBM Fork : https://github.com/finegrain-ai/LBM
[2] OmniEraser: VDOR: A Video-based Dataset for Object Removal via Sequence Consistency (2501.07397)
[3] BaiLing/RemovalBench
[4] LBM paper: LBM: Latent Bridge Matching for Fast Image-to-Image Translation (2503.07535)
A major recent paper on erasing is OmniEraser [2].
They open-sourced an evaluation dataset [3] (and I'm using it for the evaluation of our LBM-Eraser 😉).
It's not a big dataset (70 samples), but it's good quality pairs, and that's what matters !
cc @BaiLing
[1] Finegrain LBM Fork : https://github.com/finegrain-ai/LBM
[2] OmniEraser: VDOR: A Video-based Dataset for Object Removal via Sequence Consistency (2501.07397)
[3] BaiLing/RemovalBench
[4] LBM paper: LBM: Latent Bridge Matching for Fast Image-to-Image Translation (2503.07535)
Image p_13.png seems to be moving between gt and images
#1 opened 3 months ago
by
piercus
posted
an
update
4 months ago
Post
849
🚧 Reproducing LBM-Eraser… in the open [1] !
Today we have trained a LBM [2] promptless inpainter using
We use a subset of 1.25M images with
2 takeaways :
🖼 Inpainting is better compared to our RORD experiments [5]
🦶 "4 steps" outperforms single-step
[1] Finegrain LBM Fork : https://github.com/finegrain-ai/LBM
[2] LBM: Latent Bridge Matching for Fast Image-to-Image Translation (2503.07535)
[3] supermodelresearch/Re-LAION-Caption19M
[4] Resolution-robust Large Mask Inpainting with Fourier Convolutions (2109.07161)
[5] https://huggingface.co/posts/piercus/778833977889788
cc @supermodelresearch @presencesw
Today we have trained a LBM [2] promptless inpainter using
Re-LAION-Caption19M[3].We use a subset of 1.25M images with
aesthetic_score > 5.6 and pwatermark < 0.2 and LaMa [2] mask generation.2 takeaways :
🖼 Inpainting is better compared to our RORD experiments [5]
🦶 "4 steps" outperforms single-step
[1] Finegrain LBM Fork : https://github.com/finegrain-ai/LBM
[2] LBM: Latent Bridge Matching for Fast Image-to-Image Translation (2503.07535)
[3] supermodelresearch/Re-LAION-Caption19M
[4] Resolution-robust Large Mask Inpainting with Fourier Convolutions (2109.07161)
[5] https://huggingface.co/posts/piercus/778833977889788
cc @supermodelresearch @presencesw
posted
an
update
4 months ago
Post
1860
🚧 Reproducing LBM-Eraser… in progress! [1]
When repurposing a T2I model into a pure I2I model, there’s always that orphaned text path — what do we do with it? 🤔
You can reuse it as learnable embeddings in multi-task setups [2], freeze an empty text prompt, distillate or prune the corresponding part.
In LBM, they take a clever route — zeroing [3] and reshaping [4] the text-related cross-attentions into self-attentions.
This gives you fresh weights for I2I computation, nicely integrated into your SD architecture.
📎 References
[1] Our LBM Fork: https://github.com/finegrain-ai/LBM
[2] OmniPaint: OmniPaint: Mastering Object-Oriented Editing via Disentangled Insertion-Removal Inpainting (2503.08677)
[3] LBM Zeroing: https://github.com/gojasper/LBM/blob/cafebc46a9ac16dcc61691d289cc4676b5c75380/examples/training/train_lbm_surface.py#L147-L148
[4] LBM Reshaping: https://github.com/gojasper/LBM/blob/cafebc46a9ac16dcc61691d289cc4676b5c75380/examples/training/train_lbm_surface.py#L100
When repurposing a T2I model into a pure I2I model, there’s always that orphaned text path — what do we do with it? 🤔
You can reuse it as learnable embeddings in multi-task setups [2], freeze an empty text prompt, distillate or prune the corresponding part.
In LBM, they take a clever route — zeroing [3] and reshaping [4] the text-related cross-attentions into self-attentions.
This gives you fresh weights for I2I computation, nicely integrated into your SD architecture.
📎 References
[1] Our LBM Fork: https://github.com/finegrain-ai/LBM
[2] OmniPaint: OmniPaint: Mastering Object-Oriented Editing via Disentangled Insertion-Removal Inpainting (2503.08677)
[3] LBM Zeroing: https://github.com/gojasper/LBM/blob/cafebc46a9ac16dcc61691d289cc4676b5c75380/examples/training/train_lbm_surface.py#L147-L148
[4] LBM Reshaping: https://github.com/gojasper/LBM/blob/cafebc46a9ac16dcc61691d289cc4676b5c75380/examples/training/train_lbm_surface.py#L100
- 2 replies
posted
an
update
4 months ago
Post
3149
We have trained a LBM-Eraser with RORD-Dataset in the open 🔥
🚀 1-step only inference, no distillation
🪶 Light backbone :SD1.5
🧠 Light training : converge in 6k steps
Now let's improve this, especially the inpainting capabilities. Stay tuned for more :-)
LBM paper : LBM: Latent Bridge Matching for Fast Image-to-Image Translation (2503.07535)
Our LBM fork : https://github.com/finegrain-ai/LBM
🚀 1-step only inference, no distillation
🪶 Light backbone :SD1.5
🧠 Light training : converge in 6k steps
Now let's improve this, especially the inpainting capabilities. Stay tuned for more :-)
LBM paper : LBM: Latent Bridge Matching for Fast Image-to-Image Translation (2503.07535)
Our LBM fork : https://github.com/finegrain-ai/LBM
posted
an
update
4 months ago
Post
219
In LBM paper, the noise and the conditioning image are merged into a single composite image.
Unlike other inpainting methods (which typically grey-mask the missing area), LBM replaces the masked region with uniformly sampled random pixels.
Intuitively, since LBM is trained from a text-to-image (T2I) model, those random pixels act as a strong signal to the pretrained model — essentially saying: “This is where you can do your generative magic.”
LBM Paper: LBM: Latent Bridge Matching for Fast Image-to-Image Translation (2503.07535)
Our fork (work in progress): https://github.com/finegrain-ai/LBM
Unlike other inpainting methods (which typically grey-mask the missing area), LBM replaces the masked region with uniformly sampled random pixels.
Intuitively, since LBM is trained from a text-to-image (T2I) model, those random pixels act as a strong signal to the pretrained model — essentially saying: “This is where you can do your generative magic.”
LBM Paper: LBM: Latent Bridge Matching for Fast Image-to-Image Translation (2503.07535)
Our fork (work in progress): https://github.com/finegrain-ai/LBM
Webdataset format ?
1
#2 opened 4 months ago
by
piercus
replied to
their
post
4 months ago
We are starting from the "fine-mask" / "Ours (scratch)" path.
Table below from LBM paper : https://huggingface.co/papers/2503.07535
posted
an
update
4 months ago
Post
166
LBM-Eraser reproduction on going ...
We are starting with RORD Dataset-only !
Our RORD -> webdataset extraction script : https://github.com/finegrain-ai/LBM/pull/4
RORD paper : https://bmvc2022.mpi-inf.mpg.de/542/
Thanks @presencesw for presencesw/RORD
We are starting with RORD Dataset-only !
Our RORD -> webdataset extraction script : https://github.com/finegrain-ai/LBM/pull/4
RORD paper : https://bmvc2022.mpi-inf.mpg.de/542/
Thanks @presencesw for presencesw/RORD
-
- 1 reply
posted
an
update
4 months ago
Post
2938
We've just forked LBM to reproduce the LBM eraser results
Our fork : https://github.com/finegrain-ai/LBM
LBM paper: LBM: Latent Bridge Matching for Fast Image-to-Image Translation (2503.07535)
LBM relighting demo : jasperai/LBM_relighting
Our fork : https://github.com/finegrain-ai/LBM
LBM paper: LBM: Latent Bridge Matching for Fast Image-to-Image Translation (2503.07535)
LBM relighting demo : jasperai/LBM_relighting
- 2 replies
upvoted
an
article
5 months ago
Article
Finegrain Product Placement LoRA (experiment)
•
5