Comments (4)
Hi Sébastien, our research lab currently adopts a non-commercial license for research projects so it might not be suitable to directly use our codebase for commercial products. I'm closing this issue for now. If you have any further questions, feel free to start a new one or reopen this issue. Thank you!
from tokencompose.
Hi Sébastien, finetuning SD1.4 is pretty fast and can be done with a single RTX 3090 (with a batch size 1, grad. accumulation of 4 and grad. checkpointing).
We also provide SD 2.1 checkpoints finetuned on 768*768 resolutions on A6000 GPUs (requires more than >24gb vram).
Unfortunately we don't have checkpoints trained on SDXL because it's too large and likely requires GPUs such as A100. Our training pipeline requires more vram usage compared to a typical training pipeline because we have additional objectives on the cross-attention map.
If you want to train SDXL, my guess would be:
(1) increase the number of optimization steps (we observe that we need to increase this when we train 2.1 compared to 1.4)
(2) change the grounding loss ratio, e.g.,
from tokencompose.
Thanks for your insights ! I'm currently trying to reproduce your training on a finetune of SD1.5 but I'm having some errors about missing text keys in the data, I'll keep digging !
About SDXL, I'll make some tests but if you're willing to work with me on this, I can probably provide cloud GPUs (depending on the price it would cost of course). We could then opensource it, I'm seeing a lot of value to this!
from tokencompose.
Oh, I just checked your license and it's non-commercial
from tokencompose.
Related Issues (8)
- Could you please provide the metadata.jsonl or any other .json file used for training? HOT 1
- Composition capability of SDXL HOT 1
- GPU memory usage HOT 2
- Why batch size could only be 1? HOT 2
- compatibility with lora HOT 2
- Could you please provide the training log? HOT 4
- May I ask on which GPU did you train for how long HOT 1
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from tokencompose.