FAQ about License: Fair AI Public License 1.0-SD
Some might not be familiar with the Fair AI Public License 1.0-SD. Here is a FAQ regarding the Fair AI Public License 1.0-SD.
Q1: Can the images generated by the AnimagineXL 3.1 model be commercialized/sold?
A1: Since the Fair AI Public License 1.0-SD does not specify about the generated output, we reference the CreativeML Open RAIL++-M license. For those unfamiliar, CreativeML Open RAIL++-M is the license used by Stability AI's stable-diffusion-xl.
According to CreativeML Open RAIL++-M, Section III.6:"The Licensor claims no rights in the Output you generate using the Model. You are responsible for the Output you generate and its subsequent uses. None of the uses of the output should contravene any provision stated in the License."
So, essentially, it is your choice to commercialize the output generated from the AnimagineXL 3.1 model.
Q2: Can the model that I fine-tuned or merged using AnimagineXL 3.1 be made exclusive at places that provide generation services like SeaArt, Civitai, Tensor art?
A2: Before answering, let's define what an exclusive model is - it's a model that is only available at one location and cannot be downloaded or generated elsewhere.
The answer to your question is no, the models you fine-tuning or merger using AnimagineXL 3.1 should be available for everyone, which is in line with the Fair AI Public License 1.0-SD that does not allow exclusivity.
Q3: I saw keypoint number 1 about Modification Sharing, how should I share modified results from Animagine XL 3.1? Could you give me an example?
A3: For an example of how to share your modified fine-tuning results, you can follow the Training config with some modification as below are example:
Fine-tuning from Animagine XL 3.0
Using kohya-ss commit : 6b1520a46b1b6ee7c33092537dc9449d1cc4f56f | or you can share your clone repo of kohya-ss you use for fine-tuning your model
| Configuration Item | Animagine XL 3.1 |
|---|---|
| GPU | 2 x A100 80G |
| Dataset | 873,504 |
| Shuffle Separator | True |
| Num Epochs | 10 |
| Learning Rate | 1e-5 |
| Text Encoder Learning Rate | 1e-5 |
| Effective Batch Size | 16 x 3 x 2 |
| Optimizer | AdamW |
| Optimizer Args | Weight Decay: 0.1, Betas: (0.9, 0.99) |
| LR Scheduler | Cosine Annealing Warm Restart |
| LR Scheduler Args | Num Cycles: 10, Min LR: 1e-6, LR Decay: 0.9, First Cycle Steps: 9,099 |
For merged results, below are example
If using MBW
Merge parameters for Your model merge name.txt
- Animagine XL 3.1 merged to animagine-xl-3.0 to get more aesthetic using MBW
(0,1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,0.0,1,0.2,0.3,0.4,0.5,0.6,0.7,1,0.8,0.1) - Your model merge name.txt
If using non MBW
Merge parameters for Your model merge name.txt
- Name Model A + ( ANIMAGINE XL 3.1 - ANIMAGINE XL 3.0 ) x 1.0, Add difference(trainDifference)
- [1] + ( Name Model B - ANIMAGINE XL 3.0 ) x 0.55, Add difference(trainDifference)
- Your model merge name.txt
Q4: Is it necessary to include the Fair AI Public License 1.0-SD when uploading a model that I have fine-tuned or merged, and should it be detailed as in https://huggingface.co/cagliostrolab/animagine-xl-3.1#license?
A4: Yes, it is necessary to mention the Fair AI Public License 1.0-SD so others are aware that the model you've fine-tuned or merged is under the Fair AI Public License 1.0-SD. Including a detailed explanation as seen in the license section for AnimagineXL 3.1 is recommended to ensure others understand the key points of this license.
If you have another question about License fell free to ask here ^^ or to our discord server https://discord.gg/cqh9tZgbGc