AlbedoBase XL Pre
The refiner is unnecessary, and VAE is included.
Please preview the next version of the AlbedoBase XL model and provide your feedback here. Once i gather and incorporate all the feedback, i will release the official version to the AlbedoBase XL side.
v2.1pre-beta
I revised the merging process and incorporated four checkpoints, originally designed for enhancement, throughout the system.
v2.1pre-alpha
I revised the merging process and incorporated four checkpoints, originally designed for enhancement, throughout the system.
The spec grid(424.4 MB): download
v2.0pre-alpha
The algorithm has been changed, and the DPO has been excluded. This is a test 2.0 version.
v2.0pre-DPO
- The Merge formula has been updated.
- The DPO model has been added.
- Full utilization of negative prompts is now available.
- Support for Clip Skip 2 has been introduced.
- Enhanced detail clarity for all samplers is achieved as the step count increases.
The spec grid(424.3 MB): download
v1.6pre
I've changed the algorithm of the Merge itself.
v1.5pre
I have written a custom script to converge the existing AlbedoBase XL models into one. Intricately aligning the row and column weights of all U-NET and CLIP blocks according to a unique equation, may produce a merging effect different from previous methods. It is still experimental, and I am so serious that I am immersed in the challenge of writing a paper independently (though I do not have a master's or doctorate in the field).
The spec grid(399.2 MB): download
I enjoy this work, and my curiosity and passion will not cease even if a better model comes out and AGI is developed.
Description
- The Merge formula has been updated.
- The DPO model has been added.
- Full utilization of negative prompts is now available.
- Support for Clip Skip 2 has been introduced.
- Enhanced detail clarity for all samplers is achieved as the step count increases.
FAQ
Comments (10)
@albedobond, I am relatively new in this SD thing forgive me if I am not making sense here. Is it possible to merge SD/OpenDalle in your XL model? Asking because OpenDalle is very good at following prompts, almost no need regional prompting etc and generating VERY good images. Thank you.
If it's an SDXL model, any model can be merged into my AlbedoBaseXL model! Feel free to merge and if a decent model emerges, please share it on Civitai~~ Then, I can merge that model again, and in this way, we all can make progress.
@albedobond it's SD and I am pretty new, not even sure how to merge or create a model and or lora. Tried training loras on Civitai's trainer... No joy.
I have a new M3 mac, planning to buy something with decent card and run on Linux for SD.
@modusCell Merging models doesn't require a significant amount of computation. It primarily requires a substantial amount of VRAM. Regarding merging, I recommend starting with simple weighted merging, testing this and that model first. Then, move on to the SuperMerger extension in WebUI and try various things!
@albedobond Thank you, I will give try not on my computer but coLab.
Yeah, I think the other versions look better. I also tried merging the DPO with my models but in all my tests the model always came out slightly worse even with subtractive (difference) adding in the visuals. Although, DPO makes prompts behave better so it is a tradeoff I guess.
Yes... I felt that too. So, I tried overriding it by merging the 13th version last in the merging order, and it seems to have improved somewhat that way. Then, I will also upload a version excluding DPO. Thank you for your input.
@albedobond That is a great way of getting the quirkiness of DPO out! :P I'm looking forward to it!
I've only played with this for a couple of hours and used previous prompts but I can already tell that it can follow prompts a bit better than other models or previous versions. I like it.
Thank you for testing every time :) I have created several versions of 2.0 and am testing them, but it seems that a model that is satisfactory in all aspects is not quite emerging yet. Thank you.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.