What's your end goal for 10Eros-Max?
I dig what you're attempting to do here and I'm eager to see where you take it. But I'm curious, what made you think of going down this path, and what are you ultimately aiming for?
I guess another way of putting it is: What does success look like to you for this model?
With sulphur finetune happening I'm trying to restore what the LTX2.3 model was supposed to be as a shell for that to then incorporate into it, and also this work towards a base model that will always be something different v.s. what Sulphur official will release as. That end goal was always offline Grok imagine video essentially and this one was way closer since it's similar architecture with probably the best reinforcement learning pass in an open weights model. Every lora train or tune run is going to unwind all that quality and reinforcement learning and I'm applying what I know to find how to change the model around that and pretty much have the whole picture with how the model works while I've been doing this. This would've been already done if they had released documentation I've had to probe up and down and compare it to other models and use Claude to actually understand it's function. Plus idk how many others are grafting Krea2 and Wan and LTX into the model, so why not.
I didn't even consider that it's possible to graft all the other ai models into one. Definintely interested to see how this turns out.
It was not really possible cross-model to this extent before, it's a quirk of the H3 model. I was doing this grafting math scripting on LTX with LTX models and configurations. It's just luck that H3 is setup well and it has kind of an independent structure where the attention and stuff live above most of the multimodal functionality. I only got the idea from the reddit post that did cross graft from Wan2.2 to LTX 2.3, but I do it differently than that method. If you this on other models it would just break them. Even though H3 is essentially perpendicular to all of these models, full 90 degree shaping, except for actually a few parts of LTX2.3 where it's only 70-80 degrees different. It means you can do slight angular adjustments to the direction in H3 introducing finetune level changes, but the key is always how you apply, where in the model, and what you take it from, etc. I wouldn't recommend doing it, it's not easy to get it into a shape where the model is overall improved, most of the time the effect is slight or detrimental to quality. I only upload versions that have gone through like 20-30 revisions and still can do decent outputs. Also, every single attempt is a 39g model write, just for it to fail one sampling test and be deleted, not good for your SSD.