Discussion on qualities and the ta?
Based on my outputs, The acestep_v1.5_merge_sft_turbo_ta_0.5.safetensors is like ~30% better than turbo.
Note that I used the Qwen 1.7b text encoder model;
And Also Note that the ~30% is a guess, hence not based on a standard nor analytical experiment;
It just feels like 30 percent better than the turbo...
I recommend finding the best ta for the merge.... !?!?!? maybe??!?
The Turbo SFT 0.5 merge is incredible and a noticeable improvement over Turbo. Less crust, less wrong notes and better structure, I can really recommend it.
What about sft model? Which one is better?
And it seems like this model doesn't work correctly with loras trained on turbo. Just tried, most of the result are distorted.
And it seems like this model doesn't work correctly with loras trained on turbo. Just tried, most of the result are distorted.
That would be somehow reasonable! The weights are different, because the mean of the sft-turbo-ta-0.5 model is better(different) than the Turbo model; hence the lora is not optimized to work with the sft-turbo-merge .
The good news is the lora trained on turbo, can be further trained(at a lower time and cost) to work with the sft-turbo merge.
I don't have the hardware to try with so this is just a theory
Would be cool if you plan on doing a merge of xl sft and xl turbo? https://huggingface.co/Comfy-Org/ace_step_1.5_ComfyUI_files/tree/main/split_files/diffusion_models
Would be cool if you plan on doing a merge of xl sft and xl turbo? https://huggingface.co/Comfy-Org/ace_step_1.5_ComfyUI_files/tree/main/split_files/diffusion_models
Would it possibly be a viable option to fix loras without retraining by simply loading and merging a lora to the sft model at the right strength, and also in a separate run, merge it to the turbo model at the right strength, then for both of them, extract the two resulting loras against those models, and then merging those loras together?
Could have it as an automated workflow in comfyui, or adapting the merge script to do it, but I don't know enough about ace-step yet to say for sure whether this'll work or not, hence why I'm asking.
Or, maybe something similar where you have the two parts of the turbo + sft model (so i guess the turbo model and the sft model separately lol), but before merging, and you merge the lora to each, then merge them, and then extract the lora? This one would likely heavily depend on the method and strength used for each merge step, I guess, but at least in my head, it seems like a fairly clean option.
For both options, if youre doing it in comfyui, you can test how it works before extracting the lora, too, which i'd say is pretty nice!
I'd test it myself, but my pc wants to explode atm so I usually use a cheap Vast AI instance to work on diffusion models, but I'm currently so broke due to some medical expenses, that I am going without food 2 days per week in order to just afford rent this month, so I'd love it if someone else could test this.
Sorry if some of the things I'm saying make no sense, I don't have much of a chance atm to actually get context for how this will work until I figure out a way to actually get my financial situation, umm... situated.