https://civarchive.com/user/seruva19 uploaded the dataset for Moebius. Give him all the buzz for putting together such a good set.
Batch size 4, 24 epochs, 2112 steps. Learning rate: 2e-5. 12 hours running on 2 x rtx 5000 ada gpus.
Description
FAQ
Comments (6)
Nice, I like this style.
I appreciate the effort but the Qwen lora doesn't really seem to do much. I used it on longer detailed prompts and the image was almost identical, like if all i did was change the seed, even tried it at ridiculously high weights like 4. I then downloaded the first sample image you have and loaded it's workflow in comfyui. I ran it and got the same image you did. I then reran it without the lora and it retained the style and the image was almost the same so it seems like it's just using whatever built in knowledge the model is already trained on. You basically get the same results by adding "In the style of Moebius" to the prompt, with or without the lora.
True, Moebius is already heavily baked into the base model. The lora mostly changing color pallets to a darker pallet and adding a little more edge contrast and pencil work, but scene composition is not really changing all that much.
How many source images did you use?
If you go to https://civitai.com/user/seruva19 he uploaded the dataset under his Moebius model. I don't remember the actual number.
As posted previously this Lora does not seem to make a ton of difference, since the style seems to be already baked in pretty well, it changes nuances. But the end result looks really good, However, much of that is already from the base model, think of this Lora as some added refinement.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.