These are anime-style LoRAs for Qwen-image. Due to my lack of familiarity with creating LoRAs for Qwen, the time required for training, and lower-than-expected effectiveness, I'm splitting them into several LoRAs based on the style of the image and intending to use them in combination.
AnimeM1
This is a slightly cooler anime LoRA. Use it a little stronger.
AnimeL1
A very old school anime style LoRA. Unless you're aiming for an analog TV style, try using it around 0.75.
AnimeK1
This is an anime illustration style LoRA that is often seen on SDXL etc.
AnimeU1
This is a LoRA created using experimental techniques. The theme is to see how much automatic generation of style LoRA is possible. The training is limited to about one epoch, aiming for mass production.
1.0 is too strong, so use around 0.75 as a benchmark. While not a trigger word, animeU1 can be effective in some cases.
It is intended to be used in combination with other LoRAs in the series, and to extract effective components using LayeredLoraLoader.
am2_test
The first LoRA I created. Generated at Rank 16 with a miscellaneous collection of images. Resolution: 1280 dots.
Anime-H1
Generated at Rank 32 with a collection of clean, character-focused images with little background. Approximately 400 images were used. Resolution: 1280 dots.
Anime-R1
Generated at Rank 32 with images featuring relatively high line detail. 1230 images were used. Resolution: 1024 dots.
Note that while the resolution was increased during training using upscaling, the original images were still low-resolution.
Adjust the LoRA strength in conjunction with CFG scale, etc. There is no restriction on the total being 1.0 when using them in combination. In some cases, using a value of 1.5 or higher individually may be beneficial. As a general rule, if the CFG scale is high, adjust it by lowering the LoRA strength.
Example: If the amount of background information is decreasing, lower the CFG scale.
There are no trigger words, but "animeH1" or "animeR1" may be effective. Please make it clear in the prompt that the image is an anime drawing.
Description
FAQ
Comments (2)
Amazing work mate! is it possible for you train a dragon ball super anime style lora and write a tutorial on how to train character loras for qwen image/edit?
There are several methods, and they vary greatly depending on your environment (OS, GPU). A common approach is to collect images and captions. Searching the web is a good option. You can also output a large number of images using LoRA tools like SDXL.
If captions aren't available, generate them using a tagger. I use the wd-EVA02-Large-v3 with the A1111's WD14 tagger. I also use a captioning tool in Florence2. If you're not sure, you can ask ChatGPT to create them for you. The combined text file and image collection are then processed using the LoRA creation tool. If you want to specialize on characters, you'll need to add character names to the captions. If you're just styling, this isn't necessary.
I'm using the AI Toolkit, but I've only just started using it, so I'm not sure how to use it effectively. I think Musubi Tuner might have better performance, but I'm not sure how to use it. FlyMy.AI's tools don't work on Windows (unless you use WSL). If you know of any other tools, please let me know. Even the AI Toolkit gives me an error during installation on Windows. I found and installed Triton-windows, but it seems you can ignore it. The warning doesn't go away even after installing Triton-windows.
You can find instructions on how to generate Qwen Loras with low VRAM using the AI Toolkit by searching, so please refer to that. Note that low VRAM refers to 24GB. If you have anything other than 48GB or 80GB, you can assume that your VRAM is low. It may be possible to create LoRAs with less VRAM, but it will likely take a very long time.
With the AI Toolkit, outputting sample images takes time, so I recommend skipping the first sample, reducing the number of samples, and reducing the number of sample images.
If the data overflows the VRAM, it will take a very long time. In that case, you can reduce this by setting the LoRA image size to 512 or so.
I'm still experimenting with the number of images and steps used. A few dozen images and a little over 1,000 steps will work fine. For character LoRAs, even fewer should be fine.
I don't know anything about LoRA for Qwen-imade-Edit. Are there any documents that explain the control images?