■This is my testing ground for the wan2.2_5B LoRA.■My LoRAs are primarily trained on still images and are designed to be used with wan2.2_5B in a T2I → I2V workflow. They are also focused not on a single concept but on learning a wide range of unknown concepts.■I think the main feature of wan2.2_5B is that despite having 5B parameters, it remains lightweight and efficient, making both training and inference less demanding.■It looks especially promising as a practical option for video training, and static image training is also lightweight and accessible.Having both T2V and I2V integrated into a single model makes it compact and excellent. The architecture strikes a wonderful balance between quality and efficiency, giving me a strong sense of its potential. train unknown concepts is also very easy.■If you’re using wan2.2 5B for the first time, you can get the required models and workflows from the URL below—please give it a try.<a target="_blank" rel="ugc" href="https://comfyanonymous.github.io/ComfyUI_examples/wan22/">https://comfyanonymous.github.io/ComfyUI_examples/wan22/</a>■Detailed explanations for each LoRA are listed in the tabs. Workflows are also included, so please use them as a reference for inference.■In my testing, static image inference worked well even at higher resolutions such as 1920x1728 or 1280x1728, with very few issues, showing just how flexible the model is. For this reason, my workflow defaults to 1920x1280px. Vertical resolutions also produce good results, but I found that the horizontal aspect ratio, which is closer to video proportions, feels more appealing.■This model is also very good at fine details such as textures, but it does have some VAE-specific artifacts. Using i2i with models like SD1.5 to fix those is a good option. If wan2.2_5B handles the striking composition while another model refines the details, the results will be even better.My workflow also includes i2i with the models listed below, so please use them as a reference. Both realistic and anime styles are available, so feel free to choose whichever you prefer. Of course, if you have enough VRAM, SDXL is also a great option.<a target="_blank" rel="ugc" href="https://civarchive.com/models/1246353/sd15modellab">https://civarchive.com/models/1246353/sd15modellab</a>■Please feel free to ask if you have any questions!日本語での質問も大丈夫ですので気軽にお声がけください!■I trained the LoRAs using AI Toolkit. If you’re interested in training, the developer has created tutorials you can follow. You’ll find that it’s easier than you might expect.<div data-youtube-video><iframe width="640" height="480" allowfullscreen="true" autoplay="false" disablekbcontrols="false" enableiframeapi="false" endtime="0" ivloadpolicy="0" loop="false" modestbranding="false" origin playlist src="https://www.youtube.com/embed/9ATaQdin1sA?rel=1" start="0"></iframe></div>

wan2.2_5B_lora_lab

Training Data (44.97 KB):comfyui workflow■This is a LoRA trained on about 30,000 anime-style images. The base model is already strong, with anime aesthetics equivalent to or better than NovelAI v1. It also has a reasonable understanding of booru tags.■My focus was on further strengthening it and adding unknown concepts. Since it was trained with booru tags, it can be used for inference in the same way as other anime models. Because the training scale is relatively small, its capabilities are limited, but I believe it can still serve as an indicator of how much the model can learn.■There’s also a tipo workflow that automatically generates tag prompts, so please use it as a reference.■It seems that the metadata-saving extension in the workflow causes Chinese text to become garbled, so it wasn’t saved correctly. Since Wan uses Chinese in the negative prompts by default, that’s a bit inconvenient... Unfortunately, when uploading images to Civitai, you may need to paste the prompts manually instead of relying on the extension.

Files with SHA256: 03b6738b8f217cdefeb6ba4b6c7bd568fa6f1d10f970602cc4207d5edf17cb53