kudzueye

Status: Experimental/Testing Phase You should not expect great results from these LoRAs at the moment. It often involves using them together to get best results. I may be able to merge them later on for more consistent results, but I am uncertain of how good the actual quality will be. Workflow Documentation I recommend combining these loras together in whatever workflow you use. Have the general discrete lora set at a strength of 1.0 and the the portrait blend high rank model to something like 0.6 - 0.8. You will need to experiment a lot to figure out the best results. I am still uncertain of the best way to work with Qwen yet and how I should approach newer versions for this style. ComfyUI Example Workflow: https://huggingface.co/kudzueye/boreal-qwen-image/blob/main/boreal-qwen-workflow-v1.json --- This model is part of ongoing research and experimentation. Description and documentation will be updated as development progresses. You should use `photo` to trigger the image generation.

Boreal Flux Dev Version 2 This version of the Boing Reality flux dev LoRA used a different training approach for its dataset. It should also have the old latent shift dot issue fixed. This LoRA is very overtrained for how it works and the strength may need to be much lower than 1.0. It also works well with dynamic thresholding when you include a very high negative guidance. This version lacks some of the creativity of the older version, as I am still trying to switch up how I am approaching the datasets and training parameters. You should use `photo` to trigger the image generation. Weights for this model are available in Safetensors format.

This LoRA is an attempt at improving on the overall detail of generations from Hunyuan. It focus in particular on things such as improved depth of field, realistic skin texture, and better lighting. It works at generating both realistic short video clips and single frames for images. At the moment, the candidate LoRA used here is overtrained. It is recommended to use a strength of around 0.6. You will also need to experiment a lot with the seeds, guidance, steps, and resolution. Try to keep steps over 35 and minimum resolution above 512x512. Guidance seems to work to varying degrees between 3.5-12.5. Higher guidance and strength may lead to more similarities in things such as scene and characters. You will want to experiment with lowering the strength while raising the resolution and guidance when you run into lots of distortion. Also swap seeds as results vary a lot. IMPORTANT NOTES: This LoRA is still very experimental and difficult to control. Even with the settings recommended above it may still be difficult to get usable results. I am continuing doing new training runs to see what will help improve it. Also trying to see if there are any new improvements with an updated workflow. If you want an easier way of using this LoRA than with Comfyui, you might be able to try running it via Fal's Hunyuan video LoRA. Use this LoRA URL and set steps to 55 (pro mode) with 720p resolution. Expect to wait at least five minutes for it to run. Replicate should also have an option though I have not tested it out yet to verify the results. Image Examples The LoRA can also sometimes perform decently on images when you use the initial frame. Training Details I used around 150 images only for this initial LoRA. I focused on public domain photos from around the early 2010s. - epochs = 600 - gradientaccumulationsteps = 4 - warmupsteps = 100 Adapter - type = "lora" - rank = 32 - dtype = "bfloat16" - onlydoubleblocks = true Optimizer - type = "adamwoptimi" - lr = 0.0002 - betas = [ 0.9, 0.99,] - weightdecay = 0.01 - eps = 1e-8