Been gradually catching up to the current SOtA after a hiatus. Got Flux running on runpod and made some images with the help of the two LoRAs kindly shared on CivitAI by other forum members, @messg and @therealD . Thanks both, have to congratulate you on the flexibility and transparency of your models: a great job finetuning!

Also currently trying to train a WAN LoRA via the method shared by @wammypinupart

Not planning to flood the place with my images, just thought I'd post a handful as a checkpoint.

I've got a hunch that video and image generation can be somewhat self-reinforcing in terms of creating realistic results, but the idea is a bit half-baked at the moment so I'll return to it when I've got something empirical to back it up.

Glad you found this useful. I only made the lora for flux because I made a dataset for training a wan lora and thought "why not?".

I've since made a better wan lora and might use to dataset for another flux lora - will put this on civitai

Awesome! Thanks for sharing your work, these are fun!


That's been my experience for sure. These Wan loras work on image-to-video generations too, so one thing to play with is taking a starter image (or last frame of a starter video) and getting your location, pose, and aesthetic set up that way. Stacking different loras together lets you do new things too, in both Flux and Wan.

Yeah image to video is absolutely what I'm targeting. Ideally I'd have all the mess instructions removed from the prompt for the source image, as that leaves a lot more context free to focus on everything else in the initial image without overloading.

God I love open source (and now open-weights) software!

Awesome! Yeah I doubt my first pass WAN lora will be much use to others as I didn't even faff with captions on the basis I'm happy to tolerate it being super opinionated. Initial experiments looking pretty good though for my purposes, and then I'll aim to generalise later.