Blissful Tuner

[Sarania]

Hello and thank you for your suite of lovely models! I dropped by to mention that my Blissful Tuner now supports generation for Kandinsky5 Pro/Lite/Image models and LoRA training for Pro/Lite including experimental I2V training with first and first/last modes. Both can be done on consumer hardware even for Pro - I've trained LoRA for K5 Pro T2V sft 5s in 6-12 hours(depending on dataset) on my RTX 4070 Ti Super 16GB GPU with 64GB system RAM for instance and I can make a 768x512 121 frame video with Pro in ~25 mins. This is achieved through a combination of fp8 scaled quantization, torch.compile, and offloading techniques like block_swap. Inference can further be accelerated with Sage Attn, fp8 math or fp16 math and scheduled CFG optimizations as well. Previews are supported during generation using TAEHV for video modes and TAEf1 for image modes or latent2RGB for either too! All potentially accuracy decreasing optimizations are controlled with flags so you can balance quality against your hardware as much as possible. Blissful Tuner is a CLI video diffusion suite built upon kohya-ss's excellent Musubi Tuner and includes lots of extended optimizations and upgrades beyond the base tuner for several models including Kandinsky5. I put a lot of effort into Kandinsky5 specifically because I'm quite fond of these models so keep up the great work!

[Sarania]

Update: T2I and Image Edit LoRA training are now also available!