Dev-bf16
So, Im having a hard time figuring out how I can use the Dev-bf16 transformers.. I tried just plugging them in with available comfy workflow but realized that those sigmas/steps are for the comfy-int8 files and likely wont work. Does anyone have any ideas on how to put together a functioning workflow for the dev-bf16 transformers???
Hi,
The official workflows are all designed to work with the distilled version (bf16 OR int8). The dev version will give you better details and versatility but it requires 30 steps of denoising. you can apply the distilled lora on top of it (plug a load lora node after the diffusion model loader) to get faster generation with 8 steps.
Can you share more details about what you're trying to achieve by using the dev-bf16 transformers file?
Hey, thanks @adamoster for gettin back at me.. So, the distilled models definitely swap with the int8 on the current workflows. But I was trying to figure out how to set up the the full-dev model that uses the steps rather than the sampler with manual sigmas.. I remember I had one for ltx-2, but I can't remember for the life of me how that workflow was setup.. It was definitely a more complicated node graph, or at least the one I remember.