Merged Flux.1 Krea Dev Scaled FP8 with CLIP_L and T5XXL as well as the VAE into a single checkpoint file. Suitable for use with ComfyUI Checkpoint Loader node.
Description
Created with the Save Checkpoint Node in ComfyUI
FAQ
Comments (6)
Be warned- scaled FP8 has major issues on Blackwell (and maybe older architectures). In A-B testing, both main model and clip were giving bad results when scaled FP8 was used against GGUF. The 'bad' result was poor motion and poor prompt understanding - and I fear recent ComfyUI Blackwell FP8 'improvements' may be to blame (faulty accumulation of maths calculations degrading numerical accuracy).
About a year back Forge had a similar problem- the coder completely messed up the maths of adding a LoRA to a model. The problem is the devs assume that if the render image shows no 'corruption' the maths is 'working'. Forge did not corrupt images, but the LoRAs simply had no decent effect. When Scaled FP8 fails on Blackwell, likewise there is no visible frame corruption- just ignored prompt detail, and vastly reduced motion!
looks cool
how much vram do it take to have the unet completely loaded.
And how fast does it go (1.5i per secondes) ???
how much vram do it take to have the unet completely loaded.
And how fast does it go (1.5i per secondes) ???
Thanks :)