Flux Kontext Dev GGUF workflow and Experiments ComfyUI
GGUF Flux Kontext Dev workflow
- used the comfyui template kontext dev workflow
- using q4_k_m quantized model
- replaced the unet and dual clip loader with the gguf ones
- also using load image from url (mtb) node to use url of the rabbit image
Performance: 12 secs/iteration , 20 steps, 303s, 5mins 3 secs, 03:53
Basic Style Transfer
Original input image
Default style transfer output
hyper flux 8 step lora tests
- after applying the hyper flux 8 step lora, 8 steps, 0.15 strength
0.15 strength, after applying the hyper flux 8 step lora, 8 steps
hyper flux, 0.3 strength
0.4 strength, 3.5 clip, instead of 2.5
No reference latent conditioning
- changing the strength to 0.6 gave noisy image
- removed the Reference latent conditioning, and the output changes, so theres defnitely some effect.
Remove object
- removing object seems to work well even when using the hyper flux 8 step lora.
Original image from arXiv paper
Removing object works flux kontext dev, hyper flux 8 step lora
This is still W.I.P. expect more things.