Cat Experiment

Here the task is to zoom on the cat, I have taken the image and the prompt from the BFL Playground Examples.

Results Gallery

Comparison of different node configurations and their outputs:

Flux Pro Result

Baseline Configurations

Baseline

Standard Flux Kontext configuration

Baseline + Tea Cache

Tea Cache optimization enabled

Baseline + Hyper-FLUX

Hyper-FLUX LoRA acceleration

Baseline + Tea + Hyper

Combined Tea Cache + Hyper-FLUX

Detail Daemon Configurations

Detail Daemon

Detail Daemon enhancement

Daemon + Tea Cache

Detail Daemon + Tea Cache

Daemon + Hyper-FLUX

Detail Daemon + Hyper-FLUX LoRA

Daemon + Tea + Hyper

Detail Daemon + Tea Cache + Hyper-FLUX

NAG (Normalized Attention Guidance) Configurations

NAG Only

Normalized Attention Guidance

NAG + Hyper-FLUX

NAG + Hyper-FLUX LoRA

NAG + Detail Daemon

NAG + Tea Cache combination

NAG + Tea + Hyper

NAG + Tea Cache + Hyper-FLUX

NAG + Detail Daemon Combinations

NAG + Detail Daemon

NAG + Detail Daemon combination

NAG + Daemon + Tea

NAG + Detail Daemon + Tea Cache

NAG + Daemon + Hyper

NAG + Detail Daemon + Hyper-FLUX

NAG + Daemon + Tea + Hyper

All optimizations combined

Tried also DEIS Sampling for Daemon + NAG

NAG + Daemon + DEIS

NAG + Detail Daemon + DEIS Sampling

NAG + Daemon + Hyper + DEIS

NAG + Detail Daemon + Hyper-FLUX + DEIS

NAG + Daemon + Tea + DEIS

NAG + Detail Daemon + Tea Cache + DEIS

All Features + DEIS

Complete pipeline: NAG + Detail Daemon + Tea Cache + Hyper-FLUX + DEIS

Comments

This is not the easiest task, and it seems like NAG is making a big difference even with a generic negative prompt. The Flux Pro result is quite good, and it shows a very high level of detail especially visible on the cat skin. Unfortunately I have to give this to the longest run (202.34s) which is the NAG + Daemon configuration.

Generation Timings

• Next: Clay Rider
• Intro