The Flux 1.0 DEV model was tested for a few days. The number of sampling steps seems to affect the results.
The base setting is:
Webui Forge
Model: flux1-dev-bnb-nf4.safetensors
Sampler: Euler simple
CFG Scale: 1
Distilled CFG Scale: 3.5
Seed: 3487602601
Some comparisons:
a blacksmith in workshop
The model will define and optimize a focused area. In this example, the face area remains almost the same from steps 25 to 100.
Once the model finished the face, the attention moves on to the clothing and hands. These should be in place after 30 steps. Unfortunately, the hands are getting worse...
The smoke, fire and oven change between 30 to 50 steps. A feature in the lower right side in the foreground is removed in the 100th step.
The background is fixed after 25 steps.
a blacksmith in underground workshop
The workshop in the middle ground stays the same from start to finish.
The blacksmith never stays the same. In 100 steps, it transforms into a completely different bigger person.
A deer in a forest
The deer is constantly shifting and growing larger until 50 steps.
The difference between 50 and 100 steps are some background and foreground minor adjustments.
A gorgeous fairy with exquisite transparent wings is fast flying across a dark magical forest at midnight. The starry sky is moonless.
This is a more complicated prompt.
The fairy at 20, 25, and 30 steps is different from the fairy at 50 and 100 steps. The fairy is also bigger at 50 and 100 steps.
It totally ignored the moonless prompt from the beginning to the end. (Flux cannot understand negative description?)
The moon and foreground become increasingly detailed as the number of steps increases.
another comparison between 20 and 50 steps. The 50 steps output (the top one) looks closer to the prompt. https://civitai.com/posts/5427367
Summary:
The sampling steps change the focused content in the picture. Sometimes, after certain steps the focused item may be fixed, and the attention is directed to other objects. Sometimes the focused item grows larger in higher steps. It is difficult to say which steps are ideal. The items look all good but different. The breaking point might be 50 steps?
The background and foreground changes slightly with the steps. They will become slightly more detailed.
It is critical to specify which component shall be focused on in the model.