Flux sliders working for schnell , but not for dev

[NitishTRI3D]

after 50-60 steps , the image composition for val -1 and val +1 is changing a lot , like the person's face and pose is changing

could it be because of values like eta=2 is too high for dev but okay for schnell ?

[rohitgandikota]

Hi @NitishTRI3D

We have not tested it for dev model and this is an experimental code. It is possible that eta=2 is high for dev models.

could I also suggest playing with the parameter skip_slider_timestep_till during inference. I would try 1-10 if you are using total inference steps as 50:

image = pipe(
                    target_prompt,
                    height=height,
                    width=width,
                    guidance_scale=guidance_scale,
                    num_inference_steps=num_inference_steps,
                    max_sequence_length=max_sequence_length,
                    num_images_per_prompt=1,
                    generator=torch.Generator().manual_seed(seed),
                    from_timestep=0,
                    till_timestep=None,
                    output_type='pil',
                    network=networks[net], 
                    skip_slider_timestep_till=0, # this will skip adding the slider on the first step of generation ('1' will skip first 2 steps)
                )