Describe the bug
The step() function calculates prev_timestep using a hard-coded formula:
prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps.
https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_ddim.py#L402
However, if the scheduler is initialized with timestep_spacing="linspace", the preve_timestep is not match the set_timesteps.
Suggested Method to Fix:
This inconsistency could be resolved by modifying the step() function to look up the correct prev_timestep from the self.timesteps list directly, just as in DDPMScheduler.
https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_ddpm.py#L549
Reproduction
scheduler = DDIMScheduler.from_pretrained(model_id, subfolder='scheduler', timestep_spacing="linspace")
scheduler.set_timesteps(10)
prev_timesteps = []
for t in scheduler.timesteps:
prev_t = t- 1000 // num_inference_steps
prev_timesteps.append(prev_t.item())
Logs
System Info
Diffusers version: 0.35.2
Who can help?
No response