patrickvonplaten commited on
Commit
72822ef
1 Parent(s): bb7be9f

Add `clip_sample=False` to scheduler to make model compatible with DDIM.

Browse files

Hey lambdalabs 👋,

Your model repository seems to contain a stable diffusion checkpoint. We have noticed that your scheduler config currently does not correctly work with the [DDIMScheduler](https://huggingface.co/docs/diffusers/main/en/api/schedulers#diffusers.DDIMScheduler) because `clip_sample` is not set to False and will therefore [incorrectly default to True](https://github.com/huggingface/diffusers/blob/3ce6380d3a2ec5c3e3f4f48889d380d657b151bc/src/diffusers/schedulers/scheduling_ddim.py#L127).
The official stable diffusion checkpoints have `clip_sample=False` so that the scheduler config works will **all** schedulers, see: https://huggingface.co/stabilityai/stable-diffusion-2-1-base/blob/main/scheduler/scheduler_config.json#L7.

We strongly recommend that you merge this PR to make sure your model works correctly with DDIM.

Diffusingly,
Patrick.

Files changed (1) hide show
  1. scheduler/scheduler_config.json +3 -2
scheduler/scheduler_config.json CHANGED
@@ -4,8 +4,9 @@
4
  "beta_end": 0.012,
5
  "beta_schedule": "scaled_linear",
6
  "beta_start": 0.00085,
 
7
  "num_train_timesteps": 1000,
8
  "skip_prk_steps": true,
9
- "trained_betas": null,
10
- "steps_offset": 1
11
  }
 
4
  "beta_end": 0.012,
5
  "beta_schedule": "scaled_linear",
6
  "beta_start": 0.00085,
7
+ "clip_sample": false,
8
  "num_train_timesteps": 1000,
9
  "skip_prk_steps": true,
10
+ "steps_offset": 1,
11
+ "trained_betas": null
12
  }