Reverts changes to restore old DDIM

[ec111]

Restores old DIMM. Old implementations have o prefix.

[ec111]

Code for work around for new samplers to be compatible with extensions removed. They will work with original implementation.

[AUTOMATIC1111]

Doesn't it look from discussion like it's just a problem with v- prediction? I'd rather we fixed the actual problem rather than revert.

[ec111]

Doesn't it look from discussion like it's just a problem with v- prediction? I'd rather we fixed the actual problem rather than revert.

I'm not well-versed in the AI space but in reviewing the discussion, there is no conclusive answer to what the problem is, only speculation.

How can we confirm the problem lies in v-prediction?

[AUTOMATIC1111]

Can you reproduce the unwanted effect on stable diffusion's 2.0/2.1 checkpoint?

[ec111]

Can you reproduce the unwanted effect on stable diffusion's 2.0/2.1 checkpoint?

Yes. Using the standard v2-1_768-ema-pruned, the same effects can be observed.

I would make the observation that at any CFG, the new DDIM implementation produces identical or nearly identical results as Euler and Heun and other standard samplers, whilst the old DDIM implementation produces noticeably different results.

At 150 steps, CFG 3 scale.

In any case, fixing the problem is probably beyond my expertise...

[BurnZeZ]

Has there been any recent progress on this?

[AUTOMATIC1111]

Added a option to settings and wrote a response to https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/13066.

[AUTOMATIC1111]

Since there is no more discussion in https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/13066 or here, I'm closing the PR. If there's more, please reopen.