New samplers "ipndm" and "ipndm_v"; what's the difference?

[not-ski]

I saw that "ipndm" and "ipndm_v" samplers were added (d5efde8); what exactly is the difference between these two?

From scouring the linked source repo, I gather that ipndm is a fourth order sampler and that ipndm_v is the "variable-step version". But what does variable-step mean here?

Thanks <3

[zhyzhouu]

iPNDM is actually the classical Adams–Bashforth methods, which is derived for a fixed step size (uniformly sampled sigmas). But for diffusion sampling, the stride of each step is usually different, so we also provide a variable-step version of iPNDM denoted as iPNDM_v. However, we find iPNDM empirically performs better especially for small-resolution datasets. For detailed derivation, you can refer to Interval methods of Adams-Bashforth type with variable step sizes.

[not-ski]

Very interesting! Yeah I have generally noticed that in Stable Diffusion, DPM++ 3M requires many more steps than 2M to achieve usable/convergent results.

Anecdotally, at very high step counts (50+) I have preferred the results from higher-order solvers. But certainly, when it comes to speed they seem to be empirically inferior.

[zhyzhouu]

These are academic benchmarks after all so yeah, I am not surprised if the way a model was trained or fine tuned also affects it. The coefficient might make a difference but I have no idea how that can be tuned per model unfortunately.

Unfortunately, I am currently not sure how different fine-tuning will affect the model too. 🤯

[simonlui]

Anecdotally, at very high step counts (50+) I have preferred the results from higher-order solvers. But certainly, when it comes to speed they seem to be empirically inferior.

It doesn't save that much time per step but total steps needed, agreed. My anecdotal experience having played around with iPNDM and SD3 tonight and tweaking the max_order, the coherence of the prompt to not generate extra limbs and etc. with human subject seems to go down if you lower the max order so I supposed there is some correlation there as well.

[not-ski]

I'd be interested to see what results of your empirical testing on SDXL. IME, ipndm_v has noticeably outperformed ipndm (using GITS or AYS and targeting 20 steps).

Both are still generally giving lesser quality results than lower-order solvers at that step count, but ipndm_v gets much closer to a coherent/convergent result.

This is odd, considering it directly contradicts the data gathered in your previous testing. I wonder if this is another quirk of Stable Diffusion.

[kalle07]

today in webforge-ui 4 new samplers
ipndm and ipndm_v
HeunPP2
DEIS
no use for FLUX ... why : D