Separate denoised and noise estimation in Euler CFG++

[chaObserv]

This will change their behavior with the sampling CONST type. It also combines Euler CFG++ and Euler ancestral CFG++ into one main function.

The current CFG++

The current Euler CFG++ can work on the CONST sampling type (RF), which matches the algorithm from the appendix of the paper, but the guidance scale must be very low (~0.1), and it’s unstable.
It tends to oversaturate when given more steps under the same guidance scales that seem to work with fewer steps.

CFG++ for flow models

Recently, the official proposed a new approach for flow models, which separates denoised and noise estimation during sampling.

It changes the algorithm back to

eps_uncond = (x_s - alpha_s * uncond_denoised) / sigma_s
x_t = alpha_t * denoised_cfg + sigma_t * eps_uncond

instead of

x_t = denoised_cfg + sigma_t * v_uncond

With this change, the behavior of the guidance scale is more similar to the other sampling types.

This should make it more usable with flow.

Euler ancestral CFG++

The current euler_ancestral_RF is more similar to the "Overshoot Sampling" approach (from the AMO sampler), which takes a larger step along the velocity and rescales to the next state with noise compensation.
This method doesn't separate denoised and noise, so it's probably not suitable for CFG++.

alpha_ratio * (denoised_cfg + sigma_down * v_uncond) + renoise

Euler ancestral in k-diffusion seems to implement DDIM's eta parameter, which controls the interpolation between deterministic DDIM (eta = 0) and stochastic DDPM (eta = 1). A simple strategy is to preserve the original algorithm and apply RF's alpha and sigma, just like SANA's adaptation of dpm-solver to flow-dpm-solver. It can also match the "Stochastic Curved Euler Sampler", a DDPM variant for RF. (Algorithm 1 in 2506.15864).

The ancestral CFG++ seems fine in the current implementation.

All the examples are generated by SD3.5 medium with simple scheduler.