This is a zero terminal SNR (ZSNR) model and requires special setup. Read the description.

Official continuation of Fluffyrock under Project RedRocket.

Improvements:

• Fixes to classifier-free guidance (CFG) with caption dropout, which should reduce brownout issues.

• A brand-new refiner model which greatly enhances small details, especially in combination with Furception, and costs nearly nothing to run. Please be sure to look at the versions at the top to download it.

• A new timestep weighting which helps greatly with construction of scene composition.

• A larger effective batch size of 64, along with EMA weighting.

• Furception VAE packed in (licensed separately, see its page)

Please do refer to the original model's page for more instructions.

Setup instructions/basic usage:

• Use the included config (A1111) or the ModelSamplingDiscrete node (ComfyUI/Nodegraph Hellworld) to use v-prediction

• Use the Zero Terminal SNR noise schedule in settings (A1111, on dev branch, will be part of stable 1.8.0) or the zsnr setting on the ModelSamplingDiscrete node (ComfyUI) to set the zero terminal SNR noise schedule.

• Use CFG Rescale at 0.7, this is a node on ComfyUI and an extension exists for A1111. Unlike previous releases of Fluffyrock, this should not cause brownout as it did before.

Miscellaneous Recommendations:

• Consider using a CFG Scale of 10

• For FreeU, I recommend B1:1.5, S1:0.8, B2:1.6, S2:0.8. For the A1111 version, Transition Smoothness 1 also seems to help.

• Karras samplers behave weird and waste a lot of time at the start sampling the same step over and over, and will not converge very fast. You may want to stick to the normal samplers or try a lower sigma_max like 160 (which will improve convergence speed at the cost of some bias towards average brightness, but nowhere near as bad as a non-offset noise, non-ZSNR model, and which can usually be countered by increasing prompt weight on dark, to something like 1.5)

Notes on the refiner model:

In the two most common frontends (A1111 and ComfyUI), the refiner works much like it does with SDXL. For A1111, assuming that you are not using a Karras sampler/schedule, you can set the refiner to switch over at 0.8 for txt2img and not worry too much. However, there are currently some bugs with pending fixes that make it more complicated when using a Karras sampler or when using img2img/inpainting. Wait for pull requests #14978 and #14979 and then you can set it to 0.8 and forget it. For ComfyUI, use the refiner like you would use the SDXL refiner. SD.Next currently does not support refiners on 1.5/2.1 models, as Diffusers does not have the proper hooks for it -- there is a pull request pending for that in Diffusers, and once that is merged and SD Next updates to that Diffusers version, it should work.