Skip to content
BranchesTags

Latest commit

 

History

1 Commit

Folders and files

NameName
Last commit message
Last commit date
lib_res4lyf
lib_res4lyf
 
 
scripts
scripts
 
 
LICENSE
LICENSE
 
 
README.md
README.md
 
 

Repository files navigation

sd-forge-res4lyf

A port of the RES4LYF sampler family for sd-webui-forge-classic (Neo branch).

The upstream project is a ComfyUI extension. This package vendors just the parts that make sense outside of a ComfyUI node-graph (the underlying RES / DEIS samplers plus a few of the RES4LYF schedulers) and exposes them as first-class entries in forge-neo's sampler and scheduler dropdowns.

Installation

  1. Place the entire sd-forge-res4lyf folder into your forge-neo extensions/ directory:

    <your-forge-neo-install>/extensions/sd-forge-res4lyf/
<your-forge-neo-install>/extensions/sd-forge-res4lyf/scripts/forge_res4lyf.py
<your-forge-neo-install>/extensions/sd-forge-res4lyf/lib_res4lyf/...

  2. Restart the WebUI. The new samplers and schedulers appear in the standard dropdowns - no extra UI accordion is added.

What's included

Samplers

The following RES4LYF samplers are registered under the Sampling method dropdown (suffix (RES4LYF)):

RES4LYF name Forge label
res_2m RES 2M (RES4LYF)
res_3m RES 3M (RES4LYF)
res_2s RES 2S (RES4LYF)
res_3s RES 3S (RES4LYF)
res_5s RES 5S (RES4LYF)
res_6s RES 6S (RES4LYF)
res_2m_ode RES 2M ODE (RES4LYF)
res_3m_ode RES 3M ODE (RES4LYF)
res_2s_ode RES 2S ODE (RES4LYF)
res_3s_ode RES 3S ODE (RES4LYF)
res_5s_ode RES 5S ODE (RES4LYF)
res_6s_ode RES 6S ODE (RES4LYF)
deis_2m DEIS 2M (RES4LYF)
deis_3m DEIS 3M (RES4LYF)
deis_2m_ode DEIS 2M ODE (RES4LYF)
deis_3m_ode DEIS 3M ODE (RES4LYF)

All samplers use forge-neo's CFGDenoiser, so CFG, ADM/IP-Adapter, ControlNet, LoRA stack, etc. all work just like with the built-in samplers.

Schedulers

  • bong_tangent - forge-neo already ships this scheduler upstream (originally taken from RES4LYF), so this extension does not add it again. Pick Bong Tangent from the schedule dropdown.
  • tan / Tan (RES4LYF) - the single-stage tangent scheduler from RES4LYF.
  • beta57 / Beta57 (RES4LYF) - the Beta-distribution-PPF sigma schedule with alpha=0.5, beta=0.7 hard-pinned at the RES4LYF defaults. Same math family as forge-neo's built-in Beta scheduler, but doesn't read shared.opts.beta_dist_alpha/beta_dist_beta so the curve is stable across UI option changes.

What's NOT included (and why)

Most of the RES4LYF project is built around ComfyUI custom-node classes that have no equivalent in forge-neo's Gradio-based UI:

  • ClownSampler / SharkSampler nodes
  • ClownsharKSampler chain, ClownGuides, ClownGuide groups
  • Regional / temporal conditioning nodes (AttnMask, RegContext)
  • Style transfer / WCT / Retrojector / StyleMMDiT nodes
  • Flux variant patches (ReFlux, ReWan, etc.)
  • Latent manipulation nodes (LagrangeInterp, Sortpicker, ...)
  • All Adv UI dialogs and ComfyUI-side js/ widgets

Porting these would mean rebuilding forge-neo's UI around RES4LYF's node graph - well beyond the scope of "sampler + scheduler" parity. The mathematical core of the project (the Runge-Kutta sampler and its RES / DEIS coefficient tables) is what's exposed here.

How the port works

The vendored RES4LYF source under lib_res4lyf/ is unchanged from upstream. A compatibility shim (lib_res4lyf/_compat.py) installs a fake comfy package into sys.modules so that the upstream import comfy.X statements resolve against forge-neo equivalents:

Upstream import Resolved to
comfy.utils.bislerp vendored from ComfyUI
comfy.model_sampling.EPS / CONST / ... duck-typed classes that match on prediction_type
comfy.model_patcher.set_model_options_post_cfg_function backend.patcher.base.set_model_options_post_cfg_function
comfy.k_diffusion.sampling modules_forge.packages.k_diffusion.sampling
comfy.ldm.common_dit.pad_to_patch_size backend.utils.pad_to_patch_size
comfy.samplers.SCHEDULER_NAMES static list including forge-neo's scheduler names

The denoiser passed in by forge-neo is wrapped in a _DenoiserAdapter that re-exposes the ComfyUI attribute layout (model.inner_model .inner_model.model_sampling, .device, .model_config) on top of forge-neo's CFGDenoiserKDiffusion / ForgeScheduleLinker / shared.sd_model chain.

Compatibility matrix

The samplers should work with any model architecture that forge-neo itself supports - SD 1.x, SDXL, SD 3.5, Flux, Wan, etc. - because they ride on top of CFGDenoiser.

Tested against forge-neo features

These all hook in above the sampler layer, so they should compose cleanly with the samplers added by this extension:

Forge-neo feature Status / note
--sage (SageAttention) Compatible. SageAttention patches the attention layers inside the UNet/DiT, while the sampler only sees CFGDenoiser.
--fast-fp16 / --fast-fp8 Compatible. forge-neo's KModel.apply_model casts inputs to the model's computation dtype, so the fp64 latents that RES4LYF uses internally are downcast safely.
int8 diffusion Compatible for the same reason - quantised linear layers run after apply_model casts inputs.
torch.compile (max-autotune-no-cudagraphs) Compatible. forge-neo compiles KModel.apply_model / the UNet; this sits below CFGDenoiser, so the compiled graph is still invoked through the sampler.
Spectrum Integrated Should compose, but Spectrum's prediction caching was tuned against the stock samplers - the multistep RES samplers reuse previous-step model output, which Spectrum already caches, so quality at low step counts may differ vs. DPM++ 2M. Disable Spectrum if you see drift.
LoRA / ControlNet / IP-Adapter Compatible. They patch the UNet and are invoked via apply_model.
ADM / SDXL high-res fix Compatible.

dtype handling

RES4LYF promotes its sampler math to torch.float64. The wrapper casts the returned latent back to whatever dtype forge-neo handed in, so downstream code (VAE decode, hires fix, refiner, etc.) sees a tensor in the same dtype it would get from any other sampler.

Known limitations

  • eta / eta_substep are exposed as fixed values (0.5 for the ancestral SDE variants, 0.0 for the ODE variants). RES4LYF's per-step eta-curve and per-substep scheduling are not surfaced through forge-neo's UI.
  • RES4LYF's extra_options dict, regional guides, attention masks and style transfer settings are also not surfaced (see "What's NOT included" above).

Credits

  • The samplers, schedulers and all of the math here are the work of ClownsharkBatwing and the RES4LYF contributors (Apache-2.0).
  • The compatibility shim and registration glue is a thin wrapper that doesn't reimplement any of the actual sampling math.

Licence

lib_res4lyf/ is Apache-2.0 (inherited from RES4LYF). See LICENSE. The glue code under scripts/ and the _compat.py shim are released under the same Apache-2.0 terms.

About

RES4LYF's sampling method and scheduler port for forge-neo

Resources

Readme
License
Activity

Stars

3 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors

Languages