Nodes/RES4LYF/ClownStyle_Boost
ComfyUI Node Runs on cloud

ClownStyle_Boost

Push RES4LYF style transfer harder

By ClownsharkBatwing·Created 2 years ago·Updated 18 days ago· 1,222
ClownStyle_Boost
  • guides
  • guides
noise_modeupdate
recon_lureWCT
datashockscattersort
datashock_weight1.00
datashock_start_step0
datashock_end_step1
tile_h128
tile_w128

ClownStyle Boost is the "turn it up" companion to RES4LYF's style transfer. Where the base style-guide node borrows a reference image's look, ClownStyle Boost layers on more aggressive matching modes to make that transfer bite harder - it's what you reach for when a normal style guide is too polite and the reference's character isn't really coming through.

Fair warning up front: this is deep in RES4LYF's experimental end. ClownsharkBatwing built the pack solo on a single 4090 and documents most of it in Reddit comment threads rather than a manual, and the style tooling is the least-charted part of it. The mode names here - bonanza, datashock, scattersort - are his, and the community's standing complaint about the whole pack (there's a +45 thread literally titled "Someone needs to explain bongmath") applies double to these. So treat this as a knob to experiment with once the basic style guide is working, not a first stop.

How it works

It takes an existing GUIDES bundle (from a style-guide node) and augments it with a stronger transfer stage. noise_mode sets the overall strategy, and datashock picks a feature-matching operation - including scattersort, the mode the author wrote to move style without dragging the reference's color and lighting across. The boosted GUIDES flows on to your sampler.

The inputs and outputs that matter

  • guides (GUIDES) in / guides (GUIDES) out - chain this after your style-guide node and before the sampler.
  • noise_mode (default update) - the boost strategy. Options run direct, update, smart, recon, bonanza, roughly in order of how hard they push. update is the sane starting point; bonanza is the "send it" setting.
  • datashock (default scattersort) - the matching operation applied to shock the latent toward the reference's statistics. scattersort is the safest of the strong options for preserving your own colors.
  • datashock_weight (default 1) - how strong that push is. This is your main intensity dial.
  • datashock_start_step / datashock_end_step - which steps the boost runs over. Early-only keeps it from fighting the final detail pass.
  • recon_lure (default WCT) - the reconstruction operation used by the recon mode; only relevant when noise_mode is recon.

How to install it

ComfyUI Manager: search RES4LYF, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/ClownsharkBatwing/RES4LYF
cd RES4LYF
pip install -r requirements.txt

then restart and hard-refresh (F5). No downloads.

Common issues & troubleshooting

The image cooks or turns to mush. You over-boosted. Drop datashock_weight, move noise_mode down toward update or direct, and shorten the step range so the boost isn't running into the final steps.

Your own colors got hijacked by the reference. Keep datashock on scattersort - the other modes are more willing to import the reference's palette wholesale.

No visible difference. This only makes sense downstream of an actual style guide. If there's no GUIDES coming in with a real reference, there's nothing to boost - wire up ClownGuide Style first, get that working, then add this on top.

CategoryRES4LYF/sampler_extensions

Inputs (9)

NameTypeDefaultDescription
noise_modeCOMBOupdate5 options: direct, update, smart, recon, bonanza
recon_lureCOMBOWCTOnly used if noise_mode = recon. Can increase the strength of the style.
datashockCOMBOscattersortWill drastically increase the strength at low denoise levels. Use with img2img workflows.
datashock_weightFLOAT1.00-100–100Set the strength of the guide by multiplying all other weights by this value.
datashock_start_stepINT00–10000Start step for data shock.
datashock_end_stepINT11–10000End step for data shock.
tile_hINT12816–10000Tile size for tiled modes. Lower values will transfer composition more effectively. Dimensions of image must be divisible by this value.
tile_wINT12816–10000Tile size for tiled modes. Lower values will transfer composition more effectively. Dimensions of image must be divisible by this value.
guidesoptGUIDES

Outputs (1)

NameTypeDescription
guidesGUIDES