Attribution: The specific end frames used in this test were provided by @the_shadow_nyc on the banodoco Discord server.

RES4LYF Sampler Extreme Distillation Tests

Testing 119 samplers under 4-step CFG=1.0 distilled conditions

Test Date: 21 September 2025

Test Overview

NOTE: This page was vibe-coded, there may be inaccuracies or other weirdness.

This comprehensive test explores how 119 different samplers from the RES4LYF sampler pack perform under extreme distillation conditions using the Wan2.2 VACE model with LightX2V LoRA.

Research Question: While these samplers were not designed for such aggressive distillation (4 steps, CFG=1.0), some may perform remarkably well, especially those with built-in sub-steps that could provide additional refinement within the limited step budget.

Unique Setup: The test uses continuous video generation with 9-frame overlap between segments, dynamic end-frame guidance, and a split between High Noise (2 steps) and Low Noise (2 steps) VACE models.

Workflow Visualization

The workflow is embedded in the image above for reference.

📄 ComfyUI Workflow: Download Test Workflow JSON (Right-click → Save as to download)

🔬 Test Parameters

Model

Wan2.2 VACE (High→Low split)

LoRA

LightX2V Distillation

Total Steps

4 (2 High + 2 Low)

CFG Scale

1.0

Scheduler

Linear-Quadratic [1.0, 0.9, 0.8, 0.55, 0.0]

SDE Eta

0.25

Dtype

float32

Frames

81 per segment (9 overlap)

119

Samplers Tested

Analysis Notes

Initial Observations:

Samplers with built-in sub-steps (marked with _Xs notation) may perform better under these extreme conditions
The SDE eta value of 0.25 adds stochasticity which may help or hinder depending on the sampler
Some implicit methods might struggle with only 4 steps despite their theoretical advantages
The High→Low model transition at step 2 may favor certain sampler architectures

Next Steps: Detailed analysis of results to identify which sampler families perform best under distillation, followed by eta=0.0 comparison tests for top performers.