Phantom Distillation & LoRA Tests

Evaluating step-distilled and CFG-distilled LoRA performance

Test Date: August 20, 2025

Test Overview

Following the successful identification of ConditioningCombine as the optimal conditioning strategy (from previous tests), this battery evaluates distillation LoRAs designed to accelerate generation while maintaining quality.

Test Setup: All tests use the same seed (53), reference images, and prompt as the conditioning tests for direct comparison. The baseline uses ConditioningCombine with CFG=3.5 at 30 steps (our gold standard from previous testing).

LoRAs Tested:

  • LightX2V: Step and CFG distilled, designed for 4 steps at CFG=1.0
  • LightX2V Adaptive: Same as above but with adaptive rank scaling
  • CausVid V2: Modified autoregressive distillation for ~10-12 steps at CFG~2.0

🔍 Key Findings

Test Results Summary

Test Configuration Steps CFG Subject (1-10) Quality (1-10) Speed Overall
Test 01 Gold Standard (No LoRA) 30 3.5 9 8 2100s 8.5
Test 02 Baseline CFG=1.0 30 1.0 5 6 1050s 5.5
Test 03 Min CFG=2.0 30 2.0 4 7 2100s 5.5
Test 04 LightX2V Standard 4 1.0 5 4 140s 4.5
Test 05 LightX2V Adaptive 4 1.0 5 5 140s 5.0
Test 06 CausVid 12 steps 12 2.0 8 9 840s 8.5
Test 07 CausVid 8 steps 8 2.0 7 8 560s 7.5

Phase 1: Baseline Tests

Test #01: Gold Standard
Baseline
Configuration
No LoRA
Steps
30
CFG
3.5
Speed
2100s (35 min)
Subject
9
Quality
8
Notes: Very good overall result. Eyes are a bit high contrast but might be that CFG should be a bit lower, otherwise video is good.
Test #02: Pure CFG=1.0
Baseline
Configuration
No LoRA
Steps
30
CFG
1.0
Speed
1050s (17.5 min)
Subject
5
Quality
6
Notes: Video is better quality than expected with CFG of 1, but it is a little muted, and she walks a bit more like a Zombie. Resemblance is significantly lower especially on the face.
Test #03: Minimal CFG=2.0
Baseline
Configuration
No LoRA
Steps
30
CFG
2.0
Speed
360s
Subject
4
Quality
7
Notes: Video quality is much better but the resemblance of the character and dress is a bit strange, like it grabbed unexpected parts of each and didn't focus on important aspects like the face.

Phase 2: Pure Distillation Tests

Test #04: LightX2V Standard
LoRA
Configuration
LightX2V
Steps
4
CFG
1.0
Speed
140s
Subject
5
Quality
4
Notes: Video quality suffers with this one, more noise than expected which may be attributed to the scheduler or shift not being right for this distill lora. Resemblance is down too.
Test #05: LightX2V Adaptive
LoRA
Configuration
LightX2V Adaptive
Steps
4
CFG
1.0
Speed
140s
Subject
5
Quality
5
Notes: Video quality is slightly better with this version compared to the base one.
Test #06: CausVid 12 Steps
LoRA
Configuration
CausVid V2
Steps
12
CFG
2.0
Speed
840s
Subject
8
Quality
9
Notes: Very good result, almost as good resemblance as 30 steps version, and arguably better quality.
Test #07: CausVid 8 Steps
LoRA
Configuration
CausVid V2
Steps
8
CFG
2.0
Speed
560s
Subject
7
Quality
8
Notes: Some small loss of quality and resemblance compared to the 12 step version with CausVid.

Next Steps

Phase 3 Testing (In Progress): Hybrid approaches combining distillation with strategic CFG application. Initial tests include:

  • LightX2V for initial steps, then switching to CFG>1.0 for final steps
  • Reduced LoRA strength with higher CFG values
  • Staged distillation combining multiple LoRAs

The goal is finding the optimal balance between generation speed and quality, targeting a 2-3x speedup while maintaining 80%+ of baseline quality.