Share:

Parameter Sweep — Test Plan

Finding the optimal guidance × steps × LoRA scale × film stock combination • March 2026

Total Photos
992
Est. Cost
~$50
Est. Time
10-12h
Stages
4
Combos Tested
20
Subjects
8

Why This Test Exists

We've been using guidance 3.5, steps 35, LoRA scale 0.9 — but these were settled on with small sample sizes. Guidance and steps interact — guidance 3.0 at 50 steps produces a completely different image than guidance 4.0 at 30 steps. The only way to find the true optimum is a controlled sweep across all three variables, with enough samples to trust the result.

The Question: Is there a single best combo, or does it depend on the LoRA quality? Do mid/weak LoRAs need different settings than strong ones?

What's Held Constant

Variables Under Test

VariableRangeWhy It Matters
Guidance2.0, 2.5, 3.0, 3.5, 4.0, 4.5How strictly the model follows the prompt. Too low = random. Too high = plastic/saturated.
Steps25, 30, 35, 40, 45, 50Refinement passes. Too few = soft. Too many = no gain but 2× cost.
LoRA Scale0.80, 0.85, 0.90, 0.95How much the LoRA influences the output. Too low = generic face. Too high = over-fits.
Film StockPortra 400, Portra 800, Fuji 400H, EktarColour grading layer. Affects skin tones and overall mood.

The 4 Stages

1
Coarse Grid — Find the Region
READY
320 photos
~$15-18 cost
~3-4h runtime
20 combos × 2 subjects × 2 styles × 4 candidates

Tests all 20 parameter combinations on Scott and Sarah with Studio Dramatic and Smart Casual. Enough to see clear winners and eliminate the bottom half.

The 20 Combos

g2.0 s35 l0.90
g2.5 s35 l0.90
g3.0 s35 l0.90
g3.5 s35 l0.90
g4.0 s35 l0.90
g4.5 s35 l0.90
g3.0 s25 l0.90
g3.0 s30 l0.90
g3.0 s40 l0.90
g3.0 s45 l0.90
g3.0 s50 l0.90
g3.5 s40 l0.90
g3.5 s45 l0.90
g3.0 s35 l0.80
g3.0 s35 l0.85
g3.0 s35 l0.95
g3.5 s35 l0.85
g3.5 s35 l0.95
g2.5 s45 l0.95
g4.0 s30 l0.85

Output: Ranked table of all 20 combos. Top 6 automatically saved for Stage 2.

2
Full Validation — Confirm Across All Subjects & Styles
AFTER STAGE 1
480 photos
~$22-28 cost
~5-6h runtime
6 combos × 5 subjects × 4 styles × 4 candidates

Takes the top 6 combos from Stage 1 and tests them across all 5 strong LoRA subjects (Scott, Sarah, Mike, Emily, Chloe 2) and all 4 styles. This is the definitive test — if a combo wins here, it's the real deal.

Strong Subjects

SubjectGenderKnown AvgSession ID
ScottMale82%scott-tolbert-mm5jr5vd
SarahFemale89%neil-test-7-mm5jrqrx
MikeMale86%session-mmbzwxxutdpb
EmilyFemale85%emily-spence-utting-mmdwzi3o
Chloe 2Female84%chloe-spence-utting-mmdrm9l7

Output: Overall winner + per-subject breakdown + per-style breakdown. Winner saved for Stage 3.

3
Film Stock — Colour Grading Optimisation
AFTER STAGE 2
96 photos
~$5-6 cost
~1h runtime
4 film stocks × 3 subjects × 2 styles × 4 candidates

Tests 4 film stocks on the winning guidance/steps/LoRA scale combo. Film stock is likely independent of the other parameters — it's a colour grading layer — so this can be a smaller test.

Film Stocks

StockCharacter
Portra 400Natural skin tones, soft grain, warm — current default
Portra 800Slightly grainier, warmer, good in mixed light
Fuji 400HSoft pastels, slightly cool — popular for portraits
Ektar 100Fine grain, vivid saturated colours — punchy

Output: Ranked film stocks. Winner applied to final settings.

4
Mid LoRA Validation — Do Weak LoRAs Need Different Settings?
AFTER STAGE 2
96 photos
~$5-6 cost
~1h runtime
2 combos × 3 subjects × 4 styles × 4 candidates

Tests the winning combo AND the old default (g2.5/s35/l0.90) on 3 mid/weak LoRAs. If the winner also works best for weak LoRAs → one universal setting. If the old default wins for weak LoRAs → we need adaptive settings based on LoRA quality.

Mid LoRA Subjects

SubjectGenderKnown AvgWhy They're "Mid"
Neil 2Male39%Poor selfie quality — weakest LoRA
ChloeFemale80%First session — decent but not top tier
Emily 2Female80%Second session — consistent but not highest

Output: Winner vs old default on mid LoRAs. Determines if adaptive settings are needed.

Decision Rules

ScenarioAction
One combo leads by 3+ points consistently across subjects & stylesLock it — that's the winner
Two combos are within 2 pointsPick the one with fewer steps (cheaper) or lower guidance (more natural)
Winner works for mid LoRAs tooOne universal setting — simple
Different winner for mid LoRAsBuild a quality gate — score a test image first, pick settings based on result
Mid LoRAs are bad regardlessFocus on upload quality gates instead — reject bad selfies early

How to Run

# Set token first
export REPLICATE_API_TOKEN=$(powershell -Command "[System.Environment]::GetEnvironmentVariable('REPLICATE_API_TOKEN', 'User')")

# Run all stages overnight (~10-12 hours)
node test_parameter_sweep.js --stage all

# Or run individually
node test_parameter_sweep.js --stage 1
node test_parameter_sweep.js --stage 2
node test_parameter_sweep.js --stage 3
node test_parameter_sweep.js --stage 4

# View results
node test_parameter_sweep.js results
node test_parameter_sweep.js compare

Adding a New Subject

  1. Take 8 selfies: front neutral, front smile, left 45°, right 45°, chin up, chin down, half body, close-up face
  2. Upload at reports.iconicbyai.com/iconic/upload.html — enter name, gender, age, ethnicity
  3. Train LoRA: node generate_photos.js <session-id>
  4. Add session ID to test_batches.json and to test_parameter_sweep.js subject arrays

Best test: A brand new person the system has never seen gives the most unbiased validation. Take good selfies (well-lit, multiple angles, neutral expression) for a strong LoRA — this isn't about testing with bad photos, it's about confirming the winner works for new subjects.

Parameter Sweep Plan — Iconic by AI • March 2026 • 992 photos across 4 stages