01
Editorial
From Reference Images to a Reusable Creative Recipe
A four-step process for turning a folder of reference images into a reusable creative recipe: extract non-negotiables, build a structured skeleton, stress-test on new subjects, and freeze a hand-off card that another operator can run.
Workflow Kitchen · Oct 15, 2026
§ 01
This Week
Selected Reads
02
The Overfit LoRA: How I Found the Problem and Recovered
A chronological case study of diagnosing and recovering from an overfit LoRA on MiniMax H3 Turbo: late detection via out-of-distribution probes, dataset cleaning, earlier stop, and a cleaner retrain that restored usable range.03
Speed, Detail, and Consistency: Which Trade-Off Matters Most?
A controlled MiniMax H3 Turbo bench measuring the speed–detail–consistency trade-off, with clear pass rates, a decision rule that prioritizes consistency for repeatable recipes, and an example of how the rule played out on a real six-frame series.04
How to Decide Whether to Add More Training Data
A strict two-question decision gate for adding training data to a LoRA: name the specific underrepresented attribute, confirm clean examples exist, and re-test the original failure mode before keeping the new images.05
A Small-Team AI Workflow That Can Be Repeated on Monday
A practical small-team workflow for MiniMax H3 Turbo and LoRA work: three lightweight roles, a twenty-minute Monday sequence, and the communication rules that prevent silent strength drift and undocumented changes.§ 02
Index — All Entries
Archive
| No. | Title | Author | Category | Date |
|---|---|---|---|---|
| 01 | From Reference Images to a Reusable Creative Recipe | Workflow Kitchen | Oct 15, 2026 | |
| 02 | What Film Cameras Taught Me About Evaluating Generated Images | Field Notes | Oct 15, 2026 | |
| 03 | The Overfit LoRA: How I Found the Problem and Recovered | Failure Files | Oct 14, 2026 | |
| 04 | Speed, Detail, and Consistency: Which Trade-Off Matters Most? | The H3 Bench | Oct 14, 2026 | |
| 05 | How to Decide Whether to Add More Training Data | LoRA Workshop | Oct 13, 2026 | |
| 06 | A Small-Team AI Workflow That Can Be Repeated on Monday | Workflow Kitchen | Oct 13, 2026 | |
| 07 | When Better Hardware Did Not Fix the Real Problem | Failure Files | Oct 12, 2026 | |
| 08 | The Keep-or-Kill Test: Is This AI Workflow Worth Keeping? | Field Notes | Oct 12, 2026 |