Case 04 · Local generation
Z-Image-Turbo-Next
A local creative workspace rebuilt for reliable daily use, recoverable history, and measurable release gates.
Delivery Loop trail
Four gates connect the problem to public proof.
Each gate records the evidence that justified moving forward.
Understand
- Daily creation and history gaps
Design
- Legacy freeze and release gates
Deliver
- FastAPI and React workspace
Verify
- Tests and fixed-environment timings
01 · Operational problem
A local generator is not a product if history and model state are unreliable.
Daily creative work needed recoverable generations, predictable start and stop behavior, and an interface that did not silently lose prior records.
02 · Product hypothesis
A bounded local workspace could separate browsing from model loading.
Opening the browser would remain lightweight; generation would explicitly own model lifecycle, queue state, and persistent history.
03 · Technical path
FastAPI, React, and ComfyUI were isolated behind a stable product surface.
The rebuilt application preserves an immutable legacy installation while a separate runtime adds generation APIs, pagination, and explicit release behavior.
04 · AI delivery
The upgrade moved through named gates instead of an open-ended rewrite.
Each release slice repaired one operating concern—history, pagination, model lifecycle, or shutdown—before the next was authorized.
05 · Real constraints
Hardware-specific performance and visual consistency remain real limits.
Performance varies by hardware and workload; character, scene, composition, and style can drift; Turbo editing is not guaranteed pixel-precise.
06 · Verification
Release claims are tied to tests and a fixed environment.
The release passed 55 Python tests, 33 subtests, 18 frontend tests, 15 non-live browser tests, and all 15 Gate H checks.
07 · Reusable capability
Explicit model lifecycle became an operational pattern.
Browsing, generation, history recovery, pagination, and shutdown now have distinct responsibilities and testable boundaries.
08 · Limits and public boundary
The public case excludes the runtime that makes local generation possible.
Private source, model weights, portable runtimes, ComfyUI checkout, databases, unselected inputs and outputs, local paths, and device state remain excluded.
Related work
See the method move across domains.
The technology changes; the discipline around constraints and evidence remains.
Reusable lesson
Local AI becomes dependable when its operating state is explicit.
The durable value is not one generated image; it is a workspace that can be started, inspected, recovered, and stopped.
A similar operational constraint?