docs/validation.md

Live validation — 30 September 2026

Both public endpoints were called without a Hugging Face token.

Test	Result	Observed end-to-end call time
Text-to-image, 512 × 512	Valid image; ivory vase on a stone plinth	5.57 seconds
Instruction-based edit, 512 × 512	Valid image; vase changed to blue	5.31 seconds

Seed: 137. Steps: 4. Guidance: 1.0. One request per workflow; these times are illustrative single observations, not a latency benchmark or SLA.

Public example

Generation prompt: Product photograph of a single matte ivory ceramic vase on a warm beige stone plinth, soft studio light from the left, plain light gray background, no text, no people.

Edit instruction: Change only the vase color to cobalt blue. Keep the same vase shape, position, stone plinth, shadows, lighting and gray background.

Both outputs were visually inspected. Shape and composition remain broadly consistent, while the edited vase also gains a glossier appearance. This shows why exact attribute preservation is not guaranteed.

The inputs and outputs are synthetic; no client assets were used. Multi-reference quality and complex photographic edits were not exhaustively tested.