GOO-370 · Verification

Sunburst generates, edits and cuts out — proven in the running app.

GPT Image 2.5 Sunburst is wired into the FAL registry, the seed, Galleria and the workflow nodes. Four real generations were driven through the live UI by a browser agent — a text-to-image, a reference edit, a transparent cut-out, and a run from the workflow canvas node, which is a separate code path from Galleria. All seven hops of the generation path are exercised on both surfaces.

Branch ziann/goo-370-feature-add-sunburst-model Files 11 changed, uncommitted Identifier gpt-image-2.5-sunburst

Where it stands

Passed

End to end

Three generations reached completed, none stuck in queued. 15 credits charged each, three times over.

Passed

Both surfaces

Galleria and the workflow canvas node, which are separate code paths. Six quality tiers and the new Background control render on each.

Open

Before shipping

The credit figure is a placeholder and there is no production rollout script. Everything else is settled.

The generation path, segment by segment

A request crosses seven hops. The provider call was proven first with direct requests; the remaining six were proven afterwards by driving the real interface, then checking the database and service logs rather than trusting the screen.

1 Galleria panel & workflow canvas node both surfaces: model picked, settings rendered, run Verified

2 Gateway POST /api/v1/generations Verified

3 Queue worker generation:image:fal Verified

4 Vertex-AI service registry lookup resolved the FAL path Verified

5 FAL provider call fal.run/openai/gpt-image-2.5/sunburst/… Verified

6 Download & store asset persisted, served back to the feed Verified

7 Completion & credit charge credit_ledger · usage · 15.0000 Verified

Worth noting for anyone debugging later: nothing ever sat in queued. FAL image models take the synchronous fal.run path, which sidesteps the completion-webhook failure mode that strands video generations when the gateway URL is wrong.

The interface

The Galleria panel with GPT Image 2.5 Sunburst selected, showing six quality buttons and a Background control. — Signed in · model selected Every control this change introduces, rendering live: **Quality with six tiers** — Auto, Low, Medium, High, X-High, Max, against the four `gpt-image-2` offers — and the **Background** row, which no other model in the catalogue has. The button quotes **Generate (15 credits)**, read from the seeded row.

Three generated images in the feed: a lemon with no background, a green lime, and a lemon on white. — Three runs, one feed Right to left, in the order they were made: the original lemon on white (`background: auto`), the same lemon turned into a lime through the **edit** endpoint, and a transparent cut-out. The cut-out is the convincing one — **it sits straight on the dark canvas with no white card behind it**. Credits read 9,999,855, down exactly 45 from 9,999,900.

The workflow canvas node

The canvas is a different code path from Galleria — its own config, its own defaults, its own execution file. Galleria passing proved nothing about it, so it was tested separately. The node carries a risk Galleria does not: its dispatcher destructures node data into an explicit allow-list and then rebuilds it into an explicit object literal, so a key missing from either list arrives undefined and falls silently back to a default.

A Generate Image node on the workflow canvas showing a transparent-background strawberry and a Background setting reading Transparent. — Workflow canvas · Generate Image node Background set to **Transparent**, and the strawberry sits on a checkerboard rather than a white card. Note the settings differ from Galleria's by design: **Size 1:1 HD** here against Galleria's 4:3, because the node has its own defaults.

POST /api/generations → 201
"config":{"image_size":"square_hd","quality":"low","background":"transparent","output_format":"png","num_images":1}

That wire body is the point of the whole test. background arrives as "transparent" — not "auto", not absent. The silent-fallthrough failure mode did not occur; both allow-lists carry the key.
The defaults prove the paths really are distinct. The node produced 1024×1024 from square_hd, where Galleria produced 1024×768 from landscape_4_3. Same model, different surface, different result — which is exactly why testing one could not stand in for the other.
Sunburst-only controls, confirmed by comparison. A GPT Image 2 node built in the same session rendered three settings; the Sunburst node rendered four. Background is genuinely new, not inherited.

Check	Observed	Result
Generation	3e14f5b5… completed, bg=transparent, 17.3s	pass
Credits	15.0000, quoted = charged	pass
Raw asset	1024×1024 RGBA, 51.0% clear pixels	pass
Picker order	Sunburst first, GPT Image 2 second	pass

What the backend actually did

The screen can lie. These lines are from vertex-ai.log and show the path the registry resolved for each of the three runs.

[FAL Image] model=gpt-image-2.5-sunburst path=openai/gpt-image-2.5/sunburst/text-to-image edit=false images_in=0
[FAL Image] model=gpt-image-2.5-sunburst path=openai/gpt-image-2.5/sunburst/edit          edit=true  images_in=1
[FAL Image] model=gpt-image-2.5-sunburst path=openai/gpt-image-2.5/sunburst/text-to-image edit=false images_in=0

The middle line is the one that matters. Nothing in the interface names an endpoint. Attaching one reference image flipped edit=true and swapped the path on its own, which is exactly the implicit switch the registry is supposed to perform.

Check	Observed	Result
Generations	3 rows, all status = completed	pass
Stuck in queued	0	pass
Credits charged	3 × usage 15.0000 in credit_ledger	pass
Picker order	Sunburst seq 5, GPT Image 2 seq 6	pass
Gateway listener	cwd = this worktree, not a stale binary	pass

Transparency, measured rather than eyeballed

background is the parameter new in 2.5. A screenshot of a dark canvas cannot distinguish a real alpha channel from a dark fill, so the raw stored assets were decoded and their pixels counted.

Run	Encoding	Fully transparent pixels	Corner alpha
background = auto	8-bit RGB	0 (0.0%)	255
background = transparent	8-bit RGBA	364,388 (46.3%)	0

A red apple on a plain white background. — Direct provider call · generate The first evidence gathered, before the cluster was available — a direct request to the provider returning **RGB**.

The same apple, recoloured green, on a transparent background. — Direct provider call · edit The same apple fed back through the edit endpoint with `background: transparent`, returning **RGBA**. Stem, highlight and contour survive the recolour untouched — the precision editing Sunburst is chosen for.

Option sets, checked against the live provider

Rather than pay for a generation per tier, deliberately invalid values were sent and the provider's 422 responses read back. Each names its accepted set, so every dropdown is validated at once, for free.

Parameter	Accepted by FAL	Offered in UI	Match
quality	auto, low, medium, high, xhigh, max	all 6	exact
background	auto, transparent, opaque	all 3	exact
image_size	square_hd, square, portrait_4_3, portrait_16_9, landscape_4_3, landscape_16_9, auto	all 7	exact

Static checks

Check	Scope	Result
Go build	gateway, vertex-ai, go-queue	pass
Go vet	vertex-ai routes & internal	pass
Go test	./routes/external/	pass
ESLint	all 8 frontend files	pass
SQL parse	seed.sql into throwaway Postgres	pass
Sequence audit	no duplicates in image or video	pass

Open items

Before production

The credit figure is a placeholder, not an agreed price.

Seeded at 15, derived from the repo's formula against the worst verified cost — $0.211 at 1024px max. The vendor publishes no price for 4K combined with max, so that case may exceed a flat rate. Both the model row and the cost row carry a marker to substitute the real figure. If the true cost varies sharply by resolution, a model_provider_costs row would track it better than a flat default.

No production rollout script. seed.sql is development only. Shipping needs a scripts/prod-*.sql that carries the sequence renumber, not merely the new row.
Model identity: confirmed. Linear was never authorised for this session, so which model "Sunburst" meant was inferred from public sources plus the existing gpt-image-2 integration on the same provider. The ticket has since been confirmed to contain exactly one thing — a link to developers.openai.com/api/docs/models/gpt-image-2.5-sunburst. The inference was right, and this is no longer an open question.
Transparent with JPEG does nothing. The interface permits the pairing; JPEG has no alpha channel, so the result is silently opaque. Cosmetic, unguarded.
Harmless log line. No image reference limits defined for model gpt-image-2.5-sunburst — image_reference_limits is {}, matching the other FAL models. No functional effect.
Still untested. The edit path from the canvas (reference upload on a node) — it was proven from Galleria, not from the node. Also node chaining, and the failure path: every generation this database has ever held is completed, so nothing exercises the error branch.
Pre-existing bug found, not fixed. With a nameless seeded user, POST /api/workflows/{id}/session returns 400 and the canvas silently refuses to add any node, for any model — no toast, no error. frontend/stores/userStore.js:188 falls back user.Email?.String || user.Email || "Unknown User"; when Email is a Go sql.NullString, the middle term is a truthy object, so an object is sent where a string belongs and "Unknown User" is unreachable. Unrelated to this change, latent in production where Clerk populates email. Worked around with fixture data only.