Draft fast, finish high-fidelity — a priced AI image workflow

The expensive mistake in AI images is not using a pricey model. It is using a pricey model to find out what you want. Most takes in a session are decisions (is the subject left or right, is the light warm or cold, is this even the right idea), and only the last one is the deliverable. Split the two and you pay finishing prices for one take instead of ten. With eroq's current lineup that split pays handsomely in some pairings and barely at all in others, so this tutorial runs the numbers before it runs the method.

Two benches

Drafting bench, fast and cheaper per take:

• Flash Image, 10 credits / image, a few seconds a take, broad style range.
• Nano Banana 2.1, 31 credits / image, typically about 8 seconds, reads plain sentences well.
• Grok Imagine 2.0, 32 credits / image, typically about 15 seconds, characterful and quick to change looks.

Finishing bench, where the detail is:

• Nano Banana Pro, 124 credits / image, typically about 25 seconds, the highest fidelity and the most accurate text of its line.
• GPT Image 2.5 Sunburst, 31 credits / image, typically about 15 seconds, precise composition and the sharpest in-image text.
• Seedream 5.0 Pro, 36 credits / image, typically about 20 seconds, dense detail and texture.

The gap that matters sits in front of Nano Banana Pro: one take costs 4 times a Nano Banana 2.1 take and 12.4 times a Flash Image take. In front of Sunburst or Seedream, a Nano Banana 2.1 draft saves 0 credits and 5 credits a take respectively. There, draft on Flash Image or skip the bench.

What the draft decides

A draft is for the choices that survive a change of engine:

• Composition. Where the subject sits, how much sky, how close the camera is.
• Format. Pick the ratio at the draft stage and keep it. A 16:9 idea and a 9:16 idea are different pictures.
• Light and palette. Direction and temperature of the light, the two or three dominant colours.
• Room for words. If the final needs a title, leave the band empty now. Do not judge spelling on a draft.
• Style family. Photograph, painting, flat illustration.

What a draft cannot decide is exactly what you pay the finisher for: material detail, lettering, the texture of skin, stone or glass. So write the draft prompt as if it were the final one, with subject, setting, light, lens and mood. A lazy draft prompt teaches you nothing about the finished image.

The hand-off is the prompt, not the picture

The new third-party models take no reference pictures, so the finisher never sees your draft. It reads the same words and composes the frame again. Plan for that:

1. Remix the keeper. Open it from your History or Library and press Remix: the studio reloads the prompt, the model and the format.
2. Switch the model, nothing else. Same words, same format, finishing engine.
3. Write down what the draft decided. If the finisher moves things, add the decision as words: "subject on the left third, low horizon, light from the right".
4. Same family travels best. Nano Banana 2.1 to Nano Banana Pro keeps one family's reading of your sentence. Grok to Sunburst is a cross-family move: treat the draft as a direction and expect to rewrite a clause.

When the draft's exact pixels have to survive, only a reference engine can help, because it reads the picture itself. Eroq One takes the draft as Picture 1 and edits it by instruction while keeping its frame.

Three pairings, priced

Pairing	Four drafts	One finish	Per idea	Four finishing takes instead
Nano Banana 2.1 → Nano Banana Pro	124	124	248	496
Flash Image → Seedream 5.0 Pro	40	36	76	144
Grok Imagine 2.0 → Sunburst	128	31	159	124

All figures in credits. The Google ladder saves 248 credits an idea, and the Flash-to-Seedream ladder 68 credits. The Grok ladder costs more than four Sunburst takes would. That is not a mistake: Grok earns its place when you do not know yet what the image is, because it moves between looks quickly, and the comparison in Grok Imagine 2.0 vs Nano Banana 2.1 is about exactly that cost per direction.

Two worked sessions

A week of social posts, ten ideas. Forty Flash Image drafts, then one Seedream finish per idea: 760 credits, about $7.60 at the entry pack. Rendering all forty takes on Seedream instead would be 1,440 credits.

One campaign hero. Three rounds of four Nano Banana 2.1 drafts while the client argues about the light, then two Nano Banana Pro finals: 620 credits. The same twelve takes of indecision on Pro would cost 1,488 credits.

Failed takes refund themselves on either bench, so the only real waste is a draft that answered no question. Change one clause between batches and you always know what moved.

When to skip the draft

• The finisher costs about the same as the drafter. Sunburst is the obvious case: draft on Sunburst itself.
• The frame is mostly text. Spelling is decided at the finish, so render on the text model from the start; Nano Banana Pro vs Nano Banana 2.1 notes where Pro's text accuracy shows.
• A face or product must match photos. The fast third-party models cannot hold one. Start on Eroq One, or Krea2 when four references are enough.
• You already know the composition. Then the draft is a formality you are paying for.

FAQ

Why not use the best model for everything?

Because most takes are decisions, not deliverables. On the Google ladder, four Nano Banana 2.1 drafts plus one Pro finish cost 248 credits, against 496 credits for four Pro takes.

Can I turn a draft into the final image directly?

Not on the third-party models: the finish is a fresh render of the same prompt on the better engine, not an enlargement of the draft. To keep the draft's exact frame, edit it on Eroq One, which reads it as a reference.

Which fast model should I draft on?

Flash Image for the cheapest takes, Nano Banana 2.1 when the finisher is Nano Banana Pro, and Grok Imagine 2.0 when you need many different looks quickly. Per-model prices are collected in what AI images cost in 2026.

Does drafting save money before a Sunburst finish?

Not on Nano Banana 2.1, which saves 0 credits a take there. A Flash Image draft saves 21 credits a take, which is worth it for wide exploration.

Draft four, keep one, and open the image studio.