One prompt, or five?

One-shot vs chaining, defined

One-shot generation asks a single prompt to do everything at once: product, pose, scene, style, lighting, all in one go. Chaining (also called a sequential pipeline) does the opposite, building the image in small checked steps where each output becomes the next step's input, and you validate before moving on. The rule for choosing between them is simpler than most people expect. Start with one shot from the sharpest reference you have, generate three variations, and cull. Only chain when a single edit genuinely cannot do the job: an angle that has to be created before the product can be placed, or a multi-product composition. Chaining is the exception, not the route, and its territory shrinks with every new model generation.

This is the one comparison you will not find written down anywhere else, because almost nobody measured it. We did. The rest of this page is the reasoning and the numbers behind that rule.

The Small Steps Law: never bridge a big delta in one step

Underneath both approaches sits a single law. Never ask the generator to bridge a big delta in one step. A delta (a jump) is the distance between what your inputs already show and what you are asking for. If your reference is a front-facing product and you want it on a model at the beach in three-quarter view, that is three transformations, not one. Every technique in this library is one idea repeated: a way of shrinking the delta until the model can clear it.

Chaining is the literal expression of that law. You cut one impossible jump into several small ones and check each landing. But the law does not actually say "always chain." It says never ask for a jump the model cannot make. When the jump is already small (a sharp reference and a simple placement), one shot clears it in a single move, and adding steps only introduces new places for the product to drift. The law is about the size of the delta, not the number of prompts.

When one-shot is correct (and we can prove it)

The old doctrine said one-shot generation collapses for anything with real complexity. We ran a controlled eval against our own doctrine and it flipped. With a sharp reference, one jump beats chaining decisively. On the glasses test (a tortoiseshell frame with legible temple text, 48 generations, 36 blind verdicts from a judge that never sees which method produced which image), the one-jump route passed 100% best-of-three versus 25% for the chained routes, and carried about 0.7 defects per image versus about 2.8 for the two-step chain. Every extra chain step re-renders the product, and identity drift accumulates: defects tripled going from one step to two.

The reason is the reference, not the method. Round one of the same eval scored 0% on every route, and the culprit was a blurry-temple packshot: the model could never reproduce details the input never showed, and the judge unfairly penalized them. Sharpen the reference and the one-jump route jumps to 75% strict fidelity. The input, not the method, is usually the bottleneck. If you find yourself chaining to rescue a soft or low-res source, you are treating the wrong problem. Fix the input first (see Product Accuracy with Nano Banana for how).

So the house rule, now deployed inside our assistant Awa, is: sharpest possible reference, one jump, three variations. Do not strawman one-shot. When the delta is small, it is not the lazy option, it is the accurate one.

When a pipeline beats a prompt

Complex images aren't one problem. They're five sequential problems. When the delta really is too big for one move (a product that must be rotated into an angle it was never photographed from, several products composited into one scene, a structural edit the compositor refuses to touch), a planned pipeline is the correct build order. You separate the hard problems and solve them one at a time instead of asking a single exposure to nail all of them.

The canonical build chain is five steps: select your ingredients (clean portrait, clean packshot, scene, style refs), composite model and product, place that composite into the scene, apply style and action, then quality check. The step most people forget is the composite: model plus product, on its own, before any scene exists. That is where you actually solve the hard problems, so validate accuracy there before you spend a single generation on the background.

It mirrors real photography. No photographer sets up model, product, lights and set at once for one exposure. On the February eyewear drop I tried to merge scenes, models and glasses in one shot and burned about five hours before admitting the obvious: "I was trying to do too many things at once. For complex objects like glasses, you need to simplify to the max. You also need to avoid introducing visual pollution early." Skip the composite step for simple products (a t-shirt, a bottle). Do it religiously for eyewear, jewelry, transparent materials and small branded details.

The gate discipline: check before you carry

A pipeline is only as good as its gates. The discipline is a hard stop at every stage: is the product right? The fit right? The face right? Fix it HERE. A mistake carried forward poisons everything after it. If the label text is wrong at the composite stage and you push on to lighting and scene, you are now polishing a defective product, and every downstream generation inherits the flaw. The gate is where you catch it while it is still cheap to fix.

This is why we separate the stages into different sessions: models in one session, scenes in another, composites in a third. Context bleeds between stages and pollutes them, so keeping them apart keeps each generation clean. Ship the finished image only through a quality gate: score all six ingredients one to five, and ship only at four or above on every one. The gates are not bureaucracy. They are the reason one-shotting a complex image once ate eight hours while the pipelined version ships in a fraction of that.

Debugging a failing chain: Shannon Descent

Sometimes a chain breaks and you cannot tell which step did it, or every fix inside the full scene fails anyway. That is when you switch from building to diagnosing. Shannon Descent means shrinking the problem until it can't hide: throw away everything except the smallest piece that still fails, perfect it alone, then rebuild the scene around the solved anchor, checking it survives each step. It is named for Claude Shannon's method of reducing a big problem to its smallest part and progressing in tiny increments.

It works because a complex scene is a crowd of instructions competing for attention, and the piece you care about may get too small a share. Isolated, the model pours all its attention into it, and you can finally see whether it is right. On a mahjong lifestyle scene, after several failed full-scene attempts, I stopped: "ok I think I'm too ambitious, let's try something else. Using image one create a visual of these sets of tiles on a green table. That's it, forget the rest. Just the table for now." Once the table was solid it became the anchor, and the scene got rebuilt around it. You can shrink surprisingly far, down to a single element or even a single instruction inside the prompt. The sequential pipeline is the planned build order for a known-complex shot; Shannon Descent is the same instinct applied after something already broke.

Draft cheap, finish expensive: the culling half of the rule

The three-variations part of the house rule is not decoration, it is the whole reason one-shot wins in practice. Accuracy work is probabilistic. A 90% discard rate is not a failure of the prompt, it is the visual filter working exactly as intended. On one eyewear drop the real count was 137 generated, 27 shortlisted, 12 delivered. You are not trying to get the perfect image on the first spin, you are trying to give yourself enough clean shots that at least one clears the bar.

So run two dials. Draft cheap: explore with one image on a cheap model to find the prompt. Finish expensive: generate three or four on your best model, then cull to the flawless ones. In the eval, best-of-three culling is exactly what lifted the one-jump route from 75% strict fidelity to 100% per-chain. One shot plus three variations plus a hard cull beats a careful chain, because the chain spends its extra steps accumulating drift while the cull spends its extra generations buying you a clean winner.

How to choose, in one line

Measure the delta before you decide, and know that one-shot is always worth a try: it is cheap to test, and the eval says it wins far more often than the old habit assumed. If your reference is sharp and the ask is a simple placement, one-shot from that reference, generate three, and cull. Chain when you need more control than one prompt can hold: a consistent model built separately before the composite, a created angle, a multi-product build, a structural edit. And when a one-shot fails, do not keep re-rolling the same prompt out of hope; that is the moment to consider chaining, breaking the ask into steps and gating each one. If a chain breaks and you cannot see why, drop into Shannon Descent and isolate the smallest failing piece. The mistake is not picking one camp for life. The mistake is chaining a small delta out of habit, or one-shotting a big one out of hope. For the full route map of which method matches which gap, see the accuracy route map and the control vs variant pipeline for how the culling loop runs.

Related Resources.