Edit a previous generation instead of restarting

The most expensive habit in AI generation isn't picking a pricey model. It's rewriting the whole prompt every time a result comes back 80% right. You get an image where everything works except the lighting, and instead of saying "make the lighting warmer," you go back to the prompt box, add three clauses about golden hour, remove something else by accident, and get back a picture that fixed the lighting and broke the composition. Now you're two generations in and further from the target than you were after the first one.

Versely's agent exposes three different responses to a not-quite-right result, and they are not interchangeable. Picking correctly is worth more than picking a better model.

The three-way fork

Response	Tool	Use when	What it needs
Edit	edit_previous_generation	The result is mostly right, one thing is wrong	A new prompt describing the change
Remake	remake_generation	The brief is right, this particular roll isn't	Nothing beyond the reference
Restart	A fresh generation	The brief itself was wrong	A rewritten prompt

The question that sorts them is not "how good is this?" It's "is the thing that's wrong in the output, or in the instruction?" A result can be terrible and still only need a remake, if the prompt was fine and the model simply produced a bad take. A result can be quite good and still need a genuine restart, if it's good at something you no longer want.

All three assume the generation is in front of you, which it often isn't, because the good one was three days ago. get_generation_history lists what was generated in the current conversation. For anything older, describe it rather than hunt for it: the agent matches your description against past prompts by meaning, not exact keywords, so "that golden retriever image I made last week" is a workable query. Finding something you made before covers the search path, including uploads and slideshows, which live in your library rather than your generation history.

When to edit

Edit when the defect is localized and you can name it in one clause.

edit_previous_generation takes the generation you're pointing at, a new prompt describing the change, and an optional instruction about whether to keep the existing style. You don't restate the original brief. You state the delta. "Make the lighting warmer." "Remove the text on the left." "Same shot, but the product faces the camera."

The reason this beats a prompt rewrite is that it keeps everything you didn't complain about. A rewritten prompt is a fresh roll of every variable at once, including the fifteen things that were already correct. An edit is scoped. You are trading a small amount of control for a much smaller blast radius, which is usually the right trade when you're one adjustment from done.

You also don't need to track IDs. Referring to "that last one" or "the video from earlier" is enough for the agent to resolve it out of your history. Editing or redoing a previous generation is the capability page for both halves of this.

Good edit-shaped complaints:

• One attribute is off (colour, lighting, time of day, expression, camera angle)
• One element needs adding or removing
• The subject is right and the treatment is wrong

When to remake

Remake when you have no complaint about the brief, only about the roll.

This is the case people most often mishandle, because a bad output feels like evidence of a bad prompt. Often it isn't. Generative models are sampling, and some samples are worse than others for reasons that have nothing to do with your wording. If you look at a result and your honest reaction is "that's just an unlucky one," rewriting the prompt is a superstition, not a fix.

remake_generation regenerates with the same or similar settings and gives you two dials:

• variation_strength controls how far the new take is allowed to drift from the original. Low when you want essentially the same shot rolled again; higher when you want a genuinely different interpretation of the same brief.
• use_different_model re-runs the same brief somewhere else. This is the underrated one. When a prompt is clean and the output still isn't landing, the mismatch is often between the brief and the model's strengths, not between the brief and reality. Prompt adherence varies enough between models that the same words get materially different obedience.

If you're rolling the same brief repeatedly and want to hold the sample fixed while you change one thing, that's what the seed is for on a fresh generation. Remake is for the opposite: keep the brief, change the draw.

When to genuinely start over

Start over when the brief is what's wrong, and be honest about how often that is.

Three tells that you're looking at a restart rather than an edit:

1. Your list of complaints has more than two items on it. Editing sequentially through four defects is four charges and a drift risk at every step. One good rewrite is cheaper.
2. The complaints contradict each other. "Warmer, but less orange, and also a different room" is not a delta, it's a new brief wearing a delta's clothes.
3. You changed your mind about the deliverable. If the shot is now for a different placement or a different audience, its framing, ratio and pacing all move. Nothing about the previous generation is worth carrying forward.

Before committing to a rewrite, it's worth getting a second read on what's actually wrong. review_generation checks a generated image or video against its original prompt using vision and returns a structured verdict: whether it matches the prompt, a list of issues, a severity of ok, minor or broken, and a suggested fix. A severity of "minor" with one named issue is an edit. "Broken" with a list is usually a restart. Getting AI feedback on a generation is worth running when you're genuinely unsure which fork you're on — it costs credits and a beat of latency, so it earns its place on a decision you're about to spend several generations on, not on every result.

The part nobody budgets for

All three paths are priced like a new generation on the model used. Editing is not a discount; it's a different instruction with the same billing shape.

That matters because it inverts the intuition people bring from traditional editing software, where a small tweak costs less effort than a full re-render. Here, effort is flat and accuracy is what saves money. A four-step edit chain that walks toward the target costs four generations. A single well-aimed edit costs one. The saving isn't in the mechanism, it's in fewer attempts, which is why the diagnosis step matters more than the tool choice.

If you want a sense of what your own attempt rate is doing to your budget, reroll rates and budgeting against the shots you throw away treats it as a measurable number rather than a vibe, and quality per credit approaches the same question from the model side.

A worked exchange

Start with a generated product shot that's close.

"That last image is close. Make the lighting warmer, keep everything else the same."

One edit, one delta, style preserved. It comes back warmer and the composition survives. Now the second look reveals the label is illegible:

"Same image, but make the product label sharper and readable."

Second edit, still scoped. Two charges, two improvements, no drift. Compare that to the alternative, where both notes get folded into a rewritten prompt and the model returns a warmer, sharper image of a slightly different bottle on a slightly different surface.

Now a case that isn't an edit:

"Regenerate that video, try a different model this time."

Nothing about the brief changed. The complaint is with the take, so it's a remake with a model swap, not a rewrite. And a case that's neither:

"Actually this needs to be vertical for TikTok, and it should feel like a home kitchen, not a studio."

Two structural changes at once. That's a new generation, and pretending otherwise costs you an edit charge on the way to writing the prompt you were always going to write.

FAQ

Does editing a previous generation cost less than making a new one?

No. It's priced like a new generation on whichever model runs it. The savings come from needing fewer attempts, not from a cheaper per-attempt rate. That's precisely why picking the right fork matters: a correct edit gets you there in one charge where a prompt rewrite might take three.

Can I edit a video the same way I edit an image?

Yes, edit_previous_generation covers both. But for anything structural on a video — trimming, restitching, changing the aspect ratio, adding captions or music — the editor is the better path, because you can check the result with a free 480p preview (subject to a short per-user cooldown) before paying for a final export. Regenerating the clip is for when the content is wrong, not the assembly.

How do I stop an edit from changing things I liked?

Say what stays, not just what changes. "Make the lighting warmer, keep the composition and the product angle exactly as they are" constrains the edit far better than "make it warmer" alone, and asking it to keep the existing style holds the treatment steady. If a specific detail keeps drifting across edits, name that detail explicitly every time.

What if two edits in a row both make it worse?

Stop editing. Two failed deltas in a row is strong evidence that the problem is in the brief rather than the output, and a third edit is likely to compound the drift. Go back to the last version you liked, run a review on it to get a named list of issues, and write one new prompt from that.