Element re-render vs masked inpainting

Masking is the slowest part of a retouch. Drawing a clean boundary around an object, feathering it correctly, and redrawing it three times when the client changes their mind is where the hours go — and none of it is the creative work.

Reve 2.1, released 9 July 2026, ships a precision element re-render that regenerates a single named object without one. You say which element, the model finds it and redraws it. That is a real reduction in labour, and it is also a different set of guarantees. Trading a mask for a name means trading a hard boundary for the model's interpretation of your noun, and there are jobs where that trade is clearly right and jobs where it is clearly wrong.

What each one actually constrains

The distinction is not "one is easier." It is what each technique makes impossible.

A mask constrains geometry. You have declared a region. Pixels outside it are guaranteed untouched, byte for byte. The model cannot change anything you did not select, no matter how badly your instruction is phrased. Everything about how masked inpainting behaves follows from that one guarantee.

Element re-render constrains identity. You have declared a thing. The model resolves which pixels constitute that thing, and it is free to adjust anything it judges to be part of the change — including the object's shadow, its reflection, the light it bounces onto surrounding surfaces. Nothing outside your named object is guaranteed untouched.

That difference is not a bug in either tool. Each one is the correct answer to a different question. The question worth asking before every edit is: do I need a hard boundary, or do I need the change to propagate?

The decision table

Situation	Better tool	Why
One clearly nameable object, unambiguous in frame	Re-render	The name resolves cleanly and you skip the mask entirely
Two similar objects in frame ("the red mug", and there are two)	Mask	Naming cannot disambiguate; selection can
Object is partly occluded by another	Mask	Re-render tends to grab the whole occluding cluster
Soft or fibrous edges — hair, smoke, glass, fur	Re-render	Mask edges betray themselves on exactly these boundaries
The change must update shadows and reflections	Re-render	A mask confines the change and leaves the old shadow behind
A fixed region must not move at all — barcode, logo, legal copy	Mask	Only a mask gives a hard guarantee
Same instruction across 40 assets with different layouts	Re-render	Masks are per-image and do not transfer
Client will want six variants of the same object	Re-render	Re-running a name is cheap; re-masking six times is not
You must prove nothing else changed (regulated or legal review)	Mask	The guarantee is the deliverable
The object is small, distant, or low-contrast against its background	Mask	Weak identity cues make the re-render miss

Read that table as two clusters rather than ten rows. Re-render wins when the object is easy to name and the change should spread. Masking wins when the object is hard to name or the change must not spread.

The shadow problem, which is why masks lie

Here is the failure that costs the most re-work, and almost nobody anticipates it.

Swap a product inside a tight mask — a bottle for a differently shaped bottle, say — and the model dutifully redraws the bottle inside the boundary. What it cannot do is update the contact shadow on the table, because the shadow is outside your mask. You get a new object sitting in the old object's shadow. The same applies to reflections on a glossy surface, to bounce light on a nearby wall, and to any occlusion the original object created.

The result reads as wrong before anyone can say why. It fails a client review with a comment like "it looks pasted in," which is exactly what happened.

There are three ways out, in ascending order of preference:

1. Widen the mask to include the shadow and reflection zones. Works, but you have now given up the tight guarantee that was the reason to mask, and you are hand-drawing an irregular region.
2. Run a second pass on the shadow specifically. Doubles the revision count.
3. Use element re-render and let the model propagate the lighting change, then verify the fixed regions afterward.

The third is usually right for product photography, and the verification step is not optional — it is what replaces the guarantee you gave up.

The disambiguation problem, which is why re-render misses

The mirror-image failure. "Change the red mug to blue" is a perfect instruction when there is one mug. With two mugs, or a mug and a red notebook, or a mug partially behind a laptop, the model has to choose, and it will choose confidently and sometimes wrongly. Worse, it may change both.

The tells that a scene is going to disambiguate badly:

• More than one object matches your noun, even loosely.
• Your noun is generic — "the box," "the bottle," "the sign."
• The object is occluded, so its outline is ambiguous.
• The object is small in frame or low-contrast against what is behind it.
• You are relying on a colour to disambiguate and something else shares that colour.

Two of those in one image and you should mask. The added specificity of "the mug on the left, in front of the laptop" helps and is worth writing, but positional qualifiers degrade quickly in cluttered scenes, and a wrong guess on a client asset costs more than the mask would have.

A workflow that uses both

In practice you do not pick one tool per project. You pick one per edit, and the sequence that works is:

1. Classify the edit first. Is the object nameable and unambiguous? Does the change need to propagate to lighting? Two questions, thirty seconds.
2. Default to re-render for anything you will repeat. Batch work, variant generation, and anything running one instruction across forty assets with different layouts — masks do not survive any of those, because a mask is drawn per image and cannot transfer.
3. Fall back to a mask on ambiguity or on a hard constraint. Any packaging face carrying a barcode, price, ingredient list or regulatory mark goes through a mask, full stop. What happens to barcodes and legal copy on generated packaging is the reason that rule has no exceptions, and fixing brand images without reshoots is the masking technique itself.
4. Verify the regions you did not name. After every re-render, check three untouched areas: a logo, a face, and any fixed numeric or legal copy. This is the check that replaces the mask's guarantee.
5. Hold the seed steady when you are iterating on phrasing. Changing an instruction and the seed at the same time means you cannot tell which one moved the output.
6. Reach for denoising strength before you reach for a different model. On masked edits it is the dial between "barely changed" and "entirely new content in the region," and most edits that come back too aggressive are a strength setting rather than a model problem.

Reve 2.1 itself is not in the Versely catalog, but the catalog's editing models cover both behaviours. FLUX Kontext is at the instruction-driven end, best image editing model is the ranked view across the whole set with credit costs shown next to each rank, edit a photo with AI is the agent route if you would rather describe the edit than choose the model, and the AI photo editor is the direct tool.

The practical takeaway is smaller than the feature framing suggests: maskless re-render removes the slowest step from roughly the half of edits where the object is easy to name and the lighting should follow. The other half still wants a mask, and trying to force those through a name is how you end up doing the edit twice.

FAQ

Does element re-render make masking obsolete?

No. It removes the mask from edits where the object is unambiguous and the change should propagate to shadows and reflections. Edits that require a guarantee that a specific region did not change — legal copy, barcodes, logos, regulated packaging — still need a mask, because only a mask provides that guarantee.

Why did my masked product swap look pasted in?

Almost certainly the shadow. A tight mask confines the edit to the object and leaves the original contact shadow, reflection and bounce light in place, so the new object sits in the old object's lighting. Either widen the mask to include those zones or use a maskless re-render and verify afterward.

How do I stop a re-render from changing the wrong object?

Add disambiguating detail — position, relationship to another object, and a distinguishing attribute together rather than any one alone. If two objects in the frame plausibly match your noun, or the target is occluded, small or low-contrast, mask it instead. The cost of a wrong guess on a client asset exceeds the cost of the mask.

Which is cheaper in credits?

Neither technique has a fixed cost advantage — the credit cost comes from the model you run, not from whether you supplied a mask. What actually drives an editing budget is revision count, which is why the tool that produces a usable result in one pass is the cheaper one regardless of its listed cost.