Render Engines
Which model to run, and when — at a glance.
This table is the short version: what each engine is good at, how long it takes, and the one thing worth knowing before you pick it.
The first split is the only one that really matters. Engines w/ edit transform your input image — they keep the design and change the dress. Generation engines invent — they read your words and make a new image. If you have geometry to protect, you want the first kind.
| Engine | Quality | Best for | Tip |
|---|---|---|---|
| flux2-pro-edit | High | Everyday edit runs on a massing or model view | Takes several reference images — wire the design plus its context and it uses both |
| flux-kontext | Good | A single clean image you want changed, quickly | One image only; wire nothing else to IMG or it will ignore the rest |
| qwen-image-2-edit | Good | A fast second opinion on the same input | Reads material language well; keep the Structure row short |
| seedream-5-edit | High | Material and light studies with a strong result out of the box | Its default look is warm — calm the Aesthetic row if it romanticises |
| nano-banana-2-edit | High | The geometry you must not lose | The GeoLock default, and the reason: it holds geometry in both capture modes |
| nano-banana-lite | Good | Quick looks while you are still deciding | The fastest engine in the edit roster; not the one to deliver from |
| gpt-image-2-edit | High | Complex instructions that need to be understood | The slowest edit engine — send it one considered run, not five guesses |
| nano-banana-pro-edit | Very high | Style transfer where the look matters more than the outline | It may reinterpret geometry — that is why it is not the GeoLock default |
| flux2-pro | High | Text-only images with a controlled, architectural register | Set frame and camera in Tune; it follows structure better than adjectives |
| seedream-4.5 | High | Atmosphere-led text-only frames | Strong mood, loose geometry — early exploration, not documentation |
| nano-banana-pro | Very high | The best-looking text-only frame in the roster | Worth the wait for a hero image; overkill for a study |
| image-2.0 | High | Long, literal briefs where every clause must land | Slow and literal — write the brief you actually mean |
How to read the table
Speed is the expected run time, not a promise: it drives the Run button's progress fill, which approaches its end and holds until the real result lands. An estimate that lies would be worse than none.
Quality is house judgement about the finished frame, not a benchmark. Best for is the situation to reach for it in. Tip is the thing you would otherwise learn by wasting a run.
Choosing an engine
- Protect the geometry → an engine w/ edit, and by default
nano-banana-2-edit. - No geometry yet → a generation engine; see Text to Render.
- Several inputs to combine → an engine that takes many image inputs, not one that takes one.
- Still deciding → the fastest engine in the roster. Decide first, then spend the long run on the frame you keep.
House guidance above is a first pass and awaits Kevin's own words per flagship engine.
Upscale Engines
More pixels — faithfully, or creatively. Pick on purpose.
Upscaling is the last step before a print or a board, and it is where a good render can become a different one. The three engines differ in exactly one respect: how much they are allowed to invent.
| Engine | Quality | Best for | Tip |
|---|---|---|---|
| aura-sr | Good | A fast 4× when the render is already right | Fixed 4×, no settings, no reinterpretation — the safe default |
| seedvr2-upscale | High | Presentation and print, faithfully | Crisp and faithful restoration at 2× or 4×; the one to reach for by default |
| clarity-upscaler | Very high | A frame that needs more detail than it has | Creative diffusion: it can invent detail. Lower creativity, raise resemblance, and check the windows |
Faithful versus creative
A faithful upscaler makes the same image bigger. A creative one makes a plausible bigger image — inventing texture, mullions and foliage that were never in the frame. On a landscape that is usually welcome; on a facade you specified it is a change you did not ask for.
Check three places after any creative upscale: window divisions, small repeated elements, and any text. Those are where invention shows first.
How it composes
On the image, not on the rail. Upscale is a verb on the result's own under-block toolbar, like every other operation on imagery.
Before Download. Upscale, check, then download at full size.
After Modify, not before. Editing a 4× frame is slower and no more precise. Fix it small, then enlarge.
3D Engines
Imagery into geometry — one view or several.
Two engines, one question: how many views of the object do you have? A single-image engine infers the back of the building; a multi-view engine is told. Everything else follows from that.
| Engine | Quality | Best for | Tip |
|---|---|---|---|
| tripo-h3.1 | High | One clean image of one object | The default. Clay tier unless the texture is the deliverable — textures pollute captures |
| hunyuan-3.1-pro | Very high | An object you can show from several sides | Wire the extra views; given one image its advantage disappears |
Tiers
Geometry is the product. Geometry only (clay) is the default for new blocks and the right choice almost always: the render chain downstream is what dresses the object, and an untextured mesh captures cleanly. Textured tiers exist for when the mesh itself is the deliverable.
What a generated mesh is
- It is an interpretation of an image, wrong wherever the image was silent — usually the back and the base.
- It is massing, not documentation. Its job is to give the render engine a geometry to protect.
- It arrives in the workspace, so it can be placed on a fetched site, orbited, lit and captured like anything else.
How it composes
In from Image to 3D: wire one image, or several for a multi-view engine.
Out to 3D to Render: set the view, capture, wire the capture on.
Engine names above are the internal ids; public naming awaits Kevin's word.
Video Engines
A still becomes a clip — how long, how big, and how much invention.
Video is the slowest and least forgiving thing in the app: the engine invents a second of the world for every second of clip, and everything it invents is a chance to redesign your building. Short clips, one intent, and a still you already trust.
| Engine | Quality | Best for | Tip |
|---|---|---|---|
| veo-3.1-lite | Good | A first look at whether the motion works at all | Test the move here before committing to a slower engine |
| kling-v3-pro | Very high | The delivered clip — the most controlled motion in the roster | Takes a closing frame: a transition between two of your own stills beats a described one |
| kling-v3-turbo-pro | High | Longer clips where the camera does one simple thing | No closing frame — the motion must be describable, not drawn |
| veo-3.1-fast | High | Clean camera moves at 720p or 1080p | Resolution is a real setting here; set it rather than upscaling later |
| seedance-2.0-fast | Good | Quick variations of the same move | The fastest engine in the video roster; good for choosing a move |
| seedance-2.0 | High | Atmospheric motion — weather, water, light | Its largest resolutions take the longest; pick the one you will actually deliver |
Duration and resolution
Both are engine-specific — a duration one engine offers another refuses, and only some engines take a resolution at all. Set both on the block before running; the estimate on the Run button already accounts for them, because a clip's wait grows with its length.
One intent per clip
A camera move and a weather change and people arriving is three clips, not one. The first second is where invention shows: if the facade reorganises itself as the camera starts, shorten the clip and calm the move before touching the prompt.
How it composes
In from any finished still — see Still to Motion. Motion inherits every fault of its input frame, so fix the render first.
Out to History like every other run: nothing overwrites, and the still you started from stays exactly where it was.
Public engine guidance for the flagship video models awaits Kevin's word.