Front-Rear Renders

[Devin_Walker]

Not sure if this is the right group or not, I"m new here. I can’t find Veras anywhere in the forums.

I’m trying to create 2 renderings. One from the Front and one from the Rear. I’d like the landscape and surrounding areas to match. Veras is creating the exact same image for the rear as it is the front. It won’t change the background based on the view angle. I can’t get it to change no matter how I change the prompt. Is there a solution to this?

I’ve uploaded the 2 images I’m trying to render.

Thanks,

[joo_mouro]

Hi @Devin_Walker , Veras forum is not yet here at Chaos, but they are migrating it. Meanwhile use the old forum: https://forum.evolvelab.io

Do you have the model of that building? Why not give several images as reference, front and back? Background here is empty, is that the problem and Veras did not create any?

Cheers

[Devin_Walker]

Yes I do have the model. It’s a Revit model. The issue is Veras isn’t keeping continuity between the 2 scenes. I’m still learning how to create the proper prompts. But can’t figure out the correct verbiage to tell Veras how to create the exact same landscape and make them identical.

[joo_mouro]

Veras needs a first and last image image for the video making but they should be related. For example an exterior image and an interior that is visible from that exterior so to link them with some reason otherwise it will hallucinate even with nano banana 2

Two exterior images front and back is a lot to ask in my experience.

Cheers

[Devin_Walker]

Sorry for the confusion. I’m not trying to create a video. I need 2 images. 1 from the front view and 1 from the rear view. And Id like the materials and landscape to match in continuity. I was told that is a simple ask for Veras. Or was I misinformed?

Thanks

[joo_mouro]

Hi again,

You need to establish one approved visual direction first and then use Veras’s persistence (or nano banana) controls when generating the second view.

Start with the front view and use a restrained prompt describing the architectural character, materials, weather, landscape and vegetation, and lighting.

Iterate until the façade, cladding, roof, windows, ground treatment, and general landscape are acceptable. Don’t go for the second view until you are happy with the first…

Save the successful result and record its seed and settings in case you are not using nano banana.

For the next step use the result as a design reference. This way you preserve your architectural design language and landscape. You can also use the first render as a material or custom reference, depending on your Veras version.

Generate the rear view: Switch to the rear camera without changing the model or just use the second image you uploaded. Keep the same prompt, lighting description, visual style, and relevant settings. Apply the approved front image as the design reference or Design Lock.

If the building is correct but the landscape changes too much, use Veras’s selection/edit workflow to regenerate only the landscape region. Baby steps here. Avoid asking the second view to invent an entirely different landscape. Describe it as the same site, with the same paving, terrain color, walls, trees, and weather conditions.

The main limitation is that your two images contain only building geometry. If the landscape is not represented in the model, Veras must invent it independently for each camera. That makes continuity much less dependable than continuity of the building itself.

Suggested prompt…
Use essentially the same prompt for both views:

Photorealistic architectural visualization of the same contemporary building and site, preserving the supplied BIM geometry and exact façade proportions. Consistent warm light-coloured mineral render, dark metal window frames, natural timber entrance elements, and grey stone paving. The front and rear views belong to the same property and must use the same material palette, landscape design, planting character, terrain colour, boundary treatment, weather, season, and time of day. Do not redesign the building, change window positions, add floors, or introduce different materials. Maintain realistic architectural scale and restrained landscaping.

For the rear view, add only:
Rear elevation of the same building, viewed from the opposite side. Continue the same site landscape and material language from the front view.

Try it and check if it’s getting closer to what you want. I can try also and compare results…

Hope it helps!
CheerS

[joo_mouro]

Started with first image the front facade, filled in ground terrain with landscape, then made the 2nd image also a reference (front facade) and tried 2 more for the back facade. Shoot yours to compare The prompt is the one on the previous reply with the added text for the second image and used Nano Banana 2 with Cinematic and Film Photography for style chips…

Cheers

[Devin_Walker]

Thanks, I’ll give it a shot once I get back to the office.

[Devin_Walker]

So the only real issue I see is that the landscape is actually identical for the front and back. That was the issue I was also having. In reality the landscape will be very similar, but not literally the exact same.

Thanks,

[joo_mouro]

Ok, so ask for it, prompt it. The best way is to describe both because it needs it otherwise will assume it’s similar. So the second prompt needs landscape design or do it just like you did with the house by modeling it.

In my perspective the more you model, the more accurate it will be according to your ideas. But ideas don’t always need accuracy, sometimes they need the opposite. And the more you prompt the more you should be to suggestions and hallucinations…

Cheers

[Aamara_Zafar]

Good breakdown, thanks. That accuracy vs. hallucination trade-off tracks with what I’ve been seeing too — the more you lean on the model to “guess” the landscape, the more it defaults to just mirroring whatever’s already in frame, which is probably why the front/back ended up identical in the first place.

One thing that’s helped me: prompting the front and back as two completely separate scenes rather than sequential edits of the same one — different time of day or sun angle for each, even if it’s subtle. Forces the model to actually regenerate the landscape instead of pattern-matching the first render. Slower, but the results read as two distinct sides of the same building instead of a copy-paste.

Would be curious if modeling the landscape topology first (even roughly) and feeding that in as reference gets you more consistency without losing the “happy accident” variation that makes these useful for early concept work.

[BarryFishFinger]

May not be much help but in my experience using Veras (Nano Banana) what I find works is not to try to get it to do too much with a single prompt. Despite clear instructions it still hallucinates occasionally or gets it wrong so I’ve found the best way for my workflows is to generate images in a couple of passes and then composite them together in Photoshop. So my workflow is typically:

1. Generate a standard Enscape render from the model so I have control over gemoetry, lighting, textures etc

2. Ask Veras to do a smaller amount of simple things, eg add a background, or add people. Save that image.

3. As Veras to do the next thing, eg add foreground or whatever.

4. Load all into Photoshop and then erase the bits you don’t want from each layer.

So for the house example, I’d probably focus on whatever you want it to do for the house first until its consistent between front and back, then get it to do the gardens and put the two together. This way its not relying too much on the pattern of the first image.

The downside to this method is of course using more render credits.

[Devin_Walker]

unfortunately I don’t have enscape or Photoshop. I’ll just keep messing around with it.

Thanks,