#Need help simulating realistic mirror-reflected sunlight for my solar-EV AI project

1 messages · Page 1 of 1 (latest)

lucid bay
#

Hi everyone,
I'm working on a solar-powered EV technology project and developing an AI system that will manage energy in real time. My work includes a full-body photovoltaic vehicle concept (Solar Active Skin) and photovoltaic wheel rims that turn the whole car into a multi-surface solar generator.

I already have one patent granted, a second patent application in progress, and early physical prototypes built.
Right now, I’m in the proof-of-concept and R&D preparation stage, using visual simulations, 3D models, and generative graphics to test AI logic and physical behaviors—especially light interaction.

🔍 What I need help with

For my AI system to work, I need accurate visual tests of how sunlight interacts with reflective surfaces.
I’m trying to generate a scene where sunlight reflects off a large mirror surface and illuminates the opposite, shaded side of the vehicle.

I attached one of my renders as an example.

The problem:
No matter how I phrase the prompt, the model always lights only the side facing the sun.
The opposite side—which should be illuminated by mirror-reflected sunlight—remains completely dark.

🔧 What I’ve already tested

  • placing the mirror farther away
  • enlarging the reflective surface
  • “mirror-reflected sunlight” prompts
  • forcing ray direction
  • describing angles of incidence
  • secondary bounce light
  • specular reflection
  • GI-related phrasing
  • material descriptions (“high-reflective mirror panel”)

But the model still refuses to simulate the second light source via reflection.

☀️ Why this matters for my project

My startup is building an AI system that will:

  • select the optimal parking location based on direct sunlight and reflective surfaces,
  • analyze reflections from walls, glass, bright vehicles, etc.,
  • predict real energy output from the photovoltaic surfaces,
  • choose driving routes based on real-time satellite weather data,
  • avoid cloud bands and storms to maximize solar intake,
  • reduce or eliminate the need for conventional charging during travel.

To validate this concept, I need accurate light-behavior simulations.
These visuals are part of my proof-of-concept, which precedes the next R&D phase of the real hardware.

Since I already have a granted patent, another in progress, and early prototypes built, getting these simulations right is important for the development roadmap.

❓ My question for the community

Has anyone found a reliable prompting method to force realistic mirror bounce lighting that illuminates the opposite side of an object?

Is there phrasing that helps the model respect:

  • secondary reflections,
  • specular bounces,
  • global illumination,
  • or treating the mirror as a secondary light emitter?

Any insight or tricks would be incredibly valuable for my R&D process.
Thank you so much for your help!

Sebastian

c646595c-2f63-498a-a988-40db257a3f70.png file_000000004e2472438f08b9735ef01bc9.png
small edge
#

Hey, I tried to explore this. The 'closest' I got was this not at all close, but kinda pretty, image.

I searched online, and could not find any image such as you describe.

I was considering, maybe I could find images like what you need online - those could be shown to the model, and it could emulate around them - and I did not find one image similar to what you describe needing online.

Suggesting, since you need this, it may not exist - and it may not be trained into the model - maybe consider creating your own, at least with one car and whatever mirrors or other reflective surfaces you can work with. Create mockups that use reality, and maybe show and teach the model itself. Not sure how good the image program would be at this sort of 'in-chat learning', but it might be the only option at this time.

content.png
lucid bay
small edge
dim mortar
#

Hi

lucid bay
lucid bay
# small edge You're welcome! If reality matters, I would suggest **always** consider explori...

I make Real model test.
Look: 1. The truck is illuminated by the sun from one side; the other side remains in shadow despite the presence of a mirror. No reflection effect — only direct sunlight.

  1. Shadowed side of the truck. No direct sunlight, only diffuse ambient light.

  2. Sunlit side of the truck. Strong light–shadow contrast, a classic example of unidirectional lighting.

  3. A large mirror reflects sunlight onto the previously shaded side of the vehicle. Reflected light begins to effectively illuminate the body.

  4. Reflected light appears near the truck but does not hit it. The vehicle remains in shadow — reflection alone is not enough, the angle matters.

  5. The truck is partially lit and partially shaded. Adjusting the mirror angle relative to the sun clearly increases illumination.

  6. Light reflected from a doorframe or edge. Shows that not only mirrors but also ordinary surfaces can redirect light directionally.

  7. Changing angles and time increase reflected light. AI can analyze the sun’s path, shadows, and surroundings (e.g. 10 hours of parking) to recommend the best parking position for maximum PV energy.
    9/10. Conclusion: generating all scenes at once is too complex — scenes should be generated and validated individually to preserve correct physics.

Screenshot_20251221-164819.png Screenshot_20251221-164831.png Screenshot_20251221-164833.png Screenshot_20251221-164836.png Screenshot_20251221-164909.png Screenshot_20251221-164915.png Screenshot_20251221-164925.png Screenshot_20251221-164929.png Screenshot_20251221-164947.png
lucid bay
#

Single-shot test 😄 One mirror, one car, one sun — still learning how to bend physics before scaling to 8 scenes at once.” i see progress 😁😀🪞

file_00000000154c72468a8c6f8aa4242562_1.png
small edge
lucid bay
small edge
#

I think handling realistic reflection is still really hard for the model, but perhaps not impossible with guidance.

I note that the 'same spot' on the reflected apple is shiny, as if duplicated from the object apple, on the upper left. I'd expect that part of the apple to be in shadow in the reflection.

image.png image.png
lucid bay
lucid bay
# small edge I think handling realistic reflection is still really hard for the model, but pe...

In this case, the model actually handled the radiation direction and lighting sides correctly. The issue is not the direct light itself, but the fact that the model still struggles with how shadows behave when they are illuminated by reflected light.
This is especially visible in the mirror reflection in image no. 5 — you can see the real physical process where reflected light partially eliminates or softens the shadow cast by the car, and the model does not yet correctly distinguish which parts of the shadow should disappear and which should remain soft.
In my opinion, this interaction between shadow and indirect (reflected) light is the key area to focus on, because the geometry and the main light directions are already largely correct.

file_00000000846c71f5bcf954bf8177cb7e.png file_0000000087e4722f9e6f62677040d30c.png Screenshot_20251221-164909.png
lucid bay
#

"Problems encountered while creating this image"
Incorrect mirror physics – the model copied the object instead of calculating a true reflection (wrong side of the car, perspective, and character positions).
Anatomy errors (hands) – hands appeared on the wrong side, were duplicated, or disappeared.
Inconsistent car geometry – mismatched angles, scale, and wheel placement in the reflection.
Lighting and shadow issues – reflections were too sharp or brighter than the main object.
OpenAI logo orientation – the logo was incorrectly flipped or distorted in the reflection.
Core issue: the model struggled to simultaneously maintain correct 3D geometry, mirror optics, and consistent anatomy.

file_0000000080e4722fb508283d38d612d6.png file_00000000622871f49ac3df12b0f1be82_22.png file_000000002978722fa0de8a38f976141a.png file_0000000069e471f4a959e125b3c20d5d_1.png
blissful void
#

hmmm

blissful void
lucid bay
#

in a physically correct model the reflected light should appear where I marked in yellow, not where it’s currently shown in red.
In reality, the mirror would redirect the incoming solar rays according to the angle of incidence, illuminating the ground in that specific region.
The issue here isn’t the concept of reflection itself, but that the model doesn’t yet autonomously reason about light vectors and mirror normals — it needs explicit, step-by-step guidance on where and how to place reflected rays.
This shows that realistic reflection still requires manual correction or detailed prompting, rather than being inferred automatically from physical rules.

Screenshot_20251223-2025022.png
#

@blissful void

blissful void
#

i tried. Image models are not yet capable of doing that

blissful void
lucid bay
# blissful void i tried again using the prompt from a custom GPT I made...this is the best I cou...

I appreciate the effort and the work that went into this, however I have to be clear: this is not a perfect, physically accurate mirror reflection, and when it comes to sunlight rays, it is even further from correct
In practice, the GPT model is guessing, rather than simulating real physical light behavior.
I am currently working on improving my generator, specifically toward achieving perfect, physically correct mirror reflection. Reflections in the chat must be real, not heuristic or prompt-based approximations. Yes, it is possible to generate many prompts to get one image that “kind of” works — but that is not scalable. This is simply a hard physical problem.
I am also considering generating a true mirror render, similar to how GPU-based engines handle it — for example in Blender, where I previously rendered mirrors and it worked very well. I believe GPT Image 1.5 could be extended with such a capability, because in its current form it has serious limitations when it comes to physically accurate mirror reflections.
A similar issue appears with basic geometric control. For example, the model struggles with setting an object’s orientation across the full 0–360° range. Instructions such as “tilt the mast tip by 8° toward the sun” are not properly understood or executed. I am attaching images to better illustrate this limitation.
I hope that future versions of image generators will be significantly improved in terms of physical accuracy, because without this level of realism my project cannot move forward. An ideal solution would be a true 3D space, where the model can perform a real internal render instead of operating purely in 2D.
From such a 3D space, the system could then output a 2D image, but with correct angular physics — including mirror reflections, shadows, light rays, and interactions between elements. Without access to true X, Y, and Z axes, physically accurate reflections and lighting cannot be achieved```

Screenshot_20260103-231657.png file_00000000aec471f5902845f4674154dd.png Screenshot_20260103-215518.png
blissful void
#

Well, the reason these LLM AI struggle is because they can't "feel" like we do. They don't understand that actions have consequences.

elfin mango
lucid bay
lucid bay
serene hedge
lucid bay
small edge
#

Wow, mirrors + shadows, but also mirrors period still seem a very hard problem.