Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 3 min read

AnyLight: Letting Blender's Physical Renderer Decide the Lighting

AI lighting is a lottery. Ask for a specific light direction and the model gives you something plausible, not something specific. AnyLight replaces that guess with an actual physics render: build the light in 3D, render

AI lighting is a lottery. Ask for a specific light direction and the model gives you something plausible, not something specific. AnyLight replaces that guess with an actual physics render: build the light in 3D, render a reference, and let the AI model match it.

The graph

Node Value
UNETLoader qwen_image_edit_2511_bf16.safetensors
CLIPLoader qwen_2.5_vl_7b_fp8_scaled.safetensors, type qwen_image
VAELoader qwen_image_vae.safetensors
LoraLoaderModelOnly Qwen-Image-Edit-2511-Lightning-8steps-V1.0-fp32 at 1.0
EmptySD3LatentImage 928 Γ— 1664

62 nodes.

fp32 rather than bf16

The Lightning LoRA is loaded as fp32, not the bf16 used in other releases.

That is a notable choice for a distillation LoRA. FP32 is twice the memory of BF16 for this component, and the only reason to accept that is if the reduced precision measurably changed behaviour. For a lighting-control task where subtle gradients matter, that is plausible.

If you have memory pressure, BF16 is the first downgrade to try β€” but expect a difference, not equivalence.

Eight steps, not four

Lightning-8steps-V1.0 β€” eight steps, versus the four-step Lightning variants elsewhere.

Eight steps for a lighting edit makes sense: the task requires redistributing shading across every surface in the image, which is a global operation. Four steps cannot converge on a global change without visible banding.

What AnyLight actually does

The idea is to use Blender's renderer as the source of truth for lighting:

  1. Build a rough 3D scene matching the image's geometry
  2. Place lights physically
  3. Render
  4. Use the render as a lighting reference for the AI model

The AI model then has an unambiguous target: light from this direction, this intensity, this colour, with these shadows. No prompt guessing.

That is why the title claims an end to randomness. The claim is structural, not a quality claim about the model.

The cost

Building a 3D scene that matches an image's geometry is real work. For a simple object on a table it is minutes. For a person in a complex environment it is an afternoon.

So the technique pays off when:

  • The lighting must be exactly right
  • Multiple images need consistent lighting
  • The scene geometry is simple enough to model quickly

And it does not pay off for one-off images with forgiving requirements.

Where this breaks

  1. Geometry mismatch. If the 3D scene's proportions differ from the image, the shadows land in the wrong places and the model produces something incoherent.
  2. Occlusion. The release claims complex occlusion works; that depends entirely on modelling it in 3D first.
  3. 928 Γ— 1664 is fixed β€” other aspects get implicit cropping.
  4. fp32 LoRA costs memory.
  5. 62 nodes plus a 3D scene is a heavy setup for a lighting change.

Practical approach

Start with a simple scene β€” an object on a surface with one light. Confirm the pipeline works and the shadow direction matches your Blender render. Then increase complexity. The first attempt on a person in a room will fail for geometry reasons, not AI reasons, and that is easier to diagnose on a simple scene.

Workflows

One graph, published on RunningHub with invite code rh-v1111: AnyLight lighting control with Qwen 2511.

Files

https://huggingface.co/AI-KSK/ksk-video-165-comfyui-workflows

Notes

  • Workflow file only. Model weights are not included; each keeps its own licence.
  • Tutorial export, not a product.
  • Node count and loader values quoted above are read from the published JSON file.
  • Independent AI-KSK release. Not affiliated with Black Forest Labs, Blender or RunningHub.
πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.