Make a book light: the two-stage lighting cue for AI portraits (2026)

The trick: make a softbox out of two surfaces

When an AI portrait looks evenly lit but lacks shape, add a book-light cue: aim the key light into a large white surface, then send the bounced light through a diffusion panel before it reaches the subject.

The arrangement has two stages. The bounce surface spreads the light; the diffusion panel softens that bounced source again. The result is broad illumination with a gradual shadow transition across the face. The name comes from the open-book shape made by the bounce and diffusion surfaces. 1

A side-by-side illustrative portrait comparison: direct softbox light on the left and softer book-light illumination on the right — *Illustrative AI-generated comparison. Compare the cheek gradient, the shadow under the jaw, and the highlight on the jacket. The image is a visual example, not a measurement of model performance.*

Copy-paste prompt

Insert this clause into an existing portrait, fashion, or product prompt:

book-light setup: a key light aimed into a large matte-white bounce surface at camera-left, the bounced light passing through a broad translucent diffusion panel before reaching the subject, the diffusion panel angled toward the subject like an open book, broad wraparound illumination, gentle directional falloff across the face, soft-edged shadows, controlled spill, the light source and both modifiers outside the frame

The phrase that carries the technique is key light aimed into a large matte-white bounce surface ... the bounced light passing through ... diffusion panel. The phrase gives the model an action chain rather than a mood label: source, bounce, second surface, subject, and visible result.

Use defined but gradual cheek shadow after soft-edged shadows when the portrait becomes too flat. Use very broad low-contrast wrap when you want the effect to stay quiet. Change one phrase at a time.

Why the cue works

A direct key light reaches the subject in one step. A book light makes the white bounce surface the first apparent source, then uses the diffusion panel as the second modifier. Photography guides describe the same sequence as bounced light followed by diffusion, with the setup producing softer and more even illumination. 2

The setup changes three visual relationships that image models can describe clearly:

Source relationship: the key points at a surface instead of at the face.
Surface relationship: the bounced light passes through a second, translucent surface.
Result relationship: the face receives broad wrap with a readable, gradual transition into shadow.

Keep the cue in the positive prompt. The prompt needs to describe the lighting mechanism and its placement together, because soft, flattering, and cinematic alone leave the light source and its path unspecified.

Tool-specific versions

Midjourney V8.2

Midjourney's current default version is V8.2. The official documentation accepts --v 8.2 at the end of a prompt, and Midjourney documents --raw as a way to reduce its automatic styling and give a detailed prompt more control. 3 4

editorial portrait of an adult man in a charcoal jacket, chest-up, three-quarter view, warm-gray studio wall, realistic photography, book-light setup: a key light aimed into a large matte-white bounce surface at camera-left, the bounced light passing through a broad translucent diffusion panel before reaching the subject, the diffusion panel angled toward the subject like an open book, broad wraparound illumination, gentle directional falloff across the face, soft-edged shadows, controlled spill, the light source and both modifiers outside the frame --v 8.2 --raw --ar 4:5

--raw belongs at the end with the other parameters. If the face looks evenly frontal, replace gentle directional falloff with defined but gradual cheek shadow. Leave the version and Raw settings unchanged while testing the lighting clause.

FLUX.2 Dev in ComfyUI

The current ComfyUI guide includes a FLUX.2 Dev text-to-image workflow and describes it as a workflow for enhanced lighting, materials, and realistic details. The guide also says that the text-to-image workflow requires no input image. 5

Use the full scene description in the positive text prompt:

Editorial portrait of an adult man in a charcoal jacket, chest-up, three-quarter view, warm gray studio wall, realistic photography. A book-light setup shapes the portrait: a key light is aimed into a large matte-white bounce surface at camera-left, and the bounced light passes through a broad translucent diffusion panel before reaching the subject. The diffusion panel faces the subject like an open book. The result is broad wraparound illumination, a gentle directional falloff across the face, soft-edged shadows, controlled spill, and natural skin highlights. The light source and both modifiers are outside the frame.

Load the official FLUX.2 Dev text-to-image template or your existing FLUX.2 Dev workflow, then change only the lighting description for the comparison. Keep the checkpoint, seed, sampler, steps, resolution, and LoRA settings fixed. FLUX.2 Dev has its own workflow context, so treat this clause as a positive prompt addition rather than as a Midjourney parameter string.

Stable Diffusion 3.5 Medium

Stability AI identifies Stable Diffusion 3.5 Medium as an MMDiT-X text-to-image model and recommends ComfyUI or Diffusers for local use. Its Diffusers example uses num_inference_steps=40 and guidance_scale=4.5; those values provide a documented starting context for an A/B test. 6

Positive prompt:

editorial portrait of an adult man in a charcoal jacket, chest-up, three-quarter view, warm-gray studio wall, realistic photography, book-light setup: a key light aimed into a large matte-white bounce surface at camera-left, the bounced light passing through a broad translucent diffusion panel before reaching the subject, the diffusion panel angled toward the subject like an open book, broad wraparound illumination, gentle directional falloff across the face, soft-edged shadows, controlled spill, natural skin highlights, the light source and both modifiers outside the frame

For a Diffusers run, begin with the model card's 40 inference steps and 4.5 guidance scale. Keep those values, the seed, image dimensions, sampler or scheduler, and any LoRA weights fixed between A and B. Keep your existing defect list in the negative-conditioning field; the book-light change belongs in the positive prompt.

Run the controlled A/B test

Freeze the workflow. Use the same subject, pose, camera angle, lens description, aspect ratio, seed, model, sampler or scheduler, steps, guidance, LoRA weights, and upscaling path.
Render A. Replace the book-light clause with large softbox at camera-left aimed directly at the subject, broad frontal illumination, low contrast.
Render B. Paste the book-light clause into the same position. Change no other words or settings.
Compare three zones. Check the near cheek for a gradual transition, the under-jaw shadow for a soft edge, and the jacket or wall for spill that remains secondary to the face.
Tune one phrase. Use defined but gradual cheek shadow for more facial shape or very broad low-contrast wrap for a flatter, cleaner result.

The comparison image above illustrates the intended direction: B should preserve the subject and framing while changing the width and softness of the light transition. Your model, checkpoint, seed, and workflow determine the exact pixels.

The one-line takeaway

When a portrait needs softer shape, prompt the light's two-step path: aim the key into a white bounce surface, then pass the bounced light through diffusion before it reaches the subject.

The cover and comparison image are illustrative AI-generated artwork.