Problem summary
Midjourney's outputs can look like they're lit differently across renders even with the same prompt because the generator samples details stochastically, interprets vague lighting language differently, and may change composition or stylization between runs. You can reduce or nearly eliminate that variability by locking the seed, using a lighting reference image, being extremely explicit about camera and light, and fixing the generation settings.
Recommendation (short)
Create a “lighting master” image and prompt template, use --seed, low --chaos, fixed --ar and --stylize, and always include explicit camera + light phrases. Use a reference image with a high image-weight for other renders so lighting and color grading are reused.
Decision criteria (how to choose tactics)
- Fast iterations, low budget: use precise prompt templates + --seed + low --chaos; iterate with small batches.
- Production-quality or many matching images: create a master reference image (render/photograph) and use it as an image prompt, or switch to a 3D/base photo pipeline if exact consistency is required.
- Team workflows: store and share the prompt template, seed, and parameter values in your assets so everyone reproduces identical settings.
- Skill level: beginners should rely on an image reference + explicit lighting words; advanced users can add camera specs, film stock, and custom LoRAs or models.
Concrete checklist to lock lighting, angle, and grading
1) Make a Lighting Master
- Generate 1 perfect image with the exact lighting, shadows, camera angle, and color grade you want. Save the image and the final prompt + seed.
2) Use that master as a reference image
- Upload it to Midjourney and include it as the first prompt token. Increase its influence by repeating the image token or adjusting image weight (image prompt + textual prompt combo) so MJ leans on that look.
3) Fix the seed and generation settings
- Add --seed (same seed every run). Set --chaos 0–10 for minimal randomness. Fix --ar (aspect ratio), --stylize (lower if you want literal results, e.g., --stylize 50–250 in v5), and --quality (e.g., --quality 2 for higher detail).
4) Be brutally explicit about lighting and camera
- Example phrases: “three-point studio lighting, key light from camera-left at 45°, soft rim light behind right, long soft shadows, golden hour, warm key / cool fill”.
- Camera: “50mm lens, f/2.8, 1/125s, eye level, 3/4 view, medium close-up”.
5) Lock color grading in the prompt
- Specify film stock or LUT: “Kodak Portra 400, teal-and-orange grade, cinematic contrast, -2 highlights, +1 shadows”. Repeat these exact words across prompts.
6) Use negative prompts / exclusion words
- Add things you don’t want: “no harsh top light, no rim lighting, no overexposed highlights” to reduce accidental variation.
7) Batch-generate and select
- Generate a batch with identical settings and choose the best matches. When reworking, start from the chosen image and use it as the new reference.
Best-for / Avoid-if
- Best-for: projects where consistent mood, camera, and lighting across a series of stills is needed (character sheets, marketing assets, concept series).
- Avoid-if: you need wide, unpredictable variation or want many different lighting moods from one prompt.
Extra tips
- Use ChatGPT to help craft rigid prompt templates and to translate photography terms into prompts consistently. If you need pixel-perfect consistency for animation frames or product shots, move to a 3D pipeline or shoot a real reference photo and use that as the base.
Quick sample prompt template (combine with your image reference and seed):
"subject description" — "three-point studio lighting, key light camera-left 45°, soft fill camera-right, low-angle rim light behind right, soft shadows" "50mm, f/2.8, eye-level, medium close-up" "Kodak Portra 400, cinematic teal and orange grade, medium contrast" --seed 123456 --ar 4:5 --stylize 100 --chaos 5 --quality 2
Follow the checklist and keep a saved prompt+seed file—consistency becomes trivial once you standardize those elements.

Compare Midjourney and Leonardo AI