You need 20 images for a campaign — same brand, same mood, same color palette. You generate them across a week, maybe across two different scenes, and when you lay them out together they look like they were shot by 20 different photographers. Lighting drifted, color temperature drifted, sharpness drifted, even composition logic drifted. Diffusion models do not have a persistent style state. Every generation samples freshly from the model. The fix is locking style with reference images, sref, LoRAs, and palette enforcement.
Common causes
Ordered by how often each is the actual root cause.
1. No style reference passed in
Asking for “consistent style” with words alone produces drift across seeds. The model interprets style words slightly differently each generation.
How to spot it: Are you using only text prompts across your batch? Yes = this is the root cause.
2. Different scenes need different prompt structures
A product hero shot and a lifestyle scene require different prompts. Different prompts give the model different starting points and style drifts even if you used the same modifiers.
3. Model or model version drifted
Generated half the batch on Midjourney v7, half on v6.1. Or generated half on Flux Pro and half on Flux Dev. Even within the same family, versions shift behavior.
4. Resolution and aspect ratio differ
A vertical poster and a square social card from the same prompt look stylistically different because the model fits the framing differently. Different aspect ratios change perceived lighting and depth.
5. Different seeds amplify subtle prompt differences
Two prompts that differ by one adjective produce different outputs at the same seed. Add seed variation on top and drift compounds.
6. Post-processing applied unevenly
Some images upscaled, some not. Some color-graded, some not. Even if the AI output was consistent, post drifts.
Before you start
- Define the style brief in concrete terms. “Warm editorial, golden hour, shallow DOF, muted earth palette” is concrete. “Premium aesthetic” is not.
- Pick or generate one anchor image that represents the target style. This is your style reference for the whole batch.
- Save the seed, prompt template, and model of the anchor.
- Confirm you can use the same model version across the entire batch. Lock the version explicitly.
Information to collect
- The anchor image (your style north-star).
- Style brief — palette, lighting direction, lens look, mood.
- Model and exact version. For Midjourney, the version number. For SDXL, the checkpoint hash. For Flux, Pro vs Dev vs Schnell.
- A list of every scene the batch must cover, with intended aspect ratio for each.
Step-by-step fix
Ordered by ROI. Steps 1 and 2 cover most batch consistency cases.
Step 1: Lock model and version
Pick one model and one version for the whole batch. Document the exact version number. If using cloud providers, screenshot the version selector.
- Midjourney: Same
--vflag on every prompt. - SDXL: Same checkpoint file. Note the hash.
- Flux: Same tier (Pro / Dev / Schnell) on every prompt.
- DALL-E 3 / Imagen 3: Same provider, same week if possible (providers update silently).
Step 2: Pass a style reference on every generation
Replace style-by-words with style-by-image:
- Midjourney: Use
--srefwith the anchor image,--sw 100-200for strong style adherence. Same sref URL on every prompt in the batch. - SDXL: Use IP-Adapter with the anchor image. Style strength 0.6-0.8. Same anchor on every prompt.
- Flux: Use Flux Redux or IP-Adapter equivalent in ComfyUI. Same anchor.
- DALL-E 3: Less consistent without sref-equivalent. Provide an explicit style description that mirrors the anchor.
Step 3: Build a prompt template and stick to it
Create a single prompt template with placeholders for scene-specific bits:
{scene description}, warm editorial photography, golden hour side-light,
shallow depth of field, muted earth palette, slight film grain,
shot on 35mm lens, ratio {ratio}, --sref {anchor URL} --sw 150Fill in only the scene description and ratio placeholders per image. Everything else stays constant.
Step 4: Train a LoRA for high-volume work
If the batch is 50+ images or the campaign runs for months, train a style LoRA:
- Collect 20-50 reference images that represent the target style.
- Train on Replicate, OstrisAI, or local Kohya for SDXL or Flux.
- Use the LoRA on every generation at strength 0.7-1.0.
- A LoRA holds style more tightly than sref across very different scenes.
Step 5: Lock palette in post
Even with sref or LoRA, small palette drift remains. Run every image through a palette-locking post step:
- In Lightroom, save the color grade as a preset and apply to every image.
- In Photoshop, use a Color Lookup adjustment layer with a fixed LUT.
- In Davinci Resolve, build a node graph and apply to all stills.
This catches the last 10% of drift the model could not lock.
Step 6: Audit before delivery
Lay all images in a grid at the same display size. Look for:
- Color temperature outliers (one image warmer or cooler than the rest).
- Sharpness outliers (one image softer or harder).
- Composition outliers (one image with a very different rule-of-thirds setup).
Mark outliers and either regenerate or correct in post.
How to confirm the fix
- Build a contact sheet of the batch at the final display size. Style outliers should jump out immediately.
- Use a color-picker to sample the same hue across multiple images. They should be within a small delta-E range.
- Show the contact sheet to someone who did not generate the images. Ask which image looks different. None should.
- Verify the deliverable looks consistent on the actual surface (web, print, social).
Long-term prevention
- For any campaign-scale work, build a style anchor and an sref or LoRA before generating.
- Maintain a prompt template document for the project. Update it once when the brief changes; never edit prompts ad-hoc.
- Lock model versions and document them. When a provider updates, retest the anchor and decide whether to switch.
- Run every batch through a consistent post-processing recipe. Make it a preset.
- For long-running campaigns, train a project-specific LoRA. Reuse forever.
Common pitfalls
- Trusting that “consistent style” in the prompt is enough. It is not.
- Mixing model versions mid-batch because a new version dropped.
- Forgetting that aspect ratio changes affect perceived style.
- Skipping the post-processing pass and assuming the model output is grade-accurate.
- Generating across many days without re-checking provider version updates.
FAQ
Q: How many reference images do I need to train a style LoRA? A: 20-50 high-quality references are enough for SDXL or Flux. Quality matters more than quantity — choose images that all clearly share the style you want to lock.
Q: Will sref work across very different scenes (indoor product vs outdoor lifestyle)? A: Style transfers well; subject matter does not. The anchor image determines style adherence — color, lighting, lens look — but does not force the subject to match.
Q: Can I use multiple sref images? A: Midjourney supports multi-sref. Use 2-3 anchors that all share the style. This reduces dependence on any single anchor’s quirks.
Q: Why does my batch drift even with sref at maximum weight? A: Sref at max sometimes fights with scene prompts and degrades both. Drop to weight 100-150 and add explicit style descriptors. Also lock model version.