The Consistency Problem
AI image generation has a consistency problem. You generate an image. It looks great. You try to generate a second image that matches — same subject, different context. It looks like a different subject. The face changed. The colors drifted. The product proportions shifted. You are playing the prompt lottery, and the house always wins.
Nano Banana, Lovart’s image model family, was engineered to solve this. But the tools only work if you use them correctly. This guide covers the five techniques that turn inconsistent, unpredictable generation into a reliable production pipeline.
Technique 1: Anchor Your Generation — Never Start From a Blank Prompt
The most common consistency killer: describing your entire vision in one prompt and hoping the model nails it. Instead, build in layers.
The wrong way:
*”A brand mascot — a friendly fox wearing a barista apron — in a coffee shop, warm lighting, holding a latte, rustic wooden counter, chalkboard menu in background, plants on shelves, morning atmosphere, editorial photography style.”*
The model must solve for the fox, the apron, the coffee shop, the lighting, the latte, the counter, the chalkboard, the plants, the atmosphere, AND the photography style — all simultaneously. It will compromise on everything.
The right way:
1. **Anchor image:** *”A friendly fox mascot character. Simple, clean design, standing pose, front-facing. Vector illustration style.”* → Evaluate the fox. Is the face right? The proportions? The style?
2. **Add context:** *”Place this fox in a coffee shop. Warm morning lighting. Rustic wooden counter in foreground. Keep the fox’s design identical.”* → The model now generates background and lighting around an already-established character.
3. **Add details:** *”Add a barista apron to the fox. Forest green fabric with a small coffee bean logo. Keep everything else identical.”*
4. **Add props:** *”Place a latte cup in the fox’s hand. Simple white ceramic, latte art on top.”*
Each layer introduces one or two new elements while anchoring to what was already established. The consistency comes from the anchoring, not from prompt engineering. For a deeper guide on this technique, see our [over-prompting article](/blog/over-prompting-trap-novel-length-prompts-confuse-generative-ai).
Technique 2: Identity Lock — The Nuclear Option for Consistency
When you need absolute, guaranteed consistency — a brand mascot across 20 campaign images, a product across every e-commerce shot, a character across a storyboard — use **Identity Lock**.
How it works:
1. Generate or upload the definitive version of your subject — the one perfect image that you want all other images to match.
2. Enable Identity Lock (lock icon on the ChatCanvas toolbar, click your subject).
3. Nano Banana Pro extracts a mathematical fingerprint of the subject’s identity — facial structure, proportions, key features.
4. Every subsequent generation with Identity Lock active will preserve the subject’s identity. You can change everything else: lighting, angle, outfit, background, style.
When to use Identity Lock vs. anchored generation:
Pro tip: Run a “consistency check” after enabling Identity Lock. Generate the subject in 3 different contexts. Compare side by side. If any drift is visible, refine the lock by uploading an additional reference image of the same subject from a different angle.
Technique 3: Brand Kit as a Passive Consistency Layer
Identity Lock keeps specific subjects consistent. **Brand Kit** keeps your entire visual system consistent — colors, typography, lighting quality, compositional style — across every generation, even when the subjects change.
Setup (5 minutes — see our [Brand Kit setup guide](/blog/brand-kit-setup-5-minutes-lovart-best-practice)):
1. Define hex colors (exact palette).
2. Set typography preferences.
3. Upload 3-5 visual style references (brand photography, mood board images).
Once configured, Brand Kit is passive — you do not need to reference it in prompts. Every generation inherits your brand rules automatically.
The consistency multiplier: Brand Kit + Identity Lock + anchored generation = near-perfect consistency. Brand Kit handles the visual system. Identity Lock handles specific subjects. Anchored generation handles the build-up. Combined, they eliminate the prompt lottery.
Technique 4: Reference Stacking — Use the Canvas as Visual Memory
The ChatCanvas is not just a workspace. It is a visual memory system. Every image on the canvas is a reference that the Design Agent “sees” and can incorporate into new generations.
How to use reference stacking:
1. Generate your anchor subject. Place it on the left side of the canvas.
2. Generate a color palette image (abstract shapes in your brand colors). Place it below the anchor.
3. Generate a mood reference (lighting example, texture sample). Place it on the right.
4. Now prompt: *”Generate a lifestyle image using the character from the canvas, the color palette from the palette reference, and the lighting quality from the mood reference. The character should be in a modern home office, natural window light, relaxed pose.”*
The Design Agent reads all visible canvas content as context. The more visual information you give it — arranged spatially, not crammed into a text prompt — the more consistent the output.
Pro tip: Create a “consistency canvas” — a canvas with your brand mascot, color palette, logo, and style reference permanently on it. Use this canvas as the starting point for every new campaign asset. Each generation inherits visual context from everything on the canvas.
Technique 5: Session Memory — Stay in One ChatCanvas Session
Every new ChatCanvas session is a blank slate. The Design Agent has no memory of previous sessions. This is by design — it keeps sessions fast and isolated.
But within a single session, the Agent remembers. It remembers your brand’s colors because Brand Kit is active. It remembers your subject because Identity Lock is engaged. It remembers the visual context because references are on the canvas.
The rule: One campaign = one ChatCanvas session. Do not close the tab and start a new canvas for each asset. The consistency tools (Brand Kit, Identity Lock, anchored generation, reference stacking) work within a session. If you start a new session, you must re-establish the context.
If you must work across sessions: Save your canvas as a template. Export your Brand Kit configuration. Re-upload your reference images. Identity Lock embeddings persist across sessions because they are stored with your account.
FAQ
Q: How many Identity Lock subjects can I have in one session?
A: Multiple. You can lock one character, one product, and one logo simultaneously. Each locked subject appears as a pinned reference on the canvas. When you prompt, specify which locked subject to use: *”Generate a scene with the locked fox mascot and the locked coffee product.”*
Q: What if I want variation but still within brand consistency?
A: Generate multiple variants of the same prompt. Brand Kit ensures they all adhere to your palette and style. The variation comes from the model’s natural randomness within those constraints — different compositions, slightly different lighting nuances, different poses. Pick the best.
Q: Can Nano Banana 2 match Nano Banana Pro’s consistency?
A: Nano Banana 2 does not have Identity Lock — that is a Pro feature. But NB2 responds well to anchored generation, Brand Kit, and reference stacking. For non-character content (product shots, lifestyle imagery with ambiguous subjects), NB2 consistency is comparable. For character-driven content, Nano Banana Pro with Identity Lock is necessary.
Internal Links
| Anchor Text | Target |
|————-|——–|
| Brand Kit setup guide | `/blog/brand-kit-setup-5-minutes-lovart-best-practice` |
| Touch Edit guide | `/blog/touch-edit-best-practice-3-gestures-lovart` |
| over-prompting article | `/blog/over-prompting-trap-novel-length-prompts-confuse-generative-ai` |
| Nano Banana complete guide | `/blog/nano-banana-ai-complete-guide-lovart-image-model` |
| ChatCanvas getting started | `/blog/05-pillar-getting-started-lovart` |
*Best Practice article for blogs.lovart.ai.*
