Issue #11 | Platform: Nano Banana 2 (primary spotlight) / ChatGPT-Image 2.0 (established reference) | May 26, 2026
Three frames in. The face drifted.
Not slightly. The character who was wearing the red jacket in frame one is now wearing something different, standing somewhere else, lit from a different angle. The background stayed. The story did not.
[02] THE FIRST 80 WORDS
That moment, that specific failure, is the problem every creator is trying to solve right now. Not "how do I generate a good image." But "how do I generate the next image without losing what I already built."
The shift from single-image prompts to multi-image asset pipelines has forced two different disciplines onto the table at once. Google's Nano Banana 2 and OpenAI's ChatGPT-Image 2.0 each solve part of the problem. Neither solves all of it. The question is not which model wins. The question is what execution method you need to stop losing consistency between shots.
I noticed something while testing both back to back: the failures are not random. They follow patterns. And the techniques that fix them are specific enough to teach.
[03] PLATFORM SPOTLIGHT. Nano Banana 2 (Primary) + ChatGPT-Image 2.0 (Reference)
Platform: Nano Banana 2 by Google (primary spotlight) / ChatGPT-Image 2.0 by OpenAI (established reference). Nano Banana 2 launched publicly: May 2025. ChatGPT-Image 2.0 launched: April 2025. Behavioral observations noted: May 2026. Both platforms update on irregular release cycles. Verify before acting.
The Consistency Problem, Framed Honestly
Multi-image narrative consistency means one thing at the craft level: the same subject, same light, same spatial logic, across every frame in a sequence. Not approximately. Not "close enough." Exactly.
Single-frame generation has a generous margin for error. You generate, you pick the best one. A sequence does not work that way. Each frame is a dependency. When frame three fails, it usually fails because frame two already drifted from the anchor you set in frame one.
Both models get this wrong in their own way. Nano Banana 2 will hold an environment steady but let a character's details shift if you do not explicitly re-anchor them. ChatGPT-Image 2.0 holds characters better, but its "thinking" pass before generation means it will re-interpret your scene if your follow-up prompt is even slightly ambiguous. Two different drift mechanisms. Two different fixes.
[04] THE TECHNIQUE: Technique 1 The JSON Prompt Formula (Nano Banana 2)
Nano Banana 2 moves fast. Ten to fifteen seconds per asset is a real number for prototyping. The speed is not the problem. The problem is that fast generation plus a loosely written prompt equals fast drift.
The creators getting consistent results are not writing paragraphs. They are writing structured slots. This JSON structure does one thing that prose cannot do cleanly: it isolates fields. When you need to update only the lighting for the next frame, you change one field. The subject, camera, and background stay identical because they are sitting in separate slots, not buried inside a paragraph where the model has to re-parse what changed.
Reference anchoring runs on top of this. Nano Banana 2 accepts up to 14 reference images simultaneously. That means you can lock down a face, a background, a product, and a lighting reference all at once. Creators building multi-frame storyboards are using this directly. They are not describing the background. They are feeding the background back in as an image reference on every turn.
One flag: this approach works best for commercial-style output. Nano Banana 2 renders high-contrast, punchy images well. If you need something that reads like a photograph from a real shoot, you will hit its ceiling faster than you expect.
[05] THE TECHNIQUE: Technique 2 The “Preserve” Chain Method (ChatGPT-Image 2.0)
ChatGPT-Image 2.0 handles character retention through a pre-generation reasoning pass. Before it renders anything, the model thinks through your request. That thinking pass is what gives it better spatial layout reasoning and stronger character consistency on complex scenes. It is also what makes it sensitive to ambiguous follow-up instructions.
Here is the pattern that matters. On every single follow-up edit, you need three things in this order: what changes, what stays, and the physical constraint. Example: "Change the outfit to a white blazer. Keep the face, lighting, camera angle, and background identical. The blazer should have the natural weight and wrinkle of real fabric."
[06] CHOOSING THE RIGHT PIPELINE
These are not competing tools. They fit different stages of the same workflow.
Use Nano Banana 2 to prototype fast and lock down environmental consistency. Use ChatGPT-Image 2.0 when you need character fidelity across a sequence or when you are building something that has to look like it was shot in a real space.
The part that still does not fully resolve: both models drift more than they should when you extend a sequence past six or seven frames without reintroducing your anchor references. Whether that is a prompt architecture problem or a model limitation, I am not certain yet. My current practice is to re-feed the full reference block every five frames, regardless of how clean the last output looked.
The Prompt Stack for This Week
# Prompt 1 · Nano Banana 2 · Anchor Discipline Workflow
```markdown
Set your anchor frame before anything else.
Generate Frame 01 first.
Capture the output immediately.
Save the screenshot as:
ref_001
Create a structured JSON control block and reuse it for every subsequent generation.
{
"Subject": "Describe the subject in exact detail",
"Camera": "Lens, framing, focal length, movement, depth characteristics",
"Lighting": "Lighting direction, intensity, color temperature, mood",
"Background": "Environment, textures, atmosphere, depth layers",
"Reference_IDs": ["ref_001"]
}
Rules for all future turns:
- Every new frame must inherit the anchor frame.
- Reuse the JSON structure consistently.
- Preserve facial structure, wardrobe logic, object placement, and environmental continuity.
- Treat the reference image as the visual source of truth.
- Avoid introducing new visual variables unless intentionally specified.
Text rendering protocol:
- Any readable text must appear inside quotation marks.
- Define typography weight explicitly.
- Specify placement, spacing, and readability requirements.
- Never assume the model will infer typography consistency.
Re-anchoring protocol:
- Every five frames, re-anchor the sequence.
- Re-upload the latest reference images.
- Re-state the preserve list in full.
- Reconfirm identity consistency, lighting continuity, camera language, and composition rules.
- Re-anchor even if the previous outputs appeared visually correct.
Core principle:
The model may improve over time.
Anchor discipline remains constant.
```
# Prompt 2 · ChatGPT-Image 2.0 · Batch Cinematic Sequence Workflow
```markdown
Before generating any image, write the complete batch prompt.
The batch must include:
1. Every frame in the sequence.
2. Exact transitions between frames.
3. Character preservation instructions.
4. Environmental continuity rules.
5. Camera progression logic.
6. Lighting continuity.
7. Typography constraints.
Generate the entire sequence in one unified prompt instead of fragmented turns.
Structure your batch prompt like this:
PROJECT:
Define the cinematic or visual objective.
GLOBAL PRESERVE LIST:
- Character facial identity
- Hair structure and color
- Wardrobe consistency
- Accessories
- Environmental continuity
- Lighting logic
- Camera language
- Texture fidelity
- Composition rules
FRAME 01:
Describe the opening composition in full detail.
TRANSITION TO FRAME 02:
Describe camera movement, environmental changes, motion logic, and emotional continuity.
FRAME 02:
Describe the next composition.
Continue until the full sequence is complete.
Typography rendering protocol:
- Any readable text must be enclosed in quotation marks.
- Define font weight explicitly.
- Define text placement and hierarchy.
- Specify readability priority.
- Never rely on implicit typography behavior.
Sequence stabilization protocol:
- Every five frames, re-state the preserve list.
- Reintroduce anchor references.
- Reinforce character consistency and scene continuity.
- Reconfirm camera and lighting rules.
- Do this even if the outputs appear visually stable.
Core principle:
The generation quality evolves.
The anchoring system remains mandatory.
```
[07] QUESTION
When your image sequence breaks, is the drift usually happening on your subject, your lighting, or your background and which platform were you on when it happened?
[08] SOURCES USED IN THIS ISSUE
Google Nano Banana 2 announcement
OpenAI ChatGPT-Image 2.0
ChatGPT Images 2.0 vs Nano Banana 2 — Workflow Comparison
Weshop Nano Banana 2 tool overview
JSON Prompt Formula video
XDA Developers real-world comparison
Instagram field notes on structured prompting
Nano Banana speed test short
Artlist Nano Banana 2 model page
DesignMonks ChatGPT-Image 2.0 breakdown
ChatGPT Images 2.0 API prompting guide
Framia GPT-Image-2 prompt guide
Wired on OpenAI image generation
Batch storyboarding technique video
Reddit ChatGPT Images 2.0 field guide
TechRadar real-world prompt comparison
PCMag comparison review
[09] PLATFORM COMPARISON SNAPSHOT
Platform | Behavior with This Technique | Recommended Adjustment | Income Fit |
|---|---|---|---|
Nano Banana 2 | Speed: ~10-15 seconds per asset. JSON structured prompts isolate fields, preventing cross-frame drift. Reference anchoring with up to 14 simultaneous reference images locks face, background, product, and lighting together. Best for commercial-style, high-contrast output. Drift failure mode: character details shift between turns. | Use JSON prompt structure with isolated fields. Feed reference images on every turn. Prototype fast, lock environment first, then refine character. | Commercial product visuals, brand asset pipelines, multi-frame storyboards. Best for speed and volume at ~10-15 sec/asset. |
ChatGPT-Image 2.0 | Pre-generation reasoning pass before render. Character consistency is stronger than Nano Banana 2. Sensitive to ambiguous follow-up instructions. Batch storyboarding: up to 8 sequential frames from one detailed prompt. Naturalistic, photograph-like output style. Drift failure mode: scene re-interpreted on ambiguous follow-ups. | Use Preserve chain method on every follow-up. Batch all frames in one prompt. Define what changes, what stays, and physical constraints. Re-state preserve list every 5 frames. | Character-driven editorial sequences, multi-frame brand narratives, naturalistic photorealistic scenes. Best for final-quality output where consistency matters most. |
Strong style consistency without pre-planning. Anime prompts work well with --style and --ar flags. Less compositionaly coherent than Thinking mode on complex multi-element scenes. | Add --q 2 and --stylize 750 for richer outputs. Include style reference artist in prompt. | Character-focused editorial prints and atmospheric brand assets. Strong for artist-style work at $80-$200/image. | |
Adobe Firefly 3 | Responds well to descriptive style prompts. Commercially safe outputs by design. Less atmospheric than ChatGPT Thinking but strong for product and UI-adjacent visuals. | Add style reference and explicit lighting description. Avoid abstract mood words; use concrete visual descriptors. | Commercial product visuals and UI mockups for Adobe Creative Cloud clients. Best for brand-safe content at $100-$250/set. |
Google Gemini | Produces clean photorealistic outputs. Less strong on anime/stylized content. Better suited for product visuals and realistic scenes than atmospheric character work. | Focus on realistic environments and product detail. Reduce style layer. Add specific material and texture descriptors for best output quality. | Realistic product photography and environmental scene work for tech and lifestyle clients. $80-$200/image set. |
Grok (xAI) | Fast outputs with strong subject clarity. Less capable on atmospheric layering. Good for meme-adjacent and editorial concept work. | Use short, direct prompts with strong subject focus. Omit layered atmosphere descriptors. | Quick editorial concepts and social content for X platform clients. Low cost, high speed. $40-$100/set. |
Meta AI (Llama) | Free, accessible, produces consistent portraits and simple scenes. Struggles with complex atmospheric layering. Best as entry-level client work tool. | Keep prompt simple. Use clear subject + action + style format. Avoid multi-element scene descriptions. | Entry-level social media content for small business clients. Free tool advantage. $30-$80/image set. |
Platform-honest bottom line: Use Nano Banana 2 to prototype fast and lock down environmental consistency. Switch to ChatGPT-Image 2.0 when you need character fidelity across a sequence or when you are building something that has to look like it was shot in a real space.
[10] INCOME SPOTLIGHT CALLOUT
THE INCOME SIGNAL: Multi-image consistency prompting is now a billable skill. Deliverable: a 5-frame to 8-frame narrative sequence per client session. Client types: solopreneur brands, product creators, newsletter publishers, editorial directors. Revenue model: per-session rate ($150–$400) or monthly retainer ($250–$500/month for 2 sequences).
THE MINIMUM VIABLE WORKFLOW:
Step 1. Run the master prompt from this issue on ChatGPT with Thinking mode active. Set quality to high before submitting.
Step 2. Set your anchor. Generate frame one on Nano Banana 2 using the JSON formula. Screenshot it. Name it ref_001. Feed it back into every subsequent turn.
Step 3. Switch to ChatGPT-Image 2.0 for final character-fidelity sequence. Write your full batch prompt before generating. Include all frames, transitions, and preserve list. Send in one shot.
PLATFORM MATCH: Nano Banana 2 for fast environmental locking and prototyping. ChatGPT-Image 2.0 for final character-fidelity sequences and naturalistic photorealistic output.
[11] YOUR ASSIGNMENT BEFORE THE NEXT ISSUE
THE TASK: Build a 3-frame sequence using both platforms. Frame 1: establish your subject on Nano Banana 2 using the JSON formula. Frame 2: make one change to lighting or environment only. Frame 3: switch to ChatGPT-Image 2.0 and use the Preserve chain method to maintain the subject from Frame 1. Compare drift across all three frames.
THE SUCCESS CONDITION: All three frames show the same subject with consistent face structure, clothing details, and lighting direction. The Nano Banana 2 frames hold environmental consistency. The ChatGPT-Image 2.0 frame holds character consistency. No unexplained element changes across frames.
THE FAILURE SIGNAL: Character details changed between frames without instruction. Face structure, clothing color, or body proportions shifted unexpectedly. Fix: re-anchor reference images, add explicit preserve lists, and verify JSON fields were not modified between turns.
INCOME TEST: If the output meets the success condition, take this action today: screenshot it at 2x, add your WenceStudio watermark, and send it to one potential client with the subject line: "This is what AI visual direction looks like." That is your first income signal test from this issue.