# ANOMALY Knowledge ยท The Image Prompting Bible How to write image prompts that come out right the first time. Covers **Nano Banana Pro** (the workhorse) and **GPT Image 2** (the character maker). This is craft knowledge, use it every single time you write an image prompt. --- ## Which model, when | Model | Use it for | Why | |---|---|---| | **GPT Image 2** | The master character image. Clean close-ups. Logos. | Best at a believable single human face and clean text/shapes. | | **Nano Banana Pro** | Everything else: hook frames, build clips, end-result clips, recreating a competitor's frame. | Best at editing an existing frame and holding a reference character across many shots. | Rule of thumb: **GPT Image 2 creates the character once. Nano Banana Pro reuses that character forever.** Nano Banana Pro is the model id `nano_banana_2` on Higgsfield. Do not confuse it with `nano_banana_flash`, which is the weaker one. --- ## The 7-part prompt architecture (GPT Image 2) Write every prompt in this order. The order matters, the model weights the front of the prompt more heavily. 1. **Scene** what is happening, the action and context 2. **Subject** who and what: character description + the product, with exact details 3. **Details** materials, textures, colors, how finished the product is 4. **Camera** angle, shot type, handheld or locked 5. **Lighting** source, quality, time of day 6. **Style** the aesthetic and mood 7. **Negatives** what to exclude ### The build-clip formula (fill in the brackets) ``` [Character name], [age + physical description], in [setting with specific details]. [Action being performed], [the exact step being shown]. Shot on iPhone, handheld, slightly shaky, natural [light source] light. Real [workspace type] in background with [specific background elements]. [Outfit description, casual, real, matches character]. [Warm or cool] color temperature. Organic UGC style. No studio lighting. ``` ### The end-result formula ``` [Product], [specific variant or color], [exact visual description]. Held by [character description] or displayed on [surface]. [Lighting that shows the product's wow factor]. Shot on iPhone, close-up, natural light. Product fills [X]% of the frame. Background: [soft blur description]. ``` --- ## The KEEP / SWAP framework (Nano Banana Pro), the most important thing in this file When you recreate a competitor's frame, do **not** describe the whole scene from scratch. Tell the model what to keep and what to change. Being surgical is what makes it work. ``` Edit reference image 1 ([one-line scene description]). KEEP EXACTLY: [room, architecture, floor, furniture, background people and decor, composition, camera angle, and the lighting stated precisely, for example "cold fluorescent overhead, roughly 5500K"] SWAP ONLY THESE TWO THINGS: (1) Person, replace with the character from reference image 2. [Character DNA, see below]. [Outfit appropriate to the scene]. Same position, same pose, same body scale. (2) [Product or materials], replace [the original items] with [your item plus a short description, referencing your product image]. LIGHTING: preserve the exact lighting from reference 1. CAMERA: 9:16, same framing as reference 1. AVOID: wrong face, original items still visible, background changed, warped hands, watermark, on-screen text. ``` **Why it works:** the model only has to solve the small thing you asked for. Ask it to redo everything and it drifts. --- ## Character DNA (paste it into every single prompt) Write your character once as a fixed block of words and never reword it. Rewording is what causes the face to drift between clips. ``` [age range] [gender] [ethnicity], [hair, be very specific about style and how it sits], [skin tone + a real texture detail like visible pores], [face shape], [one signature accessory, for example small stud earrings] ``` Rules: - Same words, every prompt. Copy and paste, do not paraphrase. - Always pass the locked character image as a reference alongside the text. - Never write "attractive," "beautiful," or "model." Write "real, slightly imperfect, authentic." Model-perfect faces kill the organic feel and stop conversion. - Lock the outfit per scene type so clothes do not change mid-build. --- ## Product DNA Same idea for the product. Write one fixed description and reuse it. ``` [Product name], [material and texture], [color], [hardware or detail], [the one feature that makes it recognizable] ``` --- ## The realism code (how to not look AI) AI-obvious images do not convert. Add these to essentially every prompt: **Always include:** shot on iPhone, handheld, natural light, organic content, real person. **Always exclude:** no studio lighting, no perfect skin, no CGI, no ring light, no posed model, no watermark, no on-screen text. **The "make it look real" block** (works when an image looks too clean): ``` Make it look like a casual iPhone photo shot in auto mode. Natural uneven lighting, slight motion blur, soft focus, mild grain, compressed dynamic range. Keep skin with pores, minor blemishes, realistic texture. Avoid perfect symmetry or studio lighting. Handheld framing with a small tilt and imperfect composition. Subtle noise, slight over or underexposure, natural color balance. Moderate depth of field, not cinematic blur. No text or overlays. Realistic shadows, reflections, and correct anatomy. ``` --- ## Composition rules that actually drive views 1. **Leave negative space for your text hook.** Decide where the caption sits (usually center or upper-center) and compose the shot so the character and product sit *outside* that zone. Compose around the caption, not the other way around. 2. **First clip is a tight macro of the action.** Not a wide establishing shot. Close on the hands, the tool, the material. That is what stops the scroll. 3. **Something is always happening.** Cutting, pouring, rolling, inflating, stitching. Never a dead static frame. 4. **Bright, saturated, high contrast.** Eye-catching colors beat tasteful muted ones on a feed, every time. 5. **Show volume in the background.** Finished pieces stacked behind the character read as "this is a real operation." --- ## Settings - **Aspect ratio:** 9:16, and state it in the prompt text *and* the parameter. - **Resolution:** 1K for drafts, **2K for anything you will post**. - **Generate multiples.** 2 to 4 per prompt, then pick the best. Outputs vary a lot; one generation is a coin flip. - **Prompt length on Higgsfield:** keep it plain text and **under about 1100 characters**. Long prompts and JSON-structured prompts fail on that endpoint. - **References:** upload once, reuse the returned id. Do not re-upload the same character every time. --- ## Hard-learned rules (these cost real money to learn) 1. **Dark scenes get brightened against your will.** The model pulls dim scenes toward the well-lit character reference. Fix: write "do NOT brighten this scene, preserve the dim moody lighting exactly." 2. **Too many changes at once causes drift.** If a scene is hard, do the person swap first, then the product swap as a second pass on the output. 3. **Fewer references equals a better scene lock.** Do not pass product references into a scene where the product does not appear. 4. **Always run each prompt twice.** Outputs vary; picking the best of two is free quality. 5. **Hands and text are the usual failure points.** Put "warped hands" and "watermark, on-screen text" in your avoid list every time. --- ## When it keeps getting it wrong (the fix that always works) Do not just re-roll. Ask your Claude: > "Check the original prompt. Is there any reason it keeps doing [the specific problem]?" The error is almost always ambiguous language in the prompt, not bad luck in the generation. Claude re-reads its own prompt, finds the phrase pulling the wrong output, and fixes it at the source. This works for face drift, background drift, wrong product size, lighting drift, and text errors.