AI Image Troubleshooting — Fix Common Failures Fast
Diagnose ForgeEcho text-to-image failures—wrong colors, plastic skin, shape drift, garbled labels, bad crops—and fix them with one-lever prompt changes, references, and model/resolution swaps.
Diagnose before you re-roll
Blind re-rolls are the most expensive habit in AI image work. Each 2K generation is 4 credits; six confused retries are 24 credits with no learning. This playbook forces a symptom → likely cause → one lever path so you spend Chat (0.5) before Image (3–8).
Works with Nano Banana 2 family models in AI Image. Pair with Optimize Then Generate for the full Plan–Evaluate–Improve loop.
60-second triage
Answer in order:
- Did you upload a reference when identity/SKU shape matters?
- Is there exactly one primary objective in the prompt (lighting or background or retouch—not all)?
- Are there conflicting style terms (e.g. photorealistic + flat illustration)?
- Was the prompt composed for this aspect ratio?
- Are you judging at thumbnail size, not 100% zoom only?
If any answer is “no,” fix that before another generation.
Symptom table (start here)
| Symptom | Likely cause | One-lever fix |
|---|---|---|
| Product shape changed | Weak/no reference; “new angle” language | Upload reference; add “geometry unchanged from reference”; remove angle invention |
| Label / logo garbled | Model inventing text | “simplify label area” or overlay type in design tool; sharp reference + “keep label readable” |
| Wrong product color | Missing color lock | “color accurate to physical product [hex or plain name]”; avoid “vibrant boost” |
| Plastic / waxy skin | Beauty words stacked | Split passes; prefer “natural, preserve pores”; see Photo Retouch |
| Face not the same person | Text-only generation | Reference + “same person as reference, preserve facial features” |
| Busy background fights subject | No hierarchy | “subject remains hero, background uncluttered / soft gradient” |
| Cropped product on 9:16 | Landscape mental model | Rewrite safe zones for vertical; product in lower half if captions on top |
| Soft / mushy detail | Too aggressive denoise words; low res for use | Promote winner to 2K; avoid “smooth everything” |
| Random extra fingers / props | Under-constrained scene | “no hands, no extra products, single unit only” |
| Style too weak vs reference | Under-weighted style adjectives | Boost palette + lighting one step; do not rewrite subject block |
| Style destroys likeness | Over-weighted style | Reduce style words; strengthen preservation block |
| Inconsistent batch look | Prompt drift between SKUs | Freeze brand blocks; field-swap only—Brand Visual System |
Conflict scanner (run in AI Chat)
Paste your prompt and ask:
List conflicting or redundant instructions in this image prompt.
Suggest a rewritten version with: Subject / Light / Background / Constraints / Aspect.
Keep my product name and channel.Typical conflicts Chat should catch:
- photorealistic and illustration / anime
- pure white background and rich lifestyle props
- extreme beauty filter and identity preserve
- macro texture and ultra-wide environment
Reference upload rules
| Situation | Reference? | Notes |
|---|---|---|
| Amazon main / SKU-accurate pack | Required | JPG/PNG/WebP ≤3MB; sharp label |
| LinkedIn headshot | Required | Even light, face ≥50% of frame |
| Mood exploration / new campaign KV | Optional | Text-only OK for drafts |
| Style transfer | Required | Source for composition/likeness |
Bad references (motion blur, tiny product, heavy filters) propagate failure. Re-shoot or crop tighter before blaming the model.
Resolution & model levers (only after structure is OK)
| If… | Then… |
|---|---|
| Composition wrong | Stay at nano-banana-fast @ 1K; do not “fix” with 4K |
| Composition OK, detail soft | nano-banana-2 @ 2K once |
| Platform zoom / print | 4K on the single winner |
| Reference edits keep drifting | Try nano-banana-2-cl / stronger preserve lines; still one lever per run |
Credit reminder: 1K=3, 2K=4, 4K=8—see Credit Budget.
Worked examples
Example 1 — Melty Amazon label
Fail prompt (compressed): “beautify product, cinematic, neon, white background, new angle…”
Fix:
- Remove cinematic/neon/new angle
- Upload pack reference
- Prompt: white seamless BG + soft contact shadow + “preserve label text and logo, geometry unchanged”
- 1K × 2 → 2K × 1
Example 2 — Plastic LinkedIn headshot
Fail: “flawless porcelain skin, beauty filter, HDR”
Fix: two-pass method from Photo Retouch—lighting pass first, subtle skin second; keep “pores visible, same person.”
Example 3 — 9:16 cover crops the bottle
Fail: landscape product prompt forced into 9:16
Fix: rewrite—“vertical 9:16, product in lower third, top third low-texture for captions, bottle fully in frame.”
Evaluate log (print or Notion)
| Field | Entry |
|---|---|
| Date / SKU / channel | |
| Symptom | |
| Hypothesis | |
| Lever changed (one) | |
| Model + resolution | |
| Pass? | Y/N |
| Next action |
Three rows of this log teach juniors faster than twenty unexplained PNGs in a Drive folder.
When it is not a prompt problem
| Signal | Action |
|---|---|
| Same prompt suddenly worse across all jobs | Check model dropdown; note provider variance; retry once at 1K |
| Output empty / job error | Retry; if persistent, contact support with time + model |
| Legal / brand logo inventing competitors | Stop; overlay real logos in design tools only |
FAQ
Should I use negative prompts?
Prefer positive constraints (“single unit, white background, no props”). If you add avoid-lists, keep them short: no watermark, no extra fingers, no unreadble text.
How many retries before escalating?
After two one-lever retries at 1K with no improvement, go back to Chat rewrite or a better reference—not a sixth 2K.
Voice issues?
Different playbook—pacing and script shape live in AI Voice.