Skip to content

Quality Levers

How to invest more time and cost for better quality, and — equally important — when to accept "good enough" instead of perfecting. Quality is a budget, not a free dial.

The quality-vs-cost trade

flowchart LR
    A[Output quality] --> B{Acceptable?}
    B -->|yes| C[Ship]
    B -->|close but not great| D{Worth investing more?}
    D -->|yes, high-value workflow| E[Invest: prompt-tuning,<br/>more attempts, more iteration]
    D -->|no, diminishing returns| C

The pipeline can almost always produce slightly better output if you invest more. The question is whether the marginal quality improvement is worth the marginal cost.

Lever 1: Iteration count during testing

The default rhythm: iterate until 3 of 4 candidates per scene are acceptable, then graduate.

You can push higher:

Quality bar Iterations needed (typical)
3 of 4 acceptable (default) 1-3 testing iterations
4 of 4 acceptable 3-5 testing iterations
4 of 4 perfectly matching reference 5-8 testing iterations
Perfect across all variants 8-15 testing iterations (rare)

When to push higher:

  • High-value workflow — top performer, repeated use, big audience
  • Compliance-critical — small visual mistakes are costly
  • Client demands it — the brand expects polish

When NOT to push higher:

  • Experimental workflow — you're testing whether the format works at all
  • Time-constrained — budget or deadline matters
  • Single-use workflow — not going to be reused, doesn't need to be perfect

Lever 2: Per-scene targeting

Workflows have a bottleneck scene — the one that won't land. Two approaches:

Universal investment
Bump iteration count, attempts, prompt-tuning rounds for the whole workflow.
Targeted investment
Identify the bottleneck. Invest specifically in that scene. Leave the rest at default.

Targeted is almost always better. A workflow with 7 clean scenes and 1 stubborn one doesn't need universal high-quality treatment.

You: Scene 04 keeps producing bad hands. Bump attempts to 10 for just
     that node. Other scenes are fine — leave them at default.

Claude: [targeted regen with attempts=10 on Scene 04 only]

Lever 3: Choose what to accept

Quality is a function of what you'll accept. Different workflows have different acceptable thresholds:

Content type Acceptable threshold
Internal test / experiment 2 of 4 acceptable; even one good candidate works
Standard production workflow 3 of 4 acceptable; standard polish
Premium / featured workflow 4 of 4 acceptable; high polish
Client-facing premium 4 of 4 + manually reviewed by a second human

Be explicit with yourself about the threshold before generation. Without a threshold, you'll either:

  • Accept too easily because you're tired of iterating
  • Iterate forever because you're chasing perfection

Set the bar, hit it, ship.

Lever 4: Composition over perfection

Pick the composition that works rather than the most perfect rendering. Often the scene with the right energy and pose beats the scene with the most polished face — even if the polished one is anatomically cleaner.

A candidate with good composition + slight facial asymmetry > a candidate with perfect anatomy + dead-eyes posing.

The viewer notices composition (energy, lean, expression) first. They notice anatomy (fingers, eye shape) only when zooming in.

Lever 5: When to invest in prompt-tuning

The prompt-tuning skill costs 4-8 image gens per scene tuned. That's significant.

Invest in prompt-tuning when:

  • The scene is in a high-value workflow
  • Normal regen attempts haven't converged
  • The issue is specific (lighting / wardrobe / pose) — something the skill can actually fix
  • The scene matters disproportionately (hook scene, product reveal scene)

Don't invest when:

  • The scene is OK but not great — accept the OK
  • The workflow is exploratory — don't polish what you're not committed to
  • The issue is structural (wrong reference image, impossible composition) — prompt-tuning can't fix it

Lever 6: Accept hallucinated text

NanoBanana 2 routinely produces garbled text in backgrounds — signage, price tags, product labels. Ignore it.

Spending iterations trying to fix hallucinated text:

  • Won't reliably work (the model is fundamentally bad at text)
  • Burns cost
  • Distracts from things that actually matter

Things that actually matter (in priority order):

  1. Face identity — is the avatar's face right
  2. Composition — is the framing / angle correct
  3. Wardrobe — is the clothing right
  4. Pose — is the body positioning natural
  5. Setting — is the environment correct

If those five are working, ship. Garbled background text is a non-issue at scroll speed.

Lever 7: Compare to source, not to ideal

For video copies, compare your output to the source video, not to a Platonic ideal:

  • If the source has slightly cool lighting, yours can too
  • If the source's framing is imperfect, yours should match
  • If the source's pacing is unhurried, don't force yours to be tight

Trying to make a video copy "better than the original" is usually a mistake. The original worked — match it, don't surpass it.

Lever 8: Trust the auto-QA

The Generation Runner's visual QA pass catches the worst AI tells (extra fingers, melted faces, garbled limbs). It reruns them automatically up to 3 times.

You don't have to inspect every image with fresh eyes — the auto-QA has already done a pass. Spot-check, but trust the system.

Time saved: ~30-40% on manual inspection per workflow.

Lever 9: Iterate the prompt, not the model

When generation isn't producing what you want, the instinct is sometimes "try a different model." That's almost never the answer.

The model isn't the issue. The prompt is.

  • Reframe the prompt to be more specific
  • Add reference frames if the issue is composition
  • Use the prompt-tuning skill for stubborn cases
  • Split the scene if it's too ambitious in one clip

Switching models (e.g., NanoBanana → Seedance) is a fallback when prompts have been exhaustively tried and the model genuinely isn't capable. Rare.

When you're ready

You've finished Chapter 12. You can now:

  • Run a workflow automatically with the Generation Runner agent
  • Optimize for cost (mode=images, B-roll density, cache discipline)
  • Understand the pipeline's hard limits (concurrency, Veo length, file size)
  • Juggle multiple workflows without dropping context
  • Make explicit speed vs. quality trade-offs

→ Next: Chapter 14 — Troubleshooting. The final chapter — what to do when something breaks. Reference material, not necessarily linear reading.