Level 6 · In practice

6.2Express troubleshooting

Goal: turn technical mastery into reliable, reproducible and responsible production reflexes.

Estimated time: 14 min · If you only have 5 min: read “The essentials in 30s” + the diagram.

The essentials in 30 seconds

Diagnose: one symptom = likely causes = one targeted action — never relaunch blindly.
Lock in: hierarchical prompt, shared --sref / --cref, frozen version and --stylize, --seed to compare.
Fix locally: inpainting (4.1) and upscale (4.4) rather than regenerating everything; validated image for video (5.2).

Sonia thread — the table taped next to the screen

Sonia keeps this table next to her screen in Algiers. Verdict after one quarter: three rows served her most — the partly ignored prompt, the style drifting from one image to the next, and the character never quite the same. For each one, she ended up applying the sequence (symptom, causes, action) instead of relaunching at random.

Observed symptom, verified causes, actions in order: the reflex holds in three steps.

Diagnose before relaunching

A generation problem is rarely solved at random: each symptom has typical causes and one targeted action. The table below groups the most frequent situations.

Symptom	Likely causes	Actions
The result ignores part of my prompt	prompt too long/contradictory; elements drowned out; incompatible requests	shorten, prioritize (essentials first), remove contradictions; weight if needed; split into several images
Deformed hands, fingers or texts	classic model weakness on hands and typography	reframe to avoid/hide; regenerate (variations); localized inpainting (4.1); add the text on top in an editor
"Too AI" look, smooth and artificial	high --stylize; no imperfections; generic prompt	lower --stylize; add grain/imperfections/slight flaws; specify medium and light (N3)
The style changes from one image to the next	no shared reference; parameters not frozen	set a style reference (--sref); lock version/--stylize; charter (5.1)
The character is not recognizable from one scene to the next	insufficient description; no character reference	lock the description (3.5); use --cref from a validated image; vary only setting/pose
Low-resolution image / details smearing when enlarged	no upscale; unsuitable upscale	upscale (4.4, Topaz/Magnific depending on the case); regenerate at better quality rather than enlarging a weak file
The generated video "melts" or flickers	temporal coherence; movement too broad; text-to-video	start from a validated image (image-to-video, 5.2); simple slow movements; short sequences built into an edit
Inconsistent results despite the same prompt	normal probabilistic behaviour	set --seed to compare parameters; generate several times and select
Doubt about the commercial usage right of a visual	service terms; protected elements (brands, people, works)	check the terms (0.4); remove protected elements; log prompts/versions; do not present AI as reality

Common misinterpretation: "the table replaces diagnosis". No. The table gives likely causes, not a verdict: observe the symptom, verify the cause (prompt, references, parameters, terms), then apply the targeted action.

Text version of the diagram (SEO / screen readers / translation): Table method: 1) Name the observed symptom (ignored prompt, hands/texts, too AI, drifting style, unstable character, low resolution, melting video, inconsistency, commercial doubt). 2) Keep the table's likely causes without concluding alone. 3) Apply the actions in order (prioritize, inpainting 4.1, lower --stylize, --sref, --cref, upscale 4.4, image-to-video 5.2, --seed, terms 0.4). 4) Lock the reflex at the root (charter 5.1, selection over volume). Lesson: a recurring defect calls for a setting adjustment, not random relaunches.

Ready-to-say script — diagnosing without relaunching at random (30 s)

"My symptom: [ignored prompt / hands / too AI / style / character / resolution / video / inconsistency / rights]. I check the table's likely cause, I apply the targeted action (prioritize, --sref, --cref, inpainting, upscale, --seed, terms) and I check the new result."

Copy-paste template — framed regeneration after a miss

Miss: [observed symptom].
Likely cause: [prompt / reference / parameter / terms].
New attempt: [subject] + [medium + light N3] --sref [code] --cref [image] --stylize [value] --seed [value] --ar [format].
Planned local fix: [inpainting 4.1 / reframing / text in editor].

Small business / tight budget

No formal table: one symptom, the likely cause, the action and who validates, noted down.

Algiers SME

Build local realities into the diagnosis: shared files, connection for upscale, known contact person.

Remote / hybrid

Add written traces to the actions: at a distance, an unlogged defect (seed, version) comes back.

Cross-cultural

Spell out observations in writing: what seems obvious (which defect, which style) is not obvious to everyone.

Caution: Generic reminder: when in doubt about commercial use, check the terms (0.4), remove protected brands, people and works, log prompts and versions, and do not present an AI image as documentation of a reality (0.3); this guide gives no legal advice.

48h challenge — apply tomorrow

Spot 1 current symptom (ignored prompt, drifting style, unstable character), apply the table row (symptom named + cause verified + targeted action).

Section exit checklist

I can link a symptom to its likely causes.
I lock references and parameters before relaunching.
I fix locally (inpainting, upscale) instead of regenerating everything.

Self-assessment 1-5: diagnosis before relaunch?parameters and references locked?

← Previous : Guided practical cases (V3)Next : Best practices & pitfalls (V3) →

Check your understanding

Sonia's style changes from one image to the next. Which reflex does the table prescribe?

💡

Reflex: diagnose the symptom, identify the cause, apply the targeted action — do not relaunch blindly.