To choose an AI tool, start with the file you need to deliver: its use, format and the defects that would make it unusable. Then give candidates the same brief and examine every output against criteria fixed before the trial.
Here is our first episode: a PDF cover without embedded text. The comparative trial is incomplete: two images were obtained from one engine; the remaining requests were blocked by the account quota. These results support an examination of a deliverable, not a claim about the best tool.
Contents
Trial 01: a cover for a specific guide
The selected resource is the free mini-guide Clarifying an offer page (in French), already available on VIFLY. The trial concerns only a candidate illustration for a future cover. The images below have not been added to the existing PDF and do not represent a redesign delivered to a client.
The protocol was fixed on 6 November 2026 at 00 h 01, Reunion time, equivalent to 5 November at 20 h 01 UTC. Interface used: ElevenLabs Creative Studio. Two models were targeted: GPT Image 2.5 Sunburst and FLUX.2 [Pro], using the identifiers offered by the service at the time. These are two engines accessed through one interface, not a ranking of standalone applications.
| Candidate | Initial request | Observed result | What can be assessed |
|---|---|---|---|
| GPT Image 2.5 Sunburst | 4 outputs, a single brief | 2 completed outputs, 2 starts refused | Images A1 and A2 only |
| FLUX.2 [Pro] | 4 outputs, the same brief | Request blocked by insufficient quota; no image received | No visual quality can be assessed |
No retries or retouching were carried out to obtain a better image. The block concerns the account used for this trial: it demonstrates neither general model unavailability nor a quality limitation of FLUX.2 [Pro]. Missing outputs remain missing.
The archived brief, translated here from French
Create an original image for the cover of a PDF guide titled “Clarifying an offer page”. Horizontal 16:9 format. Realistic editorial photograph of an independent professional's desk: an open beige notebook, a pen, a closed navy laptop and a ceramic mug. Soft morning light, a navy and beige palette with a small cyan accent. Place the objects in the right half and leave the left half simple and low in detail so that a title can be added later. No text, numbers, logos, people or hands, no fictional interface, no futuristic rendering. A credible, sharp composition that remains readable as a thumbnail, without promising a commercial result.
The complete original text and its hash are available in the archived initial protocol (in French). The five criteria below were defined before generation. Settings actually reported for A1 and A2 were: 16:9 aspect ratio, resolution labelled “1K”, quality “high”. The received files measure 1280 × 720 pixels. No shared seed is documented.
Both received outputs, without favourable selection


These previews show generated results, not screenshots of the supplier's interface. Original PNG files are retained. WebP previews reduce loading weight; their content and framing are unchanged. The visual assessment below is an editorial judgement, separate from measured dimensions.
| Criterion | A1 | A2 |
|---|---|---|
| 16:9 format | Meets criterion: 1280 × 720 pixels | Meets criterion: 1280 × 720 pixels |
| Requested objects | Meets criterion on inspection: open notebook, pen, closed laptop and mug visible | Meets criterion on inspection: all four objects are identifiable |
| Low-detail left half | Partly meets criterion: the notebook extends slightly into this half | Meets criterion on inspection: the main objects stay on the right; shadows occupy part of the background |
| No text, logo, person or hand | Meets criterion on visual inspection; none of these elements identified | Partly meets criterion: a small text-like mark appears on the notebook; its legibility is not established |
| Credibility and palette | Partly meets criterion: the scene is broadly coherent, but no distinct cyan accent was identified | Partly meets criterion: the scene is broadly coherent, but no distinct cyan accent was identified |
The decision: keep a candidate, leave the comparison open
For a title placed on the left, A2 is a candidate to check first because of its clear area. Its small mark still needs examination at actual size. A1 offers another usable composition, with the notebook's encroachment to consider. Neither image satisfies the entire grid without reservations.
The final deliverable has not been assembled: we have not measured the contrast of an added title, exported a new PDF cover or validated print output. The next step would be a separate mock-up followed by a check at the intended size. This trial does not establish the engine's usual speed, quality on other briefs, a cover's commercial effectiveness or the rights attached to a particular use.
The service reports a technical cost for each completed output, but this episode does not compare full costs: not all requests succeeded, and subscription prices, retries and revision time are not equivalent. No reliable production duration is published. Useful technical details, file hashes and statuses are collected in the results record (in French).
A method to reuse for your own deliverable
- Define a file and a use case. For example: a 16:9 cover image with space for a title, rather than “visuals for my brand”.
- Fix the disqualifying criteria. Format, objects, accuracy, clear area, absence of unwanted text. Include “not assessable”.
- Keep the same brief and record the settings actually available. Identical text does not make two engines technically identical.
- Archive all outputs and failures. State the number of trials and permitted retries in advance. A quota block is not a poor visual score.
- Test the file in its final use. Separate raw output, retouching and deliverable validation.
- Decide for this case only. Keep an option, complete the trial or change method if no output meets the criteria.
Use the trial workbook to complete (in French) to prepare the next episode. It can be filled in on the page and printed; your answers are not sent to VIFLY.
A series grounded in the files obtained
Each new episode will keep this format: specific deliverable, date, fixed protocol, settings, all outputs, deviations, decision and limits. An announced tool does not automatically become a new episode. A repeat of this trial must carry a new version and disclose changed conditions; it will not silently replace the results above.
To prepare the work before choosing an engine, see the methods for setting a direction (in French), organising a creation session (in French) and reviewing a trial (in French). The PDF preparation journal documents the journey around the file.
Operational source consulted on 6 November 2026: official ElevenLabs Image & Video documentation, for model selection, variable settings and export. Findings about A1 and A2 come from our files, not a supplier's promise. VIFLY editorial team; AI-assisted editorial review, without independent validation of visual judgements.