Understand the technology
Turn technical knowledge into a supported decision.
For May and November 2027. This lesson is higher level only. SL students are not required to study this hybrid-model content. HL students evaluate hybrid approaches while needing only broad understanding of VAEs and flow-based models.
01 | EVALUATE MEANS REACH A SUPPORTED JUDGEMENT
Visionary Studios is considering generative AI for advertising, concept art and digital media. A hybrid could combine complementary model capabilities, but the company needs evidence that the result suits its work.
To evaluate , weigh strengths and limitations against relevant criteria and reach a justified judgement. Describing model components is necessary background, but it does not complete the evaluation.
This HL lesson uses a fictional pilot to make the decision concrete. Its architecture, targets and results are teaching examples, not findings reported in the official booklet.
02 | DEFINE THE USE BEFORE THE MODEL
Our classroom pilot concerns illustrated campaign scenes featuring the recurring explorer introduced in earlier lessons. The studio needs recognisable identity, coherent style, usable layouts and a manageable production process.
Concept exploration may tolerate some variation and manual refinement. Final campaign assets have tighter approval requirements.
Keep these uses separate. A tool can be useful for drafts without being approved for final production, and success on this campaign would not establish suitability for every studio project.
03 | STATE A PROPOSED HYBRID CLEARLY
The proposed pilot combines a compatible image representation model with a latent diffusion component. The representation model supplies an encoder and decoder; diffusion generates representations that the decoder converts into images.
Reference-based guidance could be included if the selected system supports it. That guidance is a capability to test, not an automatic consequence of being a hybrid.
We are not proposing that all four model families must appear. A flow or adversarial component needs its own defined purpose and evidence if added.
04 | TRACE THE ACTUAL PRODUCTION PATH
Brief, approved references and supported guidance.
Sample latent noise, run compatible diffusion sampling and decode.
Check requirements, correct selected outputs and obtain approval.
Training the representation and diffusion components is a separate development activity. Using pretrained components changes the studio’s preparation costs but does not remove generation or review work.
05 | CHOOSE A MEANINGFUL BASELINE
A baseline is a reference workflow for comparison. It could be the studio’s current process, a suitable standalone generative system or a simpler configuration.
Use alternatives that can actually perform the task. A deliberately unsuitable baseline would make the hybrid look stronger without proving practical value.
If the comparison changes both model structure and guidance capabilities, state that limitation. A whole-system comparison can support a purchasing decision without isolating the causal contribution of each component.
06 | AGREE CRITERIA AND ESSENTIAL GATES
Before the trial, define quality, consistency, control, time, resources, reliability and operational requirements. Also investigate data provenance and disclosure arrangements.
Some conditions are essential gates: for example, a campaign image must retain the approved character features and accommodate the required layout.
A high average aesthetic score should not cancel an essential failure. Thresholds must be clearly proposed and agreed rather than invented after seeing a preferred system’s results.
07 | TEST REPRESENTATION FIDELITY FIRST
Encode and decode approved examples before evaluating the complete generator. Inspect whether the representation path preserves required features.
If a badge becomes unclear even during reconstruction, that is evidence of a representation limitation. Do not automatically blame the diffusion sampler for every defect in the finished output.
Record both component-level findings and complete-system results. Good reconstruction alone does not prove that newly generated campaign scenes will meet the brief.
08 | TEST QUALITY AND USEFUL VARIATION
Use several representative scenes and repeated candidates. Review defects, subject relationships, style and layout with the same criteria for each alternative.
Check whether the explorer remains recognisable across new poses and settings. A set of almost identical portraits can conceal poor creative flexibility.
Where feasible, hide system labels during review and calibrate reviewers using examples. Explain disagreements rather than pretending artistic judgement has no uncertainty.
09 | TEST COMPUTATION AND DEPLOYMENT
Measure relevant processing stages and the complete user workflow. Record latency, throughput, memory limits, queue delays and failures under realistic demand.
For a hosted system, investigate usage limits, network dependence and data handling. For local execution, investigate hardware capacity, maintenance and staff expertise.
The booklet gives no prices or company hardware specification. Obtain those details before making a real deployment commitment rather than assume a universal low-cost option.
10 | READ A FICTIONAL PILOT TABLE
| Measure | Simpler baseline | Proposed hybrid |
|---|---|---|
| Candidates | 40 | 40 |
| Candidates suitable for correction and completion | 20 | 30 |
| Total generation time | 400 seconds | 320 seconds |
| Correction time per selected candidate | 30 seconds | 20 seconds |
| Additional preparation and maintenance | Not measured | Not measured |
The selected candidates are not automatically final approved assets. They still need the stated correction and final checks. The table deliberately leaves important costs unknown.
11 | CALCULATE WHAT THE TABLE SUPPORTS
The baseline selects 20/40 = 50% of candidates; the hybrid selects 30/40 = 75% . That is a 25 percentage-point increase in this trial.
Generation time per selected candidate is 400/20 = 20 seconds for the baseline and 320/30 ≈ 10.67 seconds for the hybrid.
Including the stated correction gives 50 seconds and approximately 30.67 seconds respectively. These are allocated trial averages within that scope, not guaranteed service times or complete business costs.
12 | RECOGNISE WHAT THE TABLE CANNOT SHOW
The results do not measure reference preparation, maintenance, final approval or training costs. They do not establish performance on other styles, projects or peak demand.
Forty candidates and one pilot can also be sensitive to sampling and reviewer judgements. Repeat trials and inspect whether the advantage holds across different briefs and runs.
A numerical improvement is evidence to consider, not permission to ignore missing information. A defensible conclusion states both the finding and its limits.
13 | COMPARE TOTAL COST AT DIFFERENT VOLUMES
An additional setup cost can be worthwhile at high use but difficult to justify for a one-off project. Distinguish one-time preparation from recurring work.
For an invented cost example, suppose the hybrid requires 120 additional staff minutes to set up and saves 2 staff minutes per completed asset. The simple break-even count is 120/2 = 60 assets , ignoring other differences.
Do not mix machine seconds and staff minutes as if they were automatically interchangeable. A full cost comparison needs appropriate monetary or time accounting and stated assumptions.
14 | USE COMPONENT TESTS AND VALID ABLATIONS
A whole-system pilot tells the studio whether a workflow is useful. Component tests and ablations help investigate why.
Compare valid alternatives that remove or change the relevant mechanism. If adversarial training is being tested, simply removing the discriminator after training does not erase its effect on learned parameters.
Some changes require retraining or a different interface. Record changed capacity and compute so the explanation does not claim a perfectly isolated cause without evidence.
15 | CHECK RESPONSIBILITIES AND OPERATIONAL RISKS
Investigate permissions and provenance for training data and references, possible representation bias, disclosure expectations and handling of confidential client work.
A reversible component does not make a whole lossy pipeline transparent. A compressed representation is not automatically anonymous.
Also consider dependency failures and version compatibility. An update to one checkpoint can alter the representations or behaviour expected by another component.
16 | WRITE A CONDITIONAL RECOMMENDATION
A useful recommendation identifies a use, the supporting evidence, the trade-offs and conditions for proceeding.
This is stronger than “hybrids are best” and more decisive than listing benefits and drawbacks without reaching a conclusion.
17 | PLAN A LIMITED ROLLOUT AND REVIEW
If the evidence is promising, use a limited rollout with agreed monitoring and a fallback workflow. Keep human approval for final assets and preserve version records.
Define reasons to pause or reconsider, such as repeated identity failures, unacceptable waiting times or unresolved data concerns.
Compare actual production results with pilot expectations. A trial is useful because it supports an informed next step, not because it guarantees future success.
18 | BUILD YOUR OWN EVALUATION
In your book, recommend whether to expand, limit or reject the fictional pilot. Use one quality or selection finding, one timing calculation and one significant missing cost.
Then state what evidence would change your recommendation. Keep case-study facts separate from the invented trial.
Reveal a possible response
Continue a limited trial rather than commit to full adoption. Selection rose from 50% to 75%, and allocated generation-plus-correction time fell from 50 to about 30.67 seconds per selected candidate. However, setup and maintenance were not measured, and final approval still needs review.
Test repeated projects and realistic demand, measure all relevant costs and retain the simpler workflow as a fallback. If missing costs or critical detail losses remove the advantage, prefer the simpler approach.
HL only. Read alongside the hybrid challenges and evaluation factors in the official 2027 booklet. All pilot data, thresholds and rollout proposals here are classroom illustrations.
CHECK YOUR UNDERSTANDING | BUILD AN EVIDENCE-BASED EVALUATION