IDM’s Text Scores Depend on the Backbone, Not Just the Step Count

Inclusion AI’s IDM reports 94.25% on LongText-CN and 90.20% on LongText-EN at 2 network evaluations, according to the paper. Those scores belong to Z-Image-Turbo 6B. A separate evaluation uses Qwen-Image 20B, and moving the text scores between those backbones would create a claim the published evidence does not support.

The method replaces an external feature encoder with generator EMA features. That is a specific change to the distillation setup, not proof that every resulting image service becomes cheaper or faster. For buyers interested in text inside generated images, the immediate task is to preserve the connection between method, model and benchmark before deciding whether the reported result applies to their workload.

Backbone identity is part of the result

A benchmark result is not portable simply because the same method appears in the experiment. The LongText percentages are reported on Z-Image-Turbo 6B, while AI Weekly’s detailed report puts GenEval 91 and FID 6.34 on Qwen-Image 20B at 2 NFE. These measurements describe different outcomes and model configurations, even though they appear in the same research story.

A comparison sheet should keep the backbone name beside every score. If a product brief puts the best text result next to a different model’s general image metrics without qualification, it implies a combined configuration that has not been established here. A team could choose hardware or implementation work around the wrong model because the evidence was collapsed too early.

The reported parameter counts also need their proper role. The names identify a 6B backbone and a separate 20B backbone, but the source material does not provide an operating-cost comparison between them. Parameter count alone cannot establish memory use, throughput or the delivered price of an application request. Keep those deployment questions open until the intended implementation is measured.

Two evaluations describe procedure, not elapsed time

The 2-NFE setting counts network evaluations. It is not a claim of 2 seconds per image, nor does it describe everything between submitting a prompt and receiving an accepted asset. A buyer needs a clear measurement boundary before treating a compact generation procedure as a latency commitment. The published evidence does not provide that end-to-end commitment.

IDM’s use of generator EMA features instead of an external feature encoder changes how the method obtains features. The supported claim is the replacement itself. The source material does not quantify a universal reduction in training expense, maintenance work or production resource use, so those implications should remain evaluation questions rather than reported outcomes.

The same restraint applies to headline language about short-step generation. Preserve the evaluated 2-NFE condition when discussing the supplied scores instead of drifting into a different step-count claim. Builders comparing methods should be able to reconstruct which backbone and condition produced a result without inferring them from surrounding promotional language. That makes the evaluation useful even when the implementation changes.

Text benchmarks do not approve the final asset

The reported LongText results address a relevant problem for applications that need readable text, but they do not establish perfect reproduction of every requested string. A buyer still needs to inspect examples that match the intended language and layout. That assessment should distinguish a character error from an otherwise attractive image, because those outcomes can have different consequences for publication.

Do not treat the difference between LongText-CN and LongText-EN as a universal ranking of language capability either. The cited reporting give benchmark scores, not a complete account of all language-specific workloads. The responsible interpretation is narrower: the paper reports those percentages on named tests using the named backbone. Applications with other content requirements need their own evidence.

The cost of an unsuitable asset extends beyond generation, although no numerical rework rate is supplied here. A workflow may require review or another attempt when visible text is wrong. Include that possibility in an acceptance test without inventing a projected saving. A method can have an efficient evaluation count and still fail a buyer’s actual requirement if its output cannot be used as delivered.

What To Do About It

• Preserve the model label: Put the backbone, benchmark and evaluation condition together in every comparison. Prevent a combined headline from implying that Qwen-Image produced LongText results reported for Z-Image-Turbo 6B.
• Test required text: Build a review set from the strings and layouts your application must publish. Record visible text errors separately from visual preference, then compare the relevant language result without generalizing beyond LongText-EN at 90.20%.
• Measure accepted output: Time the application request through review and any required retry instead of substituting evaluation count for latency. Keep elapsed time in its own column and label the research setting explicitly as 2 NFE.
• Inspect the method boundary: Review which feature source the implementation uses before claiming the external encoder has disappeared. Check the configuration against the paper’s stated replacement using generator EMA features and the separate Qwen-Image 20B evaluation.

The Bottom Line

IDM offers a concrete change in feature sourcing and reports useful short-evaluation results. The 94.25% LongText-CN figure belongs to Z-Image-Turbo 6B, not whichever backbone produces a convenient accompanying headline. Buyers should verify text requirements on the intended model and measure complete request behavior separately from NFE. A simpler method is valuable only if the resulting assets satisfy the job.