---
type: "post"
id: "habr-compares-four-image-generators-on-russian-prompts-5573c0"
slug: "habr-compares-four-image-generators-on-russian-prompts-5573c0"
title: "Habr compares four image generators on Russian prompts"
description: "**Thesis:** Small aggregate score gaps can conceal task-specific failure modes, so prompt and evaluation design matter more than a single winner label."
retrieval_nugget: "**Thesis:** Small aggregate score gaps can conceal task-specific failure modes, so prompt and evaluation design matter more than a single winner label."
published_at: "2026-10-14"
updated_at: "2026-10-14"
record_date: "2026-10-14"
date_kind: "published_at"
topics: []
source_urls: ["https://habr.com/ru/companies/syntx_ai/articles/1093216?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:habr-compares-four-image-generators-on-russian-prompts-5573c0"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:ca613219ce334bf1d82cf0fad64ec5f1f0843ac5c7a1918cbcf7babc91e1b806","reviewed_at":"2026-10-14T16:11:11.618Z","source_evidence_count":1,"verified_claim_count":3,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":3,"implication_count":2,"watch_condition_count":2,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"Small aggregate score gaps can conceal task-specific failure modes, so prompt and evaluation design matter more than a single winner label.","observed_facts":[{"text":"The article reports 14 Russian-prompt tasks and 76 binary checklist points across four image generators.","source_urls":["https://habr.com/ru/companies/syntx_ai/articles/1093216?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr"]},{"text":"Reported scores range from 67 to 72 points; the authors used one generation per task and one evaluator.","source_urls":["https://habr.com/ru/companies/syntx_ai/articles/1093216?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr"]},{"text":"The article says Nano Banana Pro failed markedly on an 18-element table and Nano Banana 2.1 had capitalization and punctuation errors in poster and infographic tasks.","source_urls":["https://habr.com/ru/companies/syntx_ai/articles/1093216?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr"]}],"mechanism":"Aggregate binary checklists collapse heterogeneous capabilities—exact text, spatial arrangement, counting, and image editing—into one total, allowing task-specific weaknesses to disappear inside a close overall score.","why_now":"The comparison tests a newly released model against its predecessor, Pro, and GPT Image 2.5 on Russian-language tasks, exposing a language-and-task slice not covered by the cited external benchmark described in the article.","implications":["Select evaluation cases that mirror the intended production task rather than rely on a leaderboard total.","Require repeated and independently scored runs before treating small score differences as meaningful."],"evidence_boundary":"One generation per task, one evaluator, no blind comparison, and no generation-time measurement; the reported results are not a general ranking.","watch_conditions":["Whether repeated, independently scored runs preserve the same task-level error patterns.","Whether additional samples and blinded assessment change the relative results across exact text, layout, and editing."],"related_records":[],"new_branch_reason":"No prior New Runtime record was supplied; this is a distinct comparison of Russian-prompt image-generation failure modes and evaluation design."}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/habr-compares-four-image-generators-on-russian-prompts-5573c0/","markdown":"https://newruntime.com/posts/habr-compares-four-image-generators-on-russian-prompts-5573c0.md","json":"https://newruntime.com/posts/habr-compares-four-image-generators-on-russian-prompts-5573c0.json"}
---

# Habr compares four image generators on Russian prompts

## Retrieval answer

**Thesis:** Small aggregate score gaps can conceal task-specific failure modes, so prompt and evaluation design matter more than a single winner label.

**Thesis:** Small aggregate score gaps can conceal task-specific failure modes, so prompt and evaluation design matter more than a single winner label. The article reports four systems scoring between 67 and 72 of 76 checklist points across 14 Russian-prompt tasks. It also reports a pronounced failure by Nano Banana Pro on an 18-element table, while Nano Banana 2.1 lost capitalization and punctuation in poster and infographic tasks.

**Mechanism:** Aggregate binary checklists collapse heterogeneous capabilities—exact text, spatial arrangement, counting, and image editing—into one total, allowing task-specific weaknesses to disappear inside a close overall score. **Why now:** The comparison tests a newly released model against its predecessor, Pro, and GPT Image 2.5 on Russian-language tasks, exposing a language-and-task slice not covered by the cited external benchmark described in the article.

**First implication:** Select evaluation cases that mirror the intended production task rather than rely on a leaderboard total. **Evidence boundary:** One generation per task, one evaluator, no blind comparison, and no generation-time measurement; the reported results are not a general ranking. **First watch condition:** Whether repeated, independently scored runs preserve the same task-level error patterns. The article provides prompts and checklists for attempted replication, but this account does not independently validate the outputs.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)