coulomb/intelligence-radar
5
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
intelligence-radar/digests/topics/2026-10-08-1527_Image_Generation_with_Generative_AI_Digest.md
tegwick 9c4aaa4761 Add topic digest: image generation with generative AI, model-choice trees. 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 197486@bnt-lap001
Assistant-Session: 8e588643-5640-4ad7-9309-4ed3e12b2423
2026-10-08 16:33:32 +02:00

21 KiB
Raw Blame History

title date project document_type status evidence_observed_at policy_refs tags
Image Generation with Generative AI — State of the Field and Model Choice 2026-10-08T15:27:00+02:00 intelligence-radar topic-digest initial-assessment 2026-10-08
model-ranking-v0.1
intelligence-radar
image-generation
text-to-image
image-editing
capability-advisor
openrouter
model-choice

Image Generation with Generative AI — State of the Field and Model Choice

Timestamp: 2026-10-08 15:27 CEST
Research context: intelligence-radar / Capability Advisor
Evidence snapshot: Artificial Analysis Text-to-Image and Image Editing arenas (read 2026-10-08), Arena text-to-image leaderboard (last updated 2026-10-07), OpenRouter /api/v1/models?output_modalities=image (61 entries, read 2026-10-08).

Prices and rankings in this field move monthly. Five of the top six text-to-image models were released after June 2026. Treat every number here as dated to 2026-10-08. Re-check before you commit to spending money (P7).

1. Executive summary

  • Frontier (general quality): OpenAI GPT Image 2.5 Sunburst (Sept 2026). It is #1 on both independent human-preference arenas and on the editing arena. GPT Image 2.5 Flare is statistically tied with it (rank range 1–2 on Artificial Analysis). The cost is high: about $0.21 per image at max quality.
  • Efficient (Topnotch at the lowest cost): Google Nano Banana 2.1 (≈ $0.034/img). It is close to the Frontier at about one sixth of the price, but it is only days old and its arena scores are preliminary. The well-evidenced fallback is Microsoft MAI-Image-2.6 (≈ $0.039/img).
  • Budget (credibly Sufficient for routine work): Meta Muse Image (≈ $0.010/img). It is #8 overall on Artificial Analysis and #5–8 for editing, at about one twentieth of the Frontier price. This is the stand-out value observation in this digest.
  • Open weights: Alibaba Qwen-Image-2.1 leads (Elo 1035, under a non-commercial Qwen Research licence). The open/closed gap is about 160 Elo, so the best open model is roughly as good as the closed frontier was about a year ago.
  • OpenRouter: every model above except the open-weights ones is available, so the main recommendations carry over unchanged. Missing from OpenRouter: Ideogram, Midjourney, Reve, Luma, HiDream, Imagen 4, and all self-hostable weights.
  • Structural shift: the top of the field is now natively multimodal LLM-integrated generators (GPT Image, Nano Banana/Gemini, MAI, Muse, Grok Imagine). Pure diffusion specialists (FLUX, Ideogram, Recraft, Midjourney) have moved into niches: vectors, typography, style systems, self-hosting, and very low cost.

2. The landscape (2026-10-08)

2.1 Text-to-image — Artificial Analysis Arena (AA-Image-T2I v2.0, top 20)

The anchor is FLUX.2 [dev] = 1000. Price is per 1,000 images at 1024², default settings, from the creator's API. "OR" marks models available on OpenRouter.

# (range) Model Creator Elo $/1k Released OR
1 (1–2) GPT Image 2.5 Sunburst (max) OpenAI 1198 210.7 2026-09 ✔
2 (1–2) GPT Image 2.5 Flare (max) OpenAI 1191 210.7 2026-09 ✔
3 GPT Image 2 (high) OpenAI 1172 211.0 2026-04 ✔
4 (4–6) Nano Banana 2.1 Google 1160 33.6 2026-10 ✔
5 (4–6) Grok Imagine Image 2.0 SpaceXAI 1156 60.0 2026-08 ✔
6 (4–6) MAI-Image-2.6 Microsoft AI 1151 38.9 2026-08 ✔
7 Nano Banana 2 (Gemini 3.1 Flash Image) Google 1126 67.0 2026-02 ✔
8 (8–9) Muse Image Meta 1115 10.0 2026-07 ✔
9 (8–13) GPT Image 1.5 (high) OpenAI 1107 133.0 2025-12 —
10 (9–13) MAI-Image-2.6-Flash Microsoft AI 1105 19.5 2026-09 ✔
11 MAI-Image-2.5 Microsoft AI 1104 48.1 2026-06 ✔
12 Nano Banana Pro (Gemini 3 Pro Image) Google 1102 134.0 2025-11 ✔
13 MAI-Image-2.5-Pro Microsoft AI 1100 108.5 2026-07 ✔
14 Nano Banana 2 Lite Google 1095 33.6 2026-06 ✔
15 Qwen-Image-3.0-Pro Alibaba 1088 40.0 2026-07 ✔
16 Seedream 5.0 Pro ByteDance Seed 1081 90.0 2026-07 ✔
17 Qwen-Image-3.0 Alibaba 1075 30.0 2026-07 ✔
18 grok-imagine-image-quality SpaceXAI 1046 50.0 2026-05 ✔
19 (19–24) Qwen-Image-2.1 (open weights) Alibaba 1035 no API 2026-09 —
20 MAI-Image-2.5-Flash Microsoft AI 1032 20.0 2026-06 —

Selected further reference points:

Model Elo $/1k Note
FLUX.2 [max] / [pro] / [dev] 1020 / 1004 / 1000 70 / 30 / 12 dev = open weights, arena anchor
Ideogram 4.5 (High) / 4.0 (Quality) 1016 / 1012 100 / 100 4.0 has open weights; typography specialist
Recraft V4.1 Utility / V4.1 / V4.1 Flash 1022 / 989 / 957 35 / 35 / 7 vector and brand-style specialist
FLUX.2 [dev] Flash / Turbo (Fal) 987 / 998 5 / 8 cheapest strong open-weights serving
Z-Image Turbo 941 5 open weights, very cheap
Midjourney v7 Alpha 879 no API no programmatic access; not in contention for API work
DALL·E 3 777 40 legacy; do not choose

2.2 Independent cross-check — Arena (arena.ai) text-to-image, 2026-10-07

# Model Score Votes
1 gpt-image-2.5-sunburst 1425 ± 7 (prelim.) 17,752
2 gpt-image-2.5-flare 1398 ± 7 (prelim.) 16,889
3 gpt-image-2 (medium) 1383 ± 4 93,334
4 grok-imagine-image-2.0 (canvas) 1334 ± 12 (prelim.) 3,113
5 mai-image-2.6 1332 ± 5 25,091
6 gemini-nano-banana-2.1 1328 ± 9 (prelim.) 5,310
7 reve-2.1 1302 ± 8 8,249
9 muse-image 1274 ± 5 42,262
12 seedream-5.0-pro 1255 ± 4 94,181

Agreement. Both arenas give the same top-6 set, and the same ordering of vendors at the very top. Disagreement. On Arena the OpenAI lead is about 95 points, against about 40 on Artificial Analysis. That makes the "near Frontier" status of Nano Banana 2.1 and MAI-Image-2.6 less certain than the AA numbers alone suggest. Arena also shows that GPT Image 2 at medium quality still ranks #3. Quality tier is therefore a large cost lever within the OpenAI family, but neither leaderboard publishes a per-tier price at that setting.

2.3 Image editing — Artificial Analysis Editing Arena (top tier)

Rank Model Elo $/1k OR
1 GPT Image 2.5 Sunburst (max) 1183 210.7 ✔
2 GPT Image 2.5 Flare (max) 1163 210.7 ✔
3–4 MAI-Image-2.6 1139 38.9 ✔
3–4 Nano Banana 2.1 1137 33.6 ✔
5–7 MAI-Image-2.6-Flash 1127 19.5 ✔
5–7 GPT Image 2 (high) 1125 211.0 ✔
5–8 Muse Image 1120 10.0 ✔
15–17 HunyuanImage 3.5 (Preview) 1087 24.0 ✔
18–23 Qwen-Image-2.1 (open) 1074 no API —
19–24 HunyuanImage 3.0 Instruct (open) 1068 no API —

Editing is where the cheap, LLM-integrated models are closest to the Frontier. MAI-Image-2.6-Flash and Muse Image are statistically tied with GPT Image 2 (high) for editing, at 1/10 and 1/20 of its price.

3. Benchmark selection (P4) for "general image generation"

# Instrument Weight Construct Freshness Known weaknesses
1 Artificial Analysis T2I Arena 0.35 Pairwise human preference, fixed prompt set 2026-10-08 Aesthetic bias; prompt set not task-specific; new models have few appearances
2 Arena (arena.ai) text-to-image 0.25 Pairwise preference, user-supplied prompts 2026-10-07 Self-selected voters and prompts; correlated with #1 (P5), so it is not an independent fifth of the evidence
3 Artificial Analysis Editing Arena 0.20 Preference on instructed edits 2026-10-08 Primary instrument for editing tasks, secondary for generation
4 OneIG-Bench (or a successor) 0.10 Automated alignment, text rendering, reasoning, style and diversity not refreshed Automated judges; lags new releases
5 LongText-Bench / text-rendering evals 0.10 Text rendering accuracy (Latin + CJK) not refreshed Narrow construct; vendor-reported numbers dominate

Mapping quality: medium. Instruments #1–#3 are fresh and well populated. Instruments #4 and #5 were not re-fetched for this digest. The specialist branches of the decision tree (typography, vector output, photorealism, brand style) therefore rest mostly on vendor positioning and press coverage, not on measured category scores. task-router/BENCHMARK-ATLAS.md has no image-generation family yet. That is an evidence gap to close (see §7).

4. Applying the ranking policy (model-ranking-v0.1)

The policy defines Topnotch as task_score ≥ 0.90 × task_score(frontier). Elo is an interval scale, so a ratio of Elo values is meaningless. This digest therefore uses a digest-local operationalisation, which is not a versioned policy:

task_score = expected win rate against the arena anchor (FLUX.2 [dev], Elo 1000), p = 1 / (1 + 10^(−(Elo − 1000)/400)).

Frontier (Sunburst, 1198) → p = 0.758. The Topnotch threshold of 0.90 × 0.758 = 0.683 corresponds to Elo ≥ ≈1133 on the AA board.

Model AA Elo p vs anchor ratio to Frontier Head-to-head vs Frontier $/img Class
GPT Image 2.5 Sunburst 1198 0.758 1.00 — 0.211 Frontier
GPT Image 2.5 Flare 1191 0.750 0.99 0.49 0.211 Topnotch
GPT Image 2 (high) 1172 0.729 0.96 0.46 0.211 Topnotch
Nano Banana 2.1 1160 0.715 0.94 0.45 0.034 Topnotch → Efficient
Grok Imagine Image 2.0 1156 0.711 0.94 0.44 0.060 Topnotch
MAI-Image-2.6 1151 0.705 0.93 0.43 0.039 Topnotch (Efficient fallback)
Nano Banana 2 1126 0.674 0.89 0.40 0.067 Sufficient
Muse Image 1115 0.660 0.87 0.38 0.010 Sufficient → Budget
MAI-Image-2.6-Flash 1105 0.647 0.85 0.37 0.0195 Sufficient

Sufficient needs a success threshold per task. No outcome ledger exists for image tasks yet, so this digest uses a provisional placeholder: for routine illustration and marketing imagery, a model is Sufficient at Elo ≳ 1100. That is the level of GPT Image 1.5 and Nano Banana Pro, which were the Frontier about ten months ago. Under that placeholder, Budget = Muse Image. For drafts, thumbnails and exploration (the "exploratory" 0.60 threshold), much cheaper models qualify, as §5 shows.

The anchor choice affects class membership near the threshold. If Arena's gap is used instead of AA's, Nano Banana 2.1 and MAI-Image-2.6 fall just outside Topnotch. Confidence in the Efficient pick: medium. Confidence in the Frontier and Budget picks: high.

5. Decision tree — general (any access route)

START: what does the task need?
│
├─ A. Hard constraint: weights must run on own hardware
│     (data residency, offline, fine-tuning/LoRA, no per-image fees)
│   ├─ Best quality, non-commercial use OK ........ Qwen-Image-2.1           (Elo 1035; Qwen Research licence)
│   ├─ Commercial use, quality first .............. Ideogram 4.0 (open)      (1012) or FLUX.2 [dev] (1000) — check each licence
│   ├─ Editing / instruction-following ............ Qwen-Image-2.1 (1074 edit) or HunyuanImage 3.0 Instruct (1068)
│   └─ Throughput / small GPU ..................... FLUX.2 [klein] 9B/4B, Z-Image Turbo (≈940 / 863)
│
├─ B. Output must be vector (SVG, logos, icons, scalable assets)
│   └─ Recraft V4.1 Vector (budget) / V4.1 Pro Vector (quality)       [unmeasured by arenas → evidence weak]
│
├─ C. Editing an existing image (inpaint, restyle, multi-reference, product shots)
│   ├─ Highest fidelity, cost no object ........... GPT Image 2.5 Sunburst   (1183; up to 16 reference images)
│   ├─ Near-frontier, about 6× cheaper ............ MAI-Image-2.6 (1139) ≈ Nano Banana 2.1 (1137)
│   └─ High volume ................................ Muse Image (1120, $0.010) or MAI-Image-2.6-Flash (1127, $0.020)
│
├─ D. Text-heavy images (posters, infographics, UI mockups, CJK text)
│   ├─ Default .................................... GPT Image 2.5 Sunburst   (GPT Image line is strongest on text rendering)
│   ├─ Factual infographic needing current data ... Nano Banana 2.1 / 2 with web-search grounding
│   └─ Typography-centric design, lower cost ...... Ideogram 4.5 / 4.0       [specialist; vendor claim, category score not refreshed]
│
├─ E. Brand style systems / consistent illustration style across a set
│   └─ Recraft V4 Styles (Pro), or a top model with a fixed reference-image set
│
└─ F. General generation (photo, illustration, concept art, marketing)
    ├─ Is the per-image cost irrelevant compared with the value of the best result?
    │   └─ YES → FRONTIER: GPT Image 2.5 Sunburst (max)
    │            · Flare if latency matters (tied on quality)
    │            · try GPT Image 2 at medium quality first: still #3 on Arena
    ├─ Do you need near-frontier quality at scale?
    │   └─ YES → EFFICIENT: Nano Banana 2.1 ($0.034)
    │            · fallback while 2.1 is preliminary: MAI-Image-2.6 ($0.039)
    │            · alternative look: Grok Imagine Image 2.0 ($0.060)
    ├─ Is it routine imagery where "good" is good enough?
    │   └─ YES → BUDGET: Muse Image ($0.010)
    └─ Is it drafts, thumbnails, exploration, or bulk A/B variants?
        └─ YES → FLUX.2 [dev] Flash/Turbo ($0.005–0.008), Z-Image Turbo ($0.005),
                 Recraft V4.1 Flash ($0.007)

Escalation pattern (P12). For important assets, decomposition usually beats running one model for everything. Explore with a Budget model at 10–20 variants, pick a composition, then regenerate or edit only the final one or two with the Frontier model. At these prices, 20 × Muse ($0.20) + 2 × Sunburst ($0.42) costs about the same as 3 Frontier attempts, and covers much more of the composition space. This is a hypothesis to test in the outcome ledger, not a measured result.

6. Decision tree — limited to OpenRouter

6.1 How OpenRouter bills image models

OpenRouter lists 61 image-output entries (including two openrouter/auto* routers). It expresses most per-image vendor prices as a per-token rate. For the image-token-priced vendors (Recraft, Krea, Muse, Qwen, Seedream, FLUX 3), the rate × ≈4,175 tokens reproduces Artificial Analysis's per-image price exactly. Examples: Recraft V4.1 → $0.035, Krea 2 Large → $0.060, Muse → $0.010, Qwen-Image-3 → $0.030. Prices below are derived from that rate. For OpenAI and Google the token count depends on the quality and resolution tier. Grok Imagine lists $0.04 output plus $0.01 per input image, against Artificial Analysis's $0.060 per generated image. Verify actual spend with OpenRouter's generation stats after the first runs.

Calls go through the normal chat-completions endpoint, with image output requested through modalities. One OpenRouter key therefore covers every branch below. Data handling then depends on the upstream provider, which matters for branch A.

6.2 Tree

START (OpenRouter only)
│
├─ A. Self-hosting / open weights required
│   └─ NOT SATISFIABLE on OpenRouter. Nearest hosted options:
│        black-forest-labs/flux.2-pro or flux.2-klein-4b (open-family models, hosted by BFL)
│        qwen/qwen-image-3 ($0.030) if the requirement is "non-US vendor", not "own hardware"
│
├─ B. Vector output
│   ├─ recraft/recraft-v4-styles-vector ...................... ≈ $0.050
│   ├─ recraft/recraft-v4.1-vector ........................... ≈ $0.080
│   └─ recraft/recraft-v4.1-pro-vector ....................... ≈ $0.300
│
├─ C. Editing
│   ├─ Frontier ..... openai/gpt-image-2.5-sunburst
│   ├─ Efficient .... microsoft/mai-image-2.6 or google/gemini-nano-banana-2.1
│   └─ Budget ....... meta/muse-image   (or microsoft/mai-image-2.6-flash)
│
├─ D. Text-heavy images
│   ├─ Default ....................... openai/gpt-image-2.5-sunburst
│   ├─ Grounded / current facts ...... google/gemini-nano-banana-2.1 (web_search priced separately)
│   ├─ Layered, editable design ...... inclusionai/ming-image-0.1-design-layer   (Elo ≈997; no price listed — verify)
│   └─ (Ideogram is NOT on OpenRouter; there is no typography specialist)
│
├─ E. Brand / style systems
│   └─ recraft/recraft-v4-styles (≈$0.035) / recraft-v4-styles-pro (≈$0.100)
│
└─ F. General generation
    ├─ FRONTIER ..... openai/gpt-image-2.5-sunburst      (Flare = faster twin; gpt-image-2 at lower quality to save cost)
    ├─ EFFICIENT .... google/gemini-nano-banana-2.1       (fallback: microsoft/mai-image-2.6)
    │                 x-ai/grok-imagine-image-2.0          (alternative look)
    ├─ BUDGET ....... meta/muse-image                     (≈ $0.010, Elo 1115)
    └─ BULK / DRAFT
        ├─ recraft/recraft-v4.1-flash .................. ≈ $0.007  (Elo 957)
        ├─ black-forest-labs/flux.2-klein-4b ........... ≈ $0.014  (Elo 863)
        ├─ bytedance-seed/seedream-5-0-flash ........... ≈ $0.018  (unranked, new 2026-10-01)
        └─ black-forest-labs/flux-3-image .............. ≈ $0.021  (unranked, new 2026-10-01 — watch)

6.3 What OpenRouter costs you

Need Best overall Best on OpenRouter Loss
Frontier / Efficient / Budget Sunburst / Nano Banana 2.1 / Muse same none
Self-hosting, fine-tuning, LoRA Qwen-Image-2.1, FLUX.2 [dev] — total
Typography specialist Ideogram 4.5 GPT Image 2.5 (generalist) small in quality, but no cheap typography option
Ultra-cheap bulk FLUX.2 [dev] Flash / Z-Image Turbo ($0.005) Recraft V4.1 Flash ($0.007) small
Distinctive aesthetic Reve 2.1 (Arena #7), Midjourney — depends on taste; not measurable here

Verdict: for API-driven work that does not need self-hosting, limiting yourself to OpenRouter costs almost nothing at the top of the market in October 2026.

7. Evidence gaps and confidence

Item Confidence Gap
Frontier = GPT Image 2.5 Sunburst high Both arenas agree; 14–18k votes each
Efficient = Nano Banana 2.1 medium-low 5–6k votes, preliminary on Arena; the AA vs Arena gap affects Topnotch membership
Budget = Muse Image high for routine work Strong on both arenas (42k Arena votes); Sufficient threshold is a placeholder
Specialist branches (B, D, E) low–medium No refreshed category or text-rendering scores; vendor claims dominate
OpenRouter per-image costs medium Derived from token rates; OpenAI/Google spend depends on quality tier
Price of GPT Image 2/2.5 at medium/low quality unknown The lever with the most potential savings inside the Frontier family is unmeasured here

Follow-ups worth doing:

  1. Add an image-generation family to task-router/BENCHMARK-ATLAS.md, with AA T2I, Arena T2I, AA Editing, a text-rendering benchmark and a compositional alignment benchmark.
  2. Decide on a versioned rule for applying topnotch.relative_to_frontier_minimum to pairwise-preference (Elo) scores. The anchor-win-rate mapping in §4 is a candidate.
  3. Run a small, matched outcome-ledger trial (P9). Use about 20 real prompts across branches C, D and F with Sunburst, Nano Banana 2.1, MAI-2.6 and Muse, with a human verdict per image. This would test both the Efficient pick and the §5 escalation hypothesis.
  4. Watch list: FLUX 3 Image (unranked), Seedream 5.0 Flash (unranked), Nano Banana 2.1 (vote count stabilising), Ideogram 4.5 (too few appearances, 2.4k).

Sources