Open weights vs ChatGPT image generation

Retrieved 2026-10-18. When to use GPT-Image-2.5, when to use open weights, and how to add image generation to our hosted stack. Every figure comes from open-vs-closed.md.

Open reached parity once, in Oct 2025; the gap is back to 202 Elo

arena.ai text-to-image, closed #1 vs open #1, last publish of each month. The shaded band is the gap. Months with no publish are skipped.

  • Closed #1
  • Open #1 (any licence)
  • Gap
Data table

Source: LMArena history dataset (huggingface.co/datasets/lmarena-ai/leaderboard-dataset), cross-checked against 17 Wayback captures; open-vs-closed.md §3.1. Retrieved 2026-10-18.

202 Elo

GPT-Image-2.5 wins about 76% of blind votes against the best open model.

Sunburst 1425 vs Qwen-Image-2.1 1223 on arena.ai text-to-image, board of 2026-10-07. Against Flare, ChatGPT's likely model, the gap is 175 (73%).

Executive review

Decide

Pick by the job, not by the leaderboard.

If you need this, use this, for this reason
If you need Use Why
The best single image, or a text-heavy or reasoning prompt GPT-Image-2.5: Flare by default, Sunburst for precision Top row on every board; leads by 73.6 points on GenExam (OpenAI judge) and 21.5 on UltraText (independent judge)
High volume Hosted open: FLUX.2 dev, Z-Image-Turbo, Qwen-Image-2512 $0.004 to $0.02 per MP vs $0.0527 per image for GPT high
A custom style Open weights plus a LoRA OpenAI offers no image fine-tuning; fal runs LoRA trainers for FLUX.2 klein, Qwen-Image-2512, Krea 2 and Ideogram 4
Private or self-hosted generation Self-host an open model Runs on your hardware, about $0.0004 to $0.0076 per image by arithmetic
To ship commercially on open weights Cosmos3-Super (OpenMDW), Qwen-Image-2512 (Apache-2.0), HiDream-O1 or Ming-Image (MIT) Not Qwen-Image-2.1 (research only), Ideogram 4 or FLUX dev (non-commercial)
Editing GPT-Image-2.5; open fallback Qwen-Image-Edit (Apache-2.0) or HunyuanImage 3.0 Instruct Sunburst leads 155 Elo on arena edit and 33.6 points on RISEBench++
Image generation in our hosted stack A tool call to a separately served SGLang Diffusion model No training, and it is OpenAI's own API design (see Built-in generation)

Decisions only you can make

  1. Does research-only count as open for us? If not, every gap widens by 43 to 127 Elo, and the open #1 becomes Cosmos3-Super, 260 behind.
  2. Which model and box serves our image endpoint? Qwen-Image-2512 (Apache-2.0, about 54 GB bf16 by arithmetic) or Ming-Image (MIT, fits one 80 GiB GPU), on a separate box, never co-located on DI9.
  3. Fund the hidden-state graft research? Recommendation: not now. Its gain over letting the LLM write the prompt is unproven; revisit after option 1 ships.

Scoreboard

Four boards, one answer: GPT-Image-2.5 Sunburst wins two thirds to five sixths of head-to-heads.

GPT wins 66% to 76% against the best open model, and 77% to 84% against the best commercial one

Elo-implied share of head-to-heads won by the GPT row, per board. 50% is parity.

  • vs best open, any licence
  • vs best commercially usable open
  • Parity (50%)
Data table

Source: arena.ai text-to-image and image edit boards, Artificial Analysis T2I and editing v2.1; open-vs-closed.md §2. Retrieved 2026-10-18.

Gap = GPT row minus open row. CI = √(ci₁² + ci₂²). GPT wins = 1 / (1 + 10^(−gap/400)). Against Flare, the any-licence gaps are smaller: arena T2I 175 (73%), arena edit 112 (66%), AA T2I 165.5 (72%), AA edit 98.5 (64%).

Where open sits on the ChatGPT timeline

Open has passed ChatGPT as it was in spring 2025, by about 107 Elo, but not ChatGPT as it was in late 2025.

Today's best open model sits just below GPT-Image-1.5, a model from Dec 2025

arena.ai text-to-image Elo, one axis. Dashed guides mark each ChatGPT model.

  • ChatGPT (OpenAI) models
  • Open, any licence
  • Open, commercially usable
  • Reference (Google)
Data table

Source: arena.ai text-to-image; open-vs-closed.md §1 to §3. Retrieved 2026-10-18.

Trend

Verdict: closed through 2025, parity in Oct 2025, re-opened in 2026. The Elo gap is below its May peak, but the lag is at a record. Open did not stall: open #1 rose 979 to 1223; the closed frontier rose faster, 1105 to 1425, mostly in two OpenAI steps.

The lag shrank through 2025, then stopped: nothing closed since 2025-11-20 has been matched

Months from each closed model's release until an open model matched it on LMArena, in release order.

  • Matched by an open model
  • Still unmatched (board of 2026-10-07)
Data table

Source: LMArena history dataset; release dates from Artificial Analysis, cross-checked against Hugging Face createdAt; open-vs-closed.md §3.3. Retrieved 2026-10-18.

Artificial Analysis re-scaled between the 2026-09-01 and 2026-10-01 captures, so compare within a row only. It shows the same shape: 66 to 89, then 76 (Dec 2025), about 100, then 147 to 178.

Capabilities

Open is at parity only on saturated legacy boards. On current boards the gap is widest in reasoning, dense text and edits.

Verdict per capability, from the capability matrix (§4). Gap = GPT minus open; negative means open leads.
Capability Verdict Key evidence (best GPT vs best open, benchmark, tier)
Simple text rendering Close LongText −0.02, CVTG-2K −0.013, OneIG Text −0.14 / −0.34: open self-reports lead, but on saturated boards (T0 / S) against a stale GPT Image 1 row
Dense text Far UltraText: GPT Image 2 [Low] 99.35 vs Boogu-Image-0.1-Base 77.89, +21.5 (T3-I); Qwen-Image-2512 English falls 86.5 to 42.9 from level 1 to level 3
Prompt adherence, legacy boards Parity GenEval: GPT Image 2 0.89 vs HiDream-O1 0.90; DPG-Bench 85.98 vs 89.83 (T0 / S)
Long-prompt adherence Close UniGenBench++ long: GPT Image 1.5 95.41 vs FLUX.2-dev 90.31, +5.1 (T3-I)
Precise instructions and spatial frames Far FoR-T2I: GPT-Image-2 68.2 vs Qwen-Image-2512 33.9, +34.3; UniGenBench++ short +11.0, negation 88.1 vs 64.6 (T3-I)
World knowledge and reasoning Far GenExam: Sunburst 78.9 vs Qwen-Image-2.1 5.3 strict, +73.6; WorldCast +52.7 (both OpenAI judges); T2I-CoReBench +13.8 (T3-I)
Knowledge with chain-of-thought Parity WISE_Verified: SenseNova-U1 with CoT 0.834 vs GPT-Image-1.5 0.825, open leads by 0.009 (T2)
Editing, weak judges Parity GEdit-EN: GPT Image 2 7.67 vs HiDream-O1 7.60; ImgEdit 4.73 vs 4.64 (T0 / S). Re-judged by Gemini 3.1 Pro, GEdit opens a 1.13 gap
Editing, reasoning and human votes Far RISEBench++: Sunburst 56.6 vs HunyuanImage-3.0-Instruct 23.0, +33.6; SolveEdit +34.2; arena edit +155 and AA edit +116 Elo (T2)
Holistic, LLM-judged Close Qwen-Image-Bench: Sunburst 67.01 vs Qwen-Image-2.1 60.28, +6.7 (T3-S, read from a chart); independent board +12.6
Crowd preference Far 116 to 202 Elo behind on four boards (T3 by rule, T2 for the editing arenas)

GPT leads on every current points benchmark but one; the lead is largest where reasoning is scored

GPT lead in benchmark points (GPT minus best open), only for benchmarks scored on a 0 to 100 scale. Right of zero: GPT leads. Left of zero: open leads.

  • GPT leads
  • Open leads
Data table

Source: the benchmark rows of open-vs-closed.md §4.1 to §4.5; "OpenAI judge" marks boards where an OpenAI model scores the images. Retrieved 2026-10-18.

Cost and control

Open is cheaper than GPT at high and max quality, faster on every measured host, and the only option for fine-tuning or running offline.

Most open routes cost less than GPT high; only GPT's low tier overlaps them

Dollars per image on a log scale. Open prices are per megapixel and GPT's per image; a 1024² image is about 1.05 MP, so the two are roughly comparable. Lines show a price range across hosts or settings.

  • GPT-Image-2.5 quality tiers
  • Hosted open models
  • Self-host on a rented H100
Data table

Source: OpenAI image-guide cost calculator and pricing page, fal, Replicate, Together, docs.bfl.ai, Artificial Analysis; open-vs-closed.md §5.1. The self-host row is arithmetic, not a measured run. Retrieved 2026-10-18.

Sunburst takes about two minutes an image; FLUX.2 dev takes seconds

Median generation time at 1024², seconds.

  • GPT image models
  • FLUX.2 [dev], open
Data table

Source: Artificial Analysis median generation time; open-vs-closed.md §5.2. Retrieved 2026-10-18.

Control: what each side lets you do
Axis GPT-Image-2.5 Open weights
Fine-tuning Not offered
v1/fine-tuning "Not supported" 		LoRA and full
fal trainers; day-0 LoRA for FLUX 3 Klein Base
Content filter Always on
moderation auto or low only 		Yours to set
licence bans are contractual, not technical
Privacy, on-prem Hosted only Your hardware
Licence Usage policies
bans include likeness without consent and impersonation 		Varies
top two are non-commercial
Ecosystem No adapters
no fine-tuning, so nothing to share 		Large
FLUX.1-dev 42,711 adapters; Qwen-Image-2.1 136 in 5 weeks
Price model Tokens
output $30/M, Batch $15/M 		Per MP or GPU-second

Licences

The two best open models cannot be used commercially.

The best commercially usable open row is 260 Elo below Sunburst

The open rows on arena.ai text-to-image that open-vs-closed.md reports, by Elo. Orange ticks mark two GPT models for reference.

  • Commercial use allowed
  • Restricted (research, non-commercial or community licence)
  • GPT reference
Data table

Source: arena.ai text-to-image (board of 2026-10-07) and licence terms on each Hugging Face card; open-vs-closed.md §2, §5.4 and §6.2. Retrieved 2026-10-18.

Built-in image generation

How an LLM can make images itself, who ships that in the open, and what we should build.

A normal multimodal LLM reads images but writes only text.

A unified model has an image output path inside the same network; there are four families, below.

OpenAI's API exposes image generation as a tool: the mainline model decides when to call the separately named GPT Image models. Neither OpenAI nor Google discloses the exact architecture.

No open unified model is close to GPT Image: the best on independent boards, HunyuanImage 3.0, is about 275 below Sunburst on arena text-to-image.

The four families of native image generation, with two example models each
Family Shape How it works Examples
(a) Autoregressive image tokens text LLM … The image becomes discrete codes predicted one at a time, like words Emu3, Janus-Pro
(b) One transformer, text tokens plus diffusion text one transformer text patches Next-token prediction for text, diffusion or flow matching for image patches, in one network BAGEL, Transfusion
(c) LLM, connector, diffusion decoder LLM connector diffusion Learnable queries, a connector or an MLP map the LLM's hidden states into the diffusion model's conditioning MetaQuery, Ming-Image-0.1-Design
(d) Discrete diffusion ? ? ? ? Text and image tokens start masked and are un-masked in parallel Lumina-DiMOO, MMaDA
Open unified roster, top rows (QIB = Qwen-Image-Bench overall; "self" = the model's own paper)
Model Date Params Licence Family Shared-board quality
HunyuanImage-3.0 / -Instruct 2025-09-25 83.0B MoE, 13B active tencent-hunyuan-community b (AR + MoE) arena T2I 1150 (#35); arena edit 1303 (Instruct); AA 932, Instruct 988; QIB-ZH 50.81
Ming-Image-0.1-Design 2026-09-17 6.15B MIT c, sold as T2I only AA 999, the best MLLM-bridged model on AA
HiDream-O1-Image 2026-05-08 8.8B MIT pixel-unified generator, not a chat model arena 1116; AA 974
SenseNova-U1.5-8B-MoT 2026-08-19 17.5B Apache-2.0 b (encoder- and VAE-free) QIB-EN 52.82, 60.22 with prompt enhancement (self); GenEval 0.92
LLaDA-Image 2026-08-28 6.54B DiT Apache-2.0 c + d QIB EN 53.53 / ZH 53.38 (self); GenEval 0.85
S1-Omni-Image 2026-06-12 54.0B Apache-2.0 c (S1-VL-32B to Qwen-Image) GenExam 53.2 vs Qwen-Image 23.8 vs GPT-Image-2 93.8
BAGEL-7B-MoT 2025-05-19 14.7B (7B active) Apache-2.0 b (MoT) arena T2I 898 (last, #82); arena edit 1027 ±6; AA 586 (last)

Our stack: DeepSeek writes the prompt and calls a diffusion server on a separate box

Option 1. DI9 serves DeepSeek-V4.1-Flash on 8×B200 at TP8 under SGLang, already run with tool calling. The diffusion model runs on one or two GPUs elsewhere; about 47 GiB per GPU is left on DI9 after graph capture, and that is prefill workspace, not spare memory.

Source: open-vs-closed.md §6.3; SGLang Diffusion supported-model list (docs.sglang.io, read 2026-10-18).

Ways to add image generation to DI9, ranked
Rank Option What it needs Expected quality Verdict
1 Tool call to a separately served diffusion model: DeepSeek calls an image_generation tool that hits SGLang Diffusion serving Qwen-Image-2512 (or another model the licence allows) on /v1/images/generations No training. One or two GPUs on a separate box: Qwen-Image-2512 about 54 GB bf16 by arithmetic; Ming-Image fits one 80 GiB GPU That of the served model: arena 1125 (Qwen-Image-2512) to 1223 (Qwen-Image-2.1, research licence) Do this
2 Tool call to an open unified model: SenseNova-U1.5-8B-MoT (Apache-2.0) for generate and edit; HunyuanImage-3.0-Instruct for editing No training. SenseNova's SGLang cookbook defaults to 1 GPU; HunyuanImage needs TP 2 to 8 on H200 SenseNova QIB-EN 52.82, 60.22 with prompt enhancement (self-reported) A/B it
3 S1-Omni-style graft: train an MLP from DeepSeek's 5120-dim hidden states into a diffusion model's encoder space About a 44.57M-param MLP on about 40K prompts, then MLP + DiT co-tuning on 294K samples; untested engine path on DI9 S1 lifted Qwen-Image GenExam 23.8 to 53.2, still below GPT-Image-2's 93.8 Research only
4 Learned-query connector (MetaQuery, BLIP3-o, Bifrost-1) 25 to 55M image and caption pairs; trains the connector, the DiT or both GenEval about 0.80 to 0.84, behind current open diffusion models Worse
5 "Extract" the generator from BAGEL, Janus or HunyuanImage, or retrain DeepSeek as unified Not an extraction: the generator is the shared weights. Retraining is pretraining scale (BAGEL 2,665M samples) n/a Not viable

Which benchmarks to trust

Human arenas first, current automated suites second, saturated legacy suites not at all.

Three tiers of evidence
Tier Boards How to read them
Human arenas arena.ai text-to-image and image edit; Artificial Analysis T2I and editing Crowd preference, not capability, and T3 by rule (the editing arenas are T2). They are the only boards that rank both classes on one fit with CIs.
Current automated suites UltraText, UniGenBench++, FoR-T2I, T2I-CoReBench, GenExam, WorldCast, RISEBench++, SolveEdit, WISE_Verified, Qwen-Image-Bench Independent static sets (T3-I) or refreshed sets (T2). Where the gap shows: reasoning, dense text, spatial frames and edits.
Saturated legacy suites GenEval, DPG, OneIG, LongText, CVTG, and GEdit or ImgEdit under weak GPT-4.1 or Qwen-VL judges T0: top models are within noise. Every "open beats GPT" headline lives here.

Two traps

Caveats and sources

How this could be wrong

  • Crowd Elo is preference, not capability. Sunburst, Flare, Qwen-Image-2.1 and Nano Banana 2.1 are still preliminary on arena.ai.
  • The licence choice moves the answer: excluding research and non-commercial weights widens every gap by 43 to 127 Elo.
  • Artificial Analysis re-scaled between the 2026-09-01 and 2026-10-01 captures (GPT Image 2 went 1370 to 1172); only gaps inside one capture compare.
  • LMArena's scale stability is inferred from flat legacy rows, not stated by LMArena.
  • The open and closed split is partly our labelling; prompt rewriters (Cosmos3 Agentic, qwen-image-prompt-extend) blur rows.
  • Which API model ChatGPT runs is inferred (most likely Flare); against Flare every gap is 12.5 to 43 Elo smaller.
  • Self-host cost is arithmetic, not a measured run.

Re-check around 2026-11-18

  • FLUX 3 [klein] (FLUX Kommunity licence) is on neither board yet; BFL claims +91 Elo over klein 9B internally.
  • HunyuanImage 3.5 is API-only (AA 970 for the Preview); open weights would re-enter the race.
  • Qwen-Image-2.1's preliminary scores settling, and whether any host's commercial deal makes it usable.
  • Regenerate the gap series from the LMArena dataset (one parquet download) rather than Wayback.

Sources