whatmodel.sh / image / BAGEL

BAGEL

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

BAGEL is an open-weights image model from ByteDance Seed, released May 2025. Its published size is 14B parameters. Licence: Apache 2.0.

Model information sources

Release date ↗ · Parameter count ↗

Mixture-of-Transformers: 14B total, 7B active (active_b 7 if used). Whole unified model; the card says it also uses a SigLIP encoder and FLUX.1-schnell VAE, not clearly included in the 14B.

42.0±1.1#99 of 144 image modelsfrom 22 benchmark results across 20 sources

Score within the LMArena text-to-image comparison set only.

$ how it compares

The image models ranked either side of BAGEL, out of 197 with a score, then more from ByteDance Seed.

#modelscorewhereRAMreleased
131Lumina-Image 2.0 Alpha-VLLM (Shanghai AI Lab)43±4api $0.075/image +local~13GB2025-01
132BLIP3o-NEXT-GRPO-Text-3B Salesforce Research42±3comparison only
133BAGEL ByteDance Seed42±1api $0.10/image +local~12GB2025-05
134Midjourney V6.1 Midjourney42±4comparison only2024-07
135FLUX.1 [schnell] Black Forest Labs42±2api $0.0030/image +local~7.7GB2024-08
more from ByteDance Seed
12Seedream 5.0 Pro ByteDance Seed75±2api $0.050/image2026-07
29Seedream 5.0 Lite ByteDance Seed65±2api $0.035/image2026-02
45Seedream 4.0 ByteDance Seed61±1api $0.040/image2025-09

see all image models · find the ones that fit your machine

$ what it's good at

text to image▂▅▅▃▄▄█▃▁▂▁▁ #87 of 126 · 61% of its score · 10 tests

Beats Latent Diffusion text2img-large (LDM-KL-8-G, LAION-400M), VQ-Diffusion (ITHQ), FLUX.1 Kontext [dev] here, though they rank higher overall. Trails Playground v2.5, Playground v2 (1024px aesthetic), Show-o2 7B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
LMArena text-to-image897.8 ✎▃▂▃▄▇█▄▂▄▂·▂ beats 0/75avg 1027▇▇▇▇▇▇▇ 17%typical
GenExam (relaxed overall, GPT-5 judge)12.9% ✎▄█▂▂▂▂▆▄·▂·▂ beats 5/16hard · avg 26%▇▇▇▇▇ 12%sharp
T2I-ReasonBench (overall)55.6% ✎▅▂█▂▂▂····▂▂ beats 8/13easy · avg 69%▇▇▇▇ 10%at its level
TIIF-Bench v3 (short prompts, TIIF-Evaluator)71.0% ✎▂█····▄▄▄··▂ beats 1/11easy · avg 79%▇▇▇ 8.1%at its level
ImagenWorld text-guided generation (human overall)76.0% ✎▄··█▄·▄··▄·▄ beats 2/6easy · avg 84%▇▇ 5.8%typical
MMMG knowledge-image generation (MMMG-Score, average over education levels)15.0% ✎▃··▃·█▃▃▅··▃ beats 5/9hard · avg 24%▇ 3.0%typical
WISE (WiScore overall, GPT-4o judge)70.0% ✎▄·▆▂▆█▄▆▂▄·▄ beats 21/22medium · avg 60%▇ 2.4%at its level
XTC-Bench generation (Overall Gen., scene-graph node agreement)72.5% ✎▄█·▄·······▄ beats 4/4easy · avg 70%▇ 2.2%at its level
2 more tests
testresultvs the others in its rundifficultycountswhy
GenEval lab82.0% ✎▇▃▅▇▁▄▅▃█▃▃▅ beats 30/39easy · avg 79%—not counted yet: no other model in the same run to compare it with
WISE overall (model publisher's run) lab70.0% ✎no other model yet—not counted yet: no other model in the same run to compare it with
image to image / editing▁▁·▃▂▃█▃▃▄▁▁ #56 of 62 · 31% of its score · 10 tests

Beats UNO here, though they rank higher overall. Trails OmniGen2 here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Edit-Compass (EN, overall)41.6% ✎▃▃▂·▃▆▂█·▃·▃ beats 10/20hard · avg 45%▇▇▇ 8.6%typical
Arena image edit1026.6 ✎▃▁▁▂▆▇█▄▆▅▁▂ beats 1/53avg 1153▇▇▇ 6.4%typical
UnicBench (EN, overall)69.8% ✎▂····▂▄▂▄·█▂ beats 6/11easy · avg 73%▇▇ 3.8%at its level
GEditBench v2 (overall Elo)851.0 ✎▃▂····▂█▂▃▂▃ beats 0/14avg 956▇ 2.5%blunt
GEBench GUI state generation (English GE-Score)28.0% ✎▄▄····▄▄█▄·█ beats 0/8hard · avg 34%▇ 2.5%at its level
CREval-Bench creative image editing (Overall, GPT-4o QA evaluator)68.8% ✎▄█▄··█··▄·▄▄ beats 2/8easy · avg 74%▇ 2.2%blunt · at its level
MultiTextEdit cross-lingual text-in-image editing (SemAvg, LVM judge)43.6% ✎▄·▄▄▄▄··█▄·▄ beats 3/8medium · avg 50%▇ 1.9%at its level
ImagenWorld text-guided editing (human overall)67.0% ✎█··█··██···█ beats 2/4easy · avg 73%▇ 1.6%blunt · at its level
2 more tests
testresultvs the others in its rundifficultycountswhy
OmniContext average (GPT-4.1 judge)57.3% ✎▃▃▃··▃·····█ beats 3/6medium · avg 65%▇ 1.0%at its level · noisy
GEdit-Bench-EN overall (G_O, GPT-4.1 judge) lab65.2% ✎▄·····█·▄▄·▄ beats 2/5—not counted yet: no other model in the same run to compare it with
other tests5 not in a task · 8.4% of its score
testresultvs the others in its rundifficultycountswhy
RealPref-Bench image editing (human user-study score)71.8% ✎█▄▄····▄···▄ beats 3/5easy · avg 78%▇▇ 4.0%at its level
WEAVEBench in-context multi-turn editing (Avg)44.6% ✎█▄▄·▄▄▄▄█·▄█ beats 2/12medium · avg 65%▇ 2.4%at its level
ViGoR-Bench zero-shot visual reasoning, image models (Result Avg)23.1% ✎▂···▂·▂█▆▅·▂ beats 3/15hard · avg 31%▇ 0.8%typical
PaintBench precise editing (mIoU, 20-task average)5.0% ✎▃·····█▅▃·▃▅ beats 5/9very hard · avg 4.3%▇ 0.7%too hard for it
Knowledge Visualization T2I (Brief Caption, English overall)22.2% ✎▅·▃·▃··▃·█▃▃ beats 2/9hard · avg 28%▇ 0.5%blunt

$ where to run it

Through an API

api · fal.ai · $0.10/image · fal.ai ↗ · price sourceUSD per generated image: fal bills fal-ai/bagel at $0.10 per image (billing unit images); the fal-ai/bagel/edit endpoint is also $0.10 per image. Product: fal.ai serverless text-to-image. API model: fal-ai/bagel (ByteDance-Seed BAGEL, described by fal as a 7B-parameter multimodal model). Settings: text-to-image, one image. Price basis: published per-image list rate. Not a ByteDance direct tariff. Currency: USD. fal.ai official model page, observed 2026-09-26.

Other ways to run it

version

  1. Run it as published · needs ~32 GB, NVIDIA only · from github.com/ByteDance-Seed (official)
    git clone https://github.com/bytedance-seed/BAGEL.git && cd BAGEL && conda create -n bagel python=3.10 -y && conda activate bagel && pip install -r requirements.txt && pip install flash_attn==2.5.8 --no-build-isolation && python -c "from huggingface_hub import snapshot_download; snapshot_download(cache_dir='models/BAGEL-7B-MoT/cache', local_dir='models/BAGEL-7B-MoT', repo_id='ByteDance-Seed/BAGEL-7B-MoT', local_dir_use_symlinks=False, resume_download=True, allow_patterns=['*.json', '*.safetensors', '*.bin', '*.py', '*.md', '*.txt'])" && python app.py

Official weights and documentation: huggingface.co/ByteDance-Seed ↗

$ how the score was worked out · LMArena leaderboard dataset 23%, GenExam: A Multidisciplinary… 12%, T2I-ReasonBench: Benchmarking… 10% and 17 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
LMArena leaderboard dataset2▇▇▇▇▇▇▇ 23%CC BY 4.0
GenExam: A Multidisciplinary Text-to-Image Exam (arXiv v5)1▇▇▇▇ 12%CC-BY-4.0
T2I-ReasonBench: Benchmarking Reasoning-Informed Text-to-Image Generation (Findings of ACL 2026)1▇▇▇ 10%CC-BY-4.0
Edit-Compass & EditReward-Compass: A Unified Benchmark for Image Editing and Reward Modeling1▇▇▇ 9%CC-BY-4.0
TIIF-Bench: How Does Your T2I Model Follow Your Instructions? (arXiv v3)1▇▇ 8%CC-BY-4.0
ImagenWorld: Stress-Testing Image Generation Models with Explainable Human Evaluation on Open-ended Real-World Tasks2▇▇ 7%CC-BY-4.0
HP-Edit: A Human-Preference Post-Training Framework for Image Editing (arXiv:2604.19406v1)1▇ 4%CC-BY-4.0
UnicEdit-10M: A Dataset and Benchmark Breaking the Scale-Quality Barrier via Unified Verification for Reasoning-Enriched Edits1▇ 4%CC-BY-4.0
MMMG: A Massive, Multidisciplinary, Multi-Tier Generation Benchmark for Text-to-Image Reasoning (arXiv:2506.10963v2)1▇ 3%CC-BY-SA-4.0
GEditBench v2: A Human-Aligned Benchmark for General Image Editing1▇ 3%CC-BY-4.0
GEBench: Benchmarking Image Generation Models as GUI Environments1▇ 3%CC-BY-4.0
WEAVE: Unleashing and Benchmarking the In-context Interleaved Comprehension and Generation (arXiv:2511.11434v1)1▇ 2%CC-BY-4.0
WISE: A World Knowledge-Informed Semantic Evaluation for Text-to-Image Generation (arXiv v4)1▇ 2%CC-BY-SA-4.0
CREval: An Automated Interpretable Evaluation for Creative Image Manipulation under Complex Instructions (arXiv:2603.26174v1)1▇ 2%CC-BY-4.0
Beyond Accuracy: Benchmarking Cross-Task Consistency in Unified Multimodal Models (XTC-Bench, arXiv:2604.25072v1)1▇ 2%CC-BY-4.0
MultiTextEdit: Benchmarking Cross-Lingual Degradation in Text-in-Image Editing (arXiv:2605.08163v2), Table 31▇ 2%CC BY 4.0
BAAI — OmniGen2 technical report1▇ 1%CC BY 4.0
ViGoR-Bench: How Far Are Visual Generative Models From Zero-Shot Visual Reasoners? (arXiv:2603.25823v1)1▇ 1%CC-BY-4.0
PaintBench: Deterministic Evaluation of Precise Visual Editing (arXiv:2606.00188v1)1▇ 1%CC-BY-4.0
Knowledge Visualization: A Benchmark and Method for Knowledge-Intensive Text-to-Image Generation (arXiv:2604.22302v1)1▇ 0%CC-BY-4.0
prior · pull toward the middle▇ 0%
weighted average= 42.0

not counted (3)

ByteDance Seed — BAGEL (Emerging Properties in Unified Multimodal Pretraining)GEdit-Bench-EN overall (G_O, GPT-4.1 judge) [correct]65.2%fewer than three models have this test yet, or this source ran it for this model alone
ByteDance Seed — BAGEL (Emerging Properties in Unified Multimodal Pretraining)GenEval [correct]82.0%fewer than three models have this test yet, or this source ran it for this model alone
ByteDance Seed — BAGEL (Emerging Properties in Unified Multimodal Pretraining)WISE overall (model publisher's run) [correct]70.0%fewer than three models have this test yet, or this source ran it for this model alone

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json