whatmodel.sh / image / Kolors
Kolors
↗ share
+ contribute
Kolors is an open-weights image model from Kuaishou (Kolors team), released July 2024. Its published size is 2.6B parameters. Licence: kolors-model-license.
Catalogue identity notes
Kwai-Kolors/Kolors 1.0 (not Kolors 2.x). 2.6B U-Net per the Kolors technical report Table 4; memory estimate adds the ChatGLM3-6B text encoder (~6.2B). Code is Apache-2.0; weights are under the separate Kolors model licence (free for academic use, commercial use after registration).
Score within the LMArena text-to-image comparison set only.
$ how it compares
The image models ranked either side of Kolors, out of 197 with a score.
| # | model | score | where |
|---|---|---|---|
| 148 |  Stable Diffusion XL 0.9 (base) | 40±6 | local |
| 149 | Stable Diffusion 3 Medium | 40±2 | api $0.035/image +local |
| 150 | Kolors | 39±3 | local |
| 151 |  BLIP3-o 8B | 38±3 | local |
| 152 | Stable Cascade | 38±4 | local |
$ what it's good at
text to image▂▅▅▃▄▄█▃▁▂▁▁ #96 of 126 · 96% of its score · 3 tests
Beats Latent Diffusion text2img-large (LDM-KL-8-G, LAION-400M), VQ-Diffusion (ITHQ), Stable Diffusion XL 0.9 (base) here, though they rank higher overall. Trails SDXL-Lightning (4-step UNet), Stable Cascade, BLIP3-o 8B here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| EvalMuse-40K model ranking (FGA-BLIP2 overall alignment) | 2.93 / 5 ✎ | █▄·▄██▄▄█▄▄▄ beats 7/14 | 73% |
| KolorsPrompts human overall satisfaction (Kolors team) lab | 3.59 / 5 ✎ | ▄·▄········█ beats 3/3 | 19% |
| HPSv3 image generation benchmark (Overall HPSv3 score) | 10.5500 ✎ | ▃·····▃··▃·█ beats 5/5 | 3.4% |
$ where to run it
Other ways to run it
- Run it as published · needs ~21.6 GB, NVIDIA only · from huggingface.co/Kwai-Kolors (official)
uv run --with diffusers --with transformers --with accelerate --with torch --with sentencepiece python -c "import torch; from diffusers import KolorsPipeline; p=KolorsPipeline.from_pretrained('Kwai-Kolors/Kolors-diffusers', torch_dtype=torch.float16, variant='fp16').to('cuda'); p(prompt='a cat holding a sign', guidance_scale=5.0, num_inference_steps=50).images[0].save('out.png')"
Official weights and documentation: huggingface.co/Kwai-Kolors ↗
$ how the score was worked out · EvalMuse-40K: A Reliable and… 73%, Kwai-Kolors/Kolors repository… 19%, HPSv3 repository README, imag… 3%
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
| source | results | share of the score |
|---|---|---|
 EvalMuse-40K: A Reliable and Fine-Grained Benchmark with Comprehensive Human Annotations for Text-to-Image Generation Model Evaluation (arXiv v2) | 1 | 73% |
 Kwai-Kolors/Kolors repository README, human assessment · lab-reported | 1 | 19% |
| HPSv3 repository README, image generation benchmark table (MizzenAI/HPSv3) | 1 | 3% |
| prior · pull toward the middle | 4% | |
| weighted average | = 38.6 |
$ improve this page
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