whatmodel.sh / image / Kandinsky 3
Kandinsky 3
↗ share
+ contribute
Kandinsky 3 is an open-weights image model from AI Forever, released November 2023. Its published size is 3B parameters. Licence: apache-2.0.
Catalogue identity notes
Full text-to-image pipeline including prior/decoder dependencies. No scores copied from third-party model comparisons. Memory is a conservative planning estimate including pipeline components and activation headroom; no local runtime measurement.
Model information sources
3B = the latent diffusion U-Net only. The maker also lists the Flan-UL2 text encoder (encoder part) 8.6B and MoVQ encoder/decoder 267M; whole pipeline OUR SUM = 11.87B.
Score within the LMArena text-to-image comparison set only.
$ how it compares
The image models ranked either side of Kandinsky 3, out of 197 with a score, then more from AI Forever.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 105 |  Lucid Origin | 49±3 | api $0.017/image | ||
| 106 |  Latent Diffusion text2img-large (LDM-KL-8-G, LAION-400M) | 48±15 | local | ||
| 107 | Kandinsky 3 | 48±11 | local | ||
| 108 |  Step1X-Edit v1.1 | 48±3 | local | ||
| 109 | FuseDream (BigGAN 512) | 48±15 | comparison only | ||
| more from AI Forever | |||||
| 157 | Kandinsky 2.2 | 37±5 | local | ||
$ what it's good at
text to image▂▅▅▃▄▄█▃▁▂▁▁ #72 of 126 * · 54% of its score · 7 tests
Beats Lucid Origin, Latent Diffusion text2img-large (LDM-KL-8-G, LAION-400M) here, though they rank higher overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| REAL realism benchmark for T2I (average of attribute, relationship and style scores) | 63.2% ✎ | █··········▄ beats 2/2 | 54% |
| Kandinsky 3.0 · text alignment wins against kandinsky-2-2 lab | 52.0% ✎ | no other model yet | — |
| Kandinsky 3.0 · text alignment wins against dall-e-3 lab | 36.0% ✎ | no other model yet | — |
| Kandinsky 3.0 · text alignment wins against sdxl lab | 53.0% ✎ | no other model yet | — |
| Kandinsky 3.0 · visual quality wins against kandinsky-2-2 lab | 50.0% ✎ | no other model yet | — |
| Kandinsky 3.0 · visual quality wins against dall-e-3 lab | 42.0% ✎ | no other model yet | — |
| Kandinsky 3.0 · visual quality wins against sdxl lab | 57.0% ✎ | no other model yet | — |
$ where to run it
Other ways to run it
- Run it as published · needs ~32 GB, NVIDIA only · from huggingface.co/kandinsky-community
uv run --with torch --with diffusers --with transformers --with accelerate --with safetensors python -c "import torch; from diffusers import AutoPipelineForText2Image; p=AutoPipelineForText2Image.from_pretrained('kandinsky-community/kandinsky-3', torch_dtype=torch.float16, variant='fp16').to('cuda'); p('A lighthouse at sunset', height=512, width=512).images[0].save('image.png')"
Weights and documentation: huggingface.co/kandinsky-community ↗
$ how the score was worked out · REAL: Realism Evaluation of T… 54%
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 |
|---|---|---|
 REAL: Realism Evaluation of Text-to-Image Generation Models for Effective Data Augmentation (arXiv:2502.10663v1) | 1 | 54% |
| prior · pull toward the middle | 46% | |
| weighted average | = 48.4 |
not counted (6)
| Kandinsky 3.0 — original human comparisons | Kandinsky 3.0 · text alignment wins against kandinsky-2-2 [correct] | 52.0% | fewer than three models have this test yet, or this source ran it for this model alone |
| Kandinsky 3.0 — original human comparisons | Kandinsky 3.0 · visual quality wins against kandinsky-2-2 [correct] | 50.0% | fewer than three models have this test yet, or this source ran it for this model alone |
| Kandinsky 3.0 — original human comparisons | Kandinsky 3.0 · text alignment wins against dall-e-3 [correct] | 36.0% | fewer than three models have this test yet, or this source ran it for this model alone |
| Kandinsky 3.0 — original human comparisons | Kandinsky 3.0 · visual quality wins against dall-e-3 [correct] | 42.0% | fewer than three models have this test yet, or this source ran it for this model alone |
| Kandinsky 3.0 — original human comparisons | Kandinsky 3.0 · text alignment wins against sdxl [correct] | 53.0% | fewer than three models have this test yet, or this source ran it for this model alone |
| Kandinsky 3.0 — original human comparisons | Kandinsky 3.0 · visual quality wins against sdxl [correct] | 57.0% | fewer than three models have this test yet, or this source ran it for this model alone |
$ 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