whatmodel.sh / image / IF-XL
IF-XL
↗ share
+ contribute
IF-XL is an open-weights image model from DeepFloyd, released April 2023. Its published size is 4.3B parameters. Licence: deepfloyd-if-license.
Model information sources
Release date ↗ · Weights licence ↗ · Parameter count ↗
Stage I (text-to-64x64 base module) only, matching how the catalogue records deepfloyd-if-m (0.4) and -l (0.9). The HF card also says 'Cascade Stage: I Num Parameters: 4.3B'. T5-XXL encoder and the stage II/III upscalers are not included.
Score within the LMArena text-to-image comparison set only.
$ how it compares
The image models ranked either side of IF-XL, out of 197 with a score, then more from DeepFloyd.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 164 |  Emu3 | 36±2 | local | ||
| 165 |  Ovis-U1-3B | 35±6 | comparison only | ||
| 166 | IF-XL | 35±2 | local | ||
| 167 |  Stable Diffusion XL 1.0 | 34±1 | api $0.0019/image +local | ||
| 168 |  LCM-SDXL | 34±3 | local | ||
| more from DeepFloyd | |||||
| 171 | IF-L | 33±3 | local | ||
| 177 | IF-M | 31±3 | local | ||
$ what it's good at
text to image▂▅▅▃▄▄█▃▁▂▁▁ #107 of 126 · 72% of its score · 5 tests
Beats Latent Diffusion text2img-large (LDM-KL-8-G, LAION-400M), VQ-Diffusion (ITHQ) here, though they rank higher overall. Trails SDXL-Turbo, Stable Diffusion XL 1.0 here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| EvalMuse-40K model ranking (FGA-BLIP2 overall alignment) | 2.77 / 5 ✎ | █▄·▄██▄▄█▄▄▄ beats 5/14 | 23% |
| GenEval | 61.0% ✎ | ▅█···▅▃▃▃··▃ beats 10/10 | 21% |
| ConceptMix (k=2 full-mark score) | 38.0% ✎ | ▄·▄·█▄▄▄···▄ beats 4/7 | 18% |
| ImagenHub text-guided generation (human overall) | 59.0% ✎ | ▄▄·█·······▄ beats 3/4 | 10% |
| HPS v2 benchmark (4-style average) | 27.3% ✎ | ▃··▃▅·▅·▅▅█▅ beats 11/14 | 0.1% |
other tests2 not in a task · 26% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| EvalAlign human text-image alignment score | 5.4500 mean human alignment score ✎ | █▄▄·█·▄█▄▄·▄ beats 11/11 | 21% |
| HEIM text-image alignment (aspect win rate, v1.1.0) | 58.9% ✎ | ▁····█▁·▁··▁ beats 4/9 | 5.5% |
$ where to run it
Other ways to run it
- Run it as published · needs ~29.1 GB, NVIDIA only · from github.com/deep-floyd (official)
uv run --with diffusers --with transformers --with accelerate --with torch --with safetensors --with sentencepiece --with protobuf python -c "import torch; from diffusers import DiffusionPipeline; s1=DiffusionPipeline.from_pretrained('DeepFloyd/IF-I-XL-v1.0', variant='fp16', torch_dtype=torch.float16); s1.enable_model_cpu_offload(); s2=DiffusionPipeline.from_pretrained('DeepFloyd/IF-II-L-v1.0', text_encoder=None, variant='fp16', torch_dtype=torch.float16); s2.enable_model_cpu_offload(); e,n=s1.encode_prompt('a cat holding a sign'); i=s1(prompt_embeds=e, negative_prompt_embeds=n, output_type='pt').images; s2(image=i, prompt_embeds=e, negative_prompt_embeds=n, output_type='pil').images[0].save('out.png')"
Official weights and documentation: huggingface.co/DeepFloyd ↗
$ how the score was worked out · EvalMuse-40K: A Reliable and… 23%, GenEval authors' results: djg… 21%, EvalAlign: Supervised Fine-Tu… 21% and 4 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
$ 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