whatmodel.sh / image / IF-L
IF-L
↗ share
+ contribute
IF-L is an open-weights image model from DeepFloyd, released April 2023. Its published size is 0.9B parameters. Licence: deepfloyd-if-license.
Catalogue identity notes
DeepFloyd/IF-I-L-v1.0 (900M) with IF-II-L-v1.0 (1.2B) per the DeepFloyd IF model zoo; memory as for IF-M (estimate).
Score within the LMArena text-to-image comparison set only.
$ how it compares
The image models ranked either side of IF-L, out of 197 with a score, then more from DeepFloyd.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 169 |  OmniGen (v1) | 34±4 | local | ||
| 170 |  SDXL-Turbo | 34±2 | local | ||
| 171 | IF-L | 33±3 | local | ||
| 172 |  Lumina-DiMOO | 32±4 | local | ||
| 173 |  UltraEdit (SD3, with mask) | 32±7 | comparison only | ||
| more from DeepFloyd | |||||
| 166 | IF-XL | 35±2 | local | ||
| 177 | IF-M | 31±3 | local | ||
$ what it's good at
text to image▂▅▅▃▄▄█▃▁▂▁▁ #108 of 126 * · 43% of its score · 1 test
Beats Latent Diffusion text2img-large (LDM-KL-8-G, LAION-400M), VQ-Diffusion (ITHQ), LCM-SDXL here, though they rank higher overall.
other tests2 not in a task · 54% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| EvalAlign human text-image alignment score | 5.2300 mean human alignment score ✎ | █▄▄·█·▄█▄▄·▄ beats 10/11 | 42% |
| HEIM text-image alignment (aspect win rate, v1.1.0) | 51.4% ✎ | ▁····█▁·▁··▁ beats 1/9 | 12% |
$ where to run it
Other ways to run it
- Run it as published · needs ~14 GB, NVIDIA only · from github.com/deep-floyd (official)
uv run --with diffusers --with transformers --with accelerate --with torch --with sentencepiece --with protobuf python -c "import torch; from diffusers import DiffusionPipeline; s1=DiffusionPipeline.from_pretrained('DeepFloyd/IF-I-L-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 · GenEval authors' results: djg… 43%, EvalAlign: Supervised Fine-Tu… 42%, HEIM: Holistic Evaluation of… 12%
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 |
|---|---|---|
 GenEval authors' results: djghosh13/geneval README and arXiv 2310.11513v1 | 1 | 43% |
 EvalAlign: Supervised Fine-Tuning Multimodal LLMs with Human-Aligned Data for Evaluating Text-to-Image Models (arXiv:2406.16562v3) | 1 | 42% |
| HEIM: Holistic Evaluation of Text-to-Image Models (arXiv 2311.04287v1) | 1 | 12% |
| prior · pull toward the middle | 3% | |
| weighted average | = 32.5 |
$ 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