Instructions to use cpral/IMG-116M-10pct-preview with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use cpral/IMG-116M-10pct-preview with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("cpral/IMG-116M-10pct-preview", dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Draw Things
- DiffusionBee
import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("cpral/IMG-116M-10pct-preview", dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]IMG-116M (10% data preview)
Experimental 256×256 text-to-image model trained from scratch with the AuroraIMG2-50M-Preview recipe, scaled up: 103.7M-parameter flow-matching DiT (768 wide × 12 blocks) + 12.6M text encoder trained jointly, frozen FLUX.2 small decoder. This is a pipeline check on 10% of the data. Expect it to be clearly weaker than models trained on the full set.
Eight prompts from the AuroraIMG2 card, seed 0, 50 Euler steps, CFG 3.0.
Training
- Data: 10% of QLNI/FLUX-Reason-6M-flux2-latents (evenly spaced shards, ~0.5M images), ~11 epochs: 3 098 steps × batch 2048 = 6.3M samples.
- 8× RTX 3090 Ti, DDP, bf16 autocast, torch.compile.
- Flow matching (velocity, logit-normal t), 10% CFG dropout, auxiliary MLM loss on the text encoder (0.05, every 4th step).
- AdamW (0.9, 0.99), weight decay 0.1, lrs scaled from batch 256 by √8: hidden 2.83e-4, embedding 1.41e-3, in/out 8.49e-4. 2% warmup, then constant lr.
- Released weights: WSM (Warmup-Stable-Merge, arXiv:2507.17634): uniform average of the last 6 checkpoints (last 5% of training) instead of lr decay. Held-out velocity MSE: 1.1187 (EMA: 1.1194, last raw weights: 1.1220).
Training code: https://github.com/adamo1139/trenowanie_generatora_obrazkow
Usage
import torch
from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained("$REPO", custom_pipeline="$REPO", trust_remote_code=True).to("cuda")
image = pipe("a red fox sitting in a snowy forest at sunrise, photograph",
num_inference_steps=50, guidance_scale=3.0).images[0]
image.save("out.png")
License
Weights and code: Apache-2.0 (model code from AuroraIMG2-50M-Preview, Apache-2.0).
vae/: FLUX.2 small decoder © Black Forest Labs, Apache-2.0, unchanged. Training data has its own terms, see the dataset cards.
- Downloads last month
- 18
