Supra2-IMG safetensors

Safetensors build of SupraLabs/Supra2-IMG. The announcement is Massive Release: Supra2-IMG, a tiny 100M text-to-image model on r/StableDiffusion.

The model is SupraLabs' work: a 104.1M-parameter text-to-image DiT, 256×256, trained from scratch for 10 epochs on LucasFang/FLUX-Reason-6M. The text encoder is frozen Flan-T5-Base. The VAE is SD-VAE-FT-MSE. License is Apache-2.0. Architecture and training details are on the original model card.

This repo

The original checkpoint is model_final_ema.pt. This repo hosts that checkpoint converted to FP32 safetensors.

Supra2-IMG_fp32.safetensors is a safetensors file, not a renamed PyTorch checkpoint. It holds 237 FP32 weight tensors plus the stored unconditional text embedding and mask used for classifier-free guidance. Metadata keeps patch=2 and ctx_len=128.

Flan-T5-Base and SD-VAE-FT-MSE are included under their own names. They are the original models, not part of Supra2. inference.py loads the DiT, the text encoder, the tokenizer, and the VAE from this repo.

Files the script does not read were left out: the SentencePiece model, tokenizer_config.json, special_tokens_map.json, and the original repo config.json. The tokenizer vocab is in flan-t5-base.tokenizer.json.

generate.bat and generate.sh take a prompt and an output filename.

Files

Generate

generate.bat "a boat" boat.png

./generate.sh "a boat" boat.png

A bare filename is written next to the script. A full path is used as given.

python inference.py --prompt "a boat" --seed 0 --cfg 3.0 --steps 50 --out boat.png

Recommended settings from the original model card: seed 0, CFG 3.0, steps 50.