How to Run Supra2-IMG Locally: SupraLabs' 104M-Parameter Text-to-Image Model

Supra2-IMG is a 104.1 million parameter text-to-image model from SupraLabs, trained from scratch on 5.6 million synthetic images and released under Apache 2.0. Here is what it is, how it was trained, and how to run it from the single inference script on its Hugging Face page.

6 minBeginner

Supra2-IMG is a 104.1 million parameter text-to-image model from SupraLabs, released September 21, 2026 under the Apache 2.0 license. It is a small diffusion transformer trained from scratch on 5.6 million synthetic images, pairs a frozen Flan-T5-Base text encoder with a Stable Diffusion VAE, and generates 256 by 256 pixel images from a single downloadable inference script, no clone required.

Written by Priya Raghunathan, local-AI hardware reviewer. I check parameter counts and training details against the primary model card before recommending an install path, and I say so plainly when a card leaves a number out.

Supra2-IMG entered the top 20 of Hugging Face's text-to-image trending list. At time of writing it has 149 likes and 515 downloads on the SupraLabs/Supra2-IMG model page, with a demo Space, SupraLabs/supra2-img-demo (13 likes), if you want to see sample output before installing anything. The model card lists no GitHub repository; everything needed to run it is hosted directly on the Hugging Face model page.

Specs at a glance

Field	Value
Publisher	SupraLabs
Parameters	104.1 million
Architecture	Diffusion transformer (DiT), 14 layers, 576 model width, 9 attention heads
Text encoder	Frozen Flan-T5-Base, 128-token context
Image VAE	SD-VAE-FT-MSE (Stability AI)
Output resolution	256 by 256 pixels
Training data	LucasFang/FLUX-Reason-6M, 5.6 million images, 10 epochs
License	Apache 2.0
Release date	September 21, 2026

How it was built

SupraLabs describes Supra2-IMG as a tiny diffusion transformer trained from scratch, not a fine-tune of an existing checkpoint. It reuses two off-the-shelf components rather than training its own: a frozen Flan-T5-Base model reads the text prompt, and Stability AI's SD-VAE-FT-MSE decodes the model's 32 by 32 latent output into the final 256 by 256 image. Everything in between, a 14-layer transformer with a 576-dimension model width and 9 attention heads, was trained specifically for this model.

The training set is the full LucasFang/FLUX-Reason-6M dataset, and SupraLabs ran 10 full epochs over it rather than a partial pass. Captions were chosen with a defined fallback order, composition first, then entity, then plain text, then style, then imaginative description, so the model trained on the highest-quality caption available for each image rather than a single fixed field.

Trained on one GPU in under a day

SupraLabs trained Supra2-IMG on a single Nvidia H100 SXM with 80GB of VRAM, rented from Runpod, for 9 hours including data preparation. That is a training-time figure, not an inference requirement, but it is a useful data point for how much compute a capable small text-to-image model now takes to produce from scratch.

Run it: one script, no repo clone

SupraLabs ships a single inference script directly on the Hugging Face model page rather than a GitHub repository. Create a project folder and download it with wget, exactly as shown on the model card:

bash - set up Supra2-IMG

$mkdir Supra2-IMG

$cd Supra2-IMG

Download the inference script

$wget https://huggingface.co/SupraLabs/Supra2-IMG/resolve/main/inference.py

Then generate an image by passing a prompt and the model card's recommended sampling settings, seed 0, a classifier-free guidance scale of 3.0, and 50 steps:

bash - generate an image

$python inference.py --prompt "a sea jellyfish floating in the pitch-black ocean depths" --seed 0 --cfg 3.0 --steps 50 --n 1 --out jellyfish.png

=== Supra2-IMG inference ===

[ckpt] found ./model_final_ema.pt

[model] building SupraDiT ...

[model] 104.1M parameters

[text] loading tokenizer + google/flan-t5-base ...

[sample] Euler flow, 50 steps ...

[done] saved 1 image(s) -> jellyfish.png

Recommended sampling settings

The model card's own example uses --seed 0, --cfg 3.0, and --steps 50. Start there before tuning; those are the values SupraLabs tested the model with.

Check the model page for the checkpoint file

The script's log output references a local checkpoint, model_final_ema.pt, in the same folder. The model card shows only the wget command for inference.py itself, so confirm on the Hugging Face model page which additional files the script expects before you run it, rather than assuming the script fetches the checkpoint automatically.

Licensing: Apache 2.0, no restrictions beyond attribution

Supra2-IMG is released under Apache 2.0, the same permissive license SupraLabs used for its SupraTTS speech model. Commercial use is allowed outright, with no revenue threshold and no separate enterprise tier to contact, unlike Stability AI's Community License on some SD 3.5 checkpoints.

FAQ

How many parameters does Supra2-IMG have?

104.1 million. SupraLabs describes it as a tiny diffusion transformer, far smaller than mainstream text-to-image models like SDXL (roughly 3.5 billion parameters) or Qwen-Image (20 billion).

What resolution does Supra2-IMG generate?

256 by 256 pixels. That is lower than most current text-to-image models, a trade-off that comes with the model's small parameter count and single-GPU training run.

Is Supra2-IMG free for commercial use?

Yes. It ships under Apache 2.0 with no revenue cap and no contact-required enterprise tier.

Does Supra2-IMG need a GPU to run?

The model card does not publish a minimum VRAM figure or confirm a CPU-only path. Given its small size, 104.1 million parameters, it should run on modest hardware, but test the inference script on your own machine before relying on it.

Local models run better with more VRAM. CompareRTX GPUs on Amazonbefore you upgrade.(affiliate link. We may earn a commission at no extra cost. Disclosure)