README.md

metadata

title: GenFirst Text-to-Image
emoji: 🎨
colorFrom: indigo
colorTo: gray
sdk: gradio
sdk_version: 6.31.0
app_file: app.py
short_description: GenFirst EiT MMDiT-XL/2 512px text-to-image
python_version: '3.12'
startup_duration_timeout: 30m
license: cc-by-nc-4.0
models:
  - ByteDance-Seed/GenFirst
  - openai/clip-vit-large-patch14

Demo of the GenFirst text-to-image prior (EiT, MMDiT-XL/2, variant A, 512x512) from ByteDance-Seed/GenFirst.

At startup the Space loads the official files directly from the model repo:

• the EMA weights of text-to-image/variant-a/eit-t2i-variant-a-512.pt. Only the EMA tensors (~5.3 GB of the 21 GB checkpoint) are fetched, using HTTP range reads (lazy_ckpt.py).
• the paired end-to-end FLUX VAE vae/e2e-flux-vae-sit-imagenet-s2 and its latent statistics.

Sampling runs under fp16 autocast, as in the reference infer_t2i.py, and follows the authors' GenEval protocol: 50-step Euler, CFG 7.0 on the interval [0.35, 1.0].

Weights are © ByteDance, licensed CC BY-NC 4.0 (non-commercial research use only). Code is ported from the MIT-licensed GitHub repository (the MMDiT comes from Stability-AI/sd3-ref, MIT). Example prompts come from the authors' infer_t2i.py defaults and their GenEval smoke set.