GenFirst

GenFirst

Research weights from ByteDance-Seed for image generators whose VAE and prior are trained together. They cover class-conditional and text-to-image sampling, for non-commercial research use.

Made by ByteDance-Seed on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: GenFirst is a ByteDance-Seed research release about training a latent VAE together with a generative prior while avoiding latent collapse, which tends to make joint training unstable. Training runs in two phases: the first lets the prior shape the latent space, and the second cuts both weights tenfold so the decoder can recover detail. The release holds weights for class-conditional and text-to-image generation.

Who it suits: It suits researchers and engineers who work on image generation or latent diffusion and want to study or reproduce the method under non-commercial terms.

Industry | Who uses it | How they use it | Example
Academic research | Image-generation researchers | Study joint training of a VAE and a prior | Reproduce the class-conditional results at 256x256
Machine learning engineering | Teams exploring latent diffusion | Try the released SiT-XL and MMDiT-XL priors | Sample 512x512 images from a text prompt
Benchmarking | Evaluators | Compare scores across the released checkpoints | The 800 epoch checkpoint reaches a gFID of 1.45 at 256x256

You can skip it if you need a model for commercial work, or a ready diffusers pipeline.

Where to use it:

• PyTorch, with the inference code that the card says is required
• Windows, macOS or Linux with Python, to download the files
• Research setups for class-conditional images at 256x256 and 512x512
• Text-to-image research at 512x512

Use case | Who | What they get back
Class-conditional generation | Researchers | An image for one integer label between 0 and 999
Text-to-image generation | Researchers | A 512x512 image from a prompt read by a frozen CLIP-L encoder
Comparing results | Evaluators | Published scores such as GenEval 0.8987 for variant A

How to set it up: You need Python, PyTorch and the project's inference code, because these are raw checkpoints.

• Python with the huggingface_hub package
• PyTorch, the framework the card names
• The inference code from the GenFirst repository of ByteDance-Seed on GitHub
• Enough storage for fp32 weights

1. Install the Hugging Face tools with Python.
2. Download the repository files with the second command below.
3. Get the inference code from the GenFirst repository on GitHub.
4. Pair each generative prior with the end-to-end VAE it was trained on.

pip install -U huggingface_hub
hf download ByteDance-Seed/GenFirst

How to use it: Load a prior and its VAE through the project's inference code, since the files are raw .pt checkpoints and not diffusers pipelines. For the class-conditional models, pass a single integer label. For text-to-image, give a prompt. The frozen text encoder reads it.

Tool | What it does
End-to-end VAEs | Encoder and decoder from the end-to-end stage, each with its latent normalization statistics
SiT-XL priors | Class-conditional generation from one integer label
MMDiT-XL priors | Text-to-image generation with a frozen CLIP-L text encoder

Example: You choose the 800 epoch checkpoint at 256x256 and pass the integer label for one ImageNet class. The prior samples a latent for that class, and the paired VAE decodes it into an image. The card lists a gFID of 1.45 for this checkpoint without guidance.

For newcomers and experts: Think of it as research code and weights, not an app. There is no text box to type in. A first small step is to download the files, read the card, and pick the 256x256 class-conditional checkpoint.

The VAE and the prior are trained at once. In phase 1 the prior shapes the latents. In phase 2 both weights drop tenfold so the decoder recovers detail. The priors are SiT-XL and MMDiT-XL, with 28 blocks, width 1152, 16 attention heads and about 675M parameters, each trained on a frozen VAE. Only the EiT diffusion-prior version is released, not the autoregressive EAR one. The 512 models reuse the frozen 256 f16d32 VAE and train only the prior.

Phase 1: the prior shapes the latent space
Phase 2: drop both weights tenfold
The decoder recovers detail
Freeze the VAE and train a prior on it
Sample with the prior, decode with the VAE

Good to know: The class-conditional priors were scored on ImageNet at 256x256, with and without guidance.

Title: Class-conditional results at 256x256
Checkpoint | gFID | IS | gFID (guided) | IS (guided)
EiT, 80 epochs | 2.79 | 175.4 | 1.38 | 271.3
EiT, 480 epochs | 1.60 | 216.8 | 0.988 | 270.7
EiT, 800 epochs | 1.45 | 226.7 | 0.969 | 283.4

Title: GenEval score at 512x512, no prompt rewriting
Variant A: 0.8987
Variant B: 0.8920

• Precision: weights are released in fp32, and sampling was tested in bf16
• Not diffusers: the checkpoints are raw .pt files that need the inference code
• Not released: the autoregressive EAR version described in the paper
• Text encoder: CLIP-L, frozen and not modified by the authors
• License: CC BY-NC 4.0 for the weights, stricter than the MIT license on the code
• Cost to run: none named
• Data it reaches: none named beyond the files you download
• Not for: commercial use, since the weights are for non-commercial research only

On Hugging Face, 13 November 2026

• Task: text to image
• Library: pytorch
• Formats: Safetensors
• Downloads in the last 30 days: 0
• Likes: 7
• License: cc-by-nc-4.0
• Tags: pytorch, diffusers, safetensors, diffusion, latent-diffusion, end-to-end-training, variational-autoencoder, text-to-image, class-conditional-image-generation, mmdit

Links

• Model: https://huggingface.co/ByteDance-Seed/GenFirst
• Files: https://huggingface.co/ByteDance-Seed/GenFirst/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ByteDance-Seed/GenFirst

Then follow the usage in its model card. Nothing in this kit runs by itself.