Skip to content
Other

Primal VAE

компактный VAE без scaling factor

2 min read

Primal VAE от Ostris — это VAE для diffusion, который делает только 2x spatial down до 1 или 2 каналов. После patch 4 получаются 16 или 32 канала при 8x — та же форма, что у Flux.1 VAE и Flux.2 VAE. Обе версии — примерно 80M параметров.

Ключевые особенности

  • Два варианта: 2x1 (1 ch @ 2x → 16 ch @ 8x) и 2x2 (2 ch @ 2x → 32 ch @ 8x) после patch 4.
  • Латент обучен быть похожим на гауссов шум: без scaling_factor, shift_factor и latent normalization.
  • patchify/unpatchify принимают общий compression; compression=8 — это патч 4×4.
  • Conv U-Net с RMSNorm по каналам, без attention; тайлы кодируются и декодируются так же, как полное изображение, при наличии overlap.
  • Через decode можно пускать backprop с gradient checkpointing; инференс в bf16 стоит около 0.05 dB.
  • Лицензия Apache 2.0, веса fp32, архитектура читается из формы тензоров.

Пример

import torch
from diffusers import AutoModel

vae = AutoModel.from_pretrained("ostris/primal_vae", variant="2x1", trust_remote_code=True, torch_dtype=torch.bfloat16).to("cuda")
z = vae.encode(x)       # [B, 1, H/2, W/2], ~N(0, 1)
x_hat = vae.decode(z)   # [B, 3, H, W]
ПодробнееPSNR на трёх 512 px изображениях: Primal 2x2 — 37.5 / 27.1 / 27.4 dB, Flux.2 VAE — 37.2 / 26.5 / 25.9 dB; Primal 2x1 уступает Flux.1 на текстурах: 22.0 против 24.2 dB на img02.
2x2 обучен 147k шагов, 2x1 — 62k; у 2x1 std латента ещё дрейфует (1.07–1.21), у 2x2 — в пределах нескольких процентов от 1.
Стороны изображения должны быть кратны 16. Для инференса bf16 стоит ~0.05 dB. Лицензия Apache 2.0.
Три изображения и PSNR — не полноценный бенчмарк.
Discuss Telegram 44 964

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.5 GB · ready files1.0 GB
  • FP8 / INT8recommendedweights 0.2 GB0.8 GB
  • 4-bitweights 0.1 GB0.6 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # ostris/primal_vae
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download ostris/primal_vae --local-dir ~/models/primal_vae
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to primal_vae
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of primal_vae (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 1All on Hugging Face

About the model

ostris/primal_vae
Author
ostris
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

96444

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments