---
license: other
private: false
tags:
- imagenet-1k
- image-generation
- flux2
- vae-latents
- safetensors
- pdm3-ht
---

# PDM-3-HT FLUX.2 VAE latents for ImageNet-256 train

Public research artifact for PDM-3-HT VAE-backend experiments. This repository contains **latent cache shards only**. It intentionally does **not** contain raw ImageNet images, ADM-cropped uint8 images, PAE latents, PAE checkpoints, or training checkpoints.

## Source and preprocessing

- Source dataset: ImageNet-1k train via `ILSVRC/imagenet-1k`; access requires accepting the upstream ImageNet terms.
- Image preprocessing before VAE encode: ADM-style center crop to 256x256, then uint8 `[0,255]` converted to float `[-1,1]`.
- Raw/cropped images are not redistributed here.

## Encoder

- Diffusers model repo: `diffusers/FLUX.2-dev-bnb-4bit`
- Subfolder: `vae`
- Diffusers class: `AutoencoderKLFlux2`
- Latent mode: deterministic posterior `mode()`
- VAE forward dtype used for cache build: fp32
- Saved latent dtype: bf16

## Schema

Each `latents_rank00_shard*.safetensors` file contains:

| key | dtype | shape | meaning |
| --- | --- | --- | --- |
| `latents` | bf16 | `[N, 32, 32, 32]` | VAE latent for the cropped image |
| `latents_flip` | bf16 | `[N, 32, 32, 32]` | VAE latent for the horizontal flip |
| `labels` | int64 | `[N]` | ImageNet class label |

Shard size is `4096` samples except the final partial shard (`3215` samples).

## Verified cache summary

- Samples: `1,281,167`
- Shards: `313`
- Total safetensors size: `167.936` GB decimal
- Last shard: `latents_rank00_shard000312.safetensors`
- Generated/upload-prepared UTC: `2026-05-28T03:18:07+00:00`

## Intended use

This cache is for testing whether PDM-3-HT's HT + MeanFlow trajectory mechanism is VAE-backend agnostic. PAE is treated as a baseline latent backend; this repository is the FLUX.2 VAE backend cache for swap-in ablations.