This release provides **two specialized, post-trained <code>AutoencoderKLFlux2</code> VAEs** for **FLUX 3 Image Klein (5.6B)** and **FLUX.2** in ComfyUI, inspired by [Teaching a FLUX.2 VAE to See Transparency <code>trmz/flux2-klein-alpha</code>)](<a target="_blank" rel="ugc" href="https://tr-mz.github.io/papers/flux2-klein-alpha/)*">https://tr-mz.github.io/papers/flux2-klein-alpha/)</a> and [UltraFlux <code>arXiv:2511.18050</code>)](<a target="_blank" rel="ugc" href="https://arxiv.org/abs/2511.18050)*">https://arxiv.org/abs/2511.18050)</a>. Both VAEs keep their **Encoder and 128-channel packed latent space 100% frozen**, making them **100% drop-in compatible** with existing models, LoRAs, and seeds without altering composition.---## 📦 Which Version Should I Download?| Version / File | Channels | Best Used For | Key Characteristics || :--- | :---: | :--- | :--- || *<code>flux3_ultra_rgba_vae.safetensors</code>** (Recommended for RGBA) | **4-Channel <code>RGBA</code>)** | Native transparent PNG generation &amp; foreground extraction with the **FLUX 3 RGBA Transparent &amp; Extract LoRAs** | **+6.55 dB** Foreground RGB PSNR <code>32.74</code> $\rightarrow$ <code>39.29 dB</code>), **+3.42 dB** Alpha PSNR, and **-55% less hidden background color bleed** on fine hair, fur, and semi-transparent glass/wings. | | *<code>flux3_ultra_rgb_vae.safetensors</code>** (Experimental Study) | **3-Channel <code>RGB</code>)** | Standard **FLUX 3 Image Klein** photography workflows (via vanilla ComfyUI <code>Load VAE</code>) | **Experimental decoder fine-tune** initialized 100% from native <code>ae_flux_3.safetensors</code>. Uses Phase-Coherent YCbCr Luminance-Wavelet distillation on 2K DSLR <code>DIV2K</code>) &amp; editorial <code>Vogue Runway</code>) photos. Delivers **subtle micro-detail refinement** <code>~+16–20%</code> Laplacian micro-contrast on 1–2 px pores/hair at 100–200% zoom) with **0.00% color shift**. |---## 🧪 Honest Technical Notes &amp; What to Expect### 1. <code>flux3_ultra_rgba_vae.safetensors</code> (4-Channel RGBA Transparency VAE)Built on top of **[trmz/flux2-klein-alpha](<a target="_blank" rel="ugc" href="https://huggingface.co/trmz/flux2-klein-alpha/tree/main/vae)**">https://huggingface.co/trmz/flux2-klein-alpha/tree/main/vae)**</a> and post-trained on **1,500 multi-layer RGBA matting assets** <code>MULAN</code> + <code>LayerDecomp</code> + high-res photography) using:* **Visibility-Weighted Premultiplied Reconstruction:** Prioritizes color accuracy on visible and semi-transparent foreground pixels ($\alpha &gt; 0$) so foreground colors never look washed out.* **Hidden-Background Suppression ($\alpha &lt; 0.01 \rightarrow \text{RGB}=0$):** Cuts invisible background RGB residue in half <code>-55%</code>), preventing dark/colored halos around fine hair strands and translucent edges.* **2D Haar Wavelet &amp; Sobel Alpha-Edge Loss:** Sharpens high-frequency alpha transitions <code>39.95 dB</code> $\rightarrow$ <code>43.37 dB</code> Alpha PSNR). ### 2. <code>flux3_ultra_rgb_vae.safetensors</code> (Experimental 3-Channel RGB Photo VAE)&gt; **⚠️ Realistic Expectations (No Hype):**&gt; Unlike older 4-channel SD1.5/SDXL VAEs, Black Forest Labs' native 32-channel <code>128</code> packed) <code>ae_flux_3.safetensors</code> <code>flux2-vae</code>) is already near the theoretical reconstruction limit. Because the encoder must remain frozen to stay compatible with the diffusion model, **the visual difference between <code>ae_flux_3.safetensors</code> and <code>flux3_ultra_rgb_vae.safetensors</code> is very subtle at normal viewing distance.*** **Why separated it from the RGBA VAE:** In <code>trmz</code>'s 4-channel RGBA VAE, the encoder's <code>quant_conv</code> weights were shifted by <code>+16%</code> to accommodate the alpha channel. Using an RGBA-derived decoder on standard FLUX 3 latents causes a slight magenta shadow cast. Therefore, <code>flux3_ultra_rgb_vae.safetensors</code> was initialized **100% from native <code>ae_flux_3.safetensors</code>** <code>0.00%</code> color shift).* **How it was trained:** Only the final full-resolution decoding stage <code>decoder.up.3</code>, <code>norm_out</code>, <code>conv_out</code>) was fine-tuned on uncompressed 2K/4k DSLR reference photos <code>DIV2K_train_HR</code>) and fashion photography <code>Vogue Runway</code> — zero cartoons/illustrations) using **Phase-Coherent YCbCr Luminance-Wavelet Distillation**.* **Where you can spot the difference:** When comparing 2K–4K <code>2736×1536</code> / <code>3856×2176</code>) portrait renders at **100%–200% zoom** in an Image Compare slider, 1–2 px luminance micro-textures (skin pores, peach fuzz, beard hairs, eyelashes, glitter) show <code>~+16–20%</code> higher Laplacian acutance while colors, skin tones, and shadows remain 100% identical to <code>ae_flux_3</code>.---## 🛠️ Installation &amp; Usage in ComfyUI1. Place the <code>.safetensors</code> file(s) into: ```text ComfyUI/models/vae/

Files with SHA256: afc887129d0d2febc4af38de2fa5fda5ed30e951eb1c90c23fcba6dadc9ea3e7