# ✨ FLUX 3 Image Klein – Native RGBA Transparent Layer LoRAGenerate **true 4-channel transparent PNGs (RGBA)** directly from text prompts with **FLUX 3 Image Klein (5.6B)** no background removal post-processing <code>rembg</code> / <code>BiRefNet</code>) required!Instead of generating a flat RGB image and guessing a mask afterward, this LoRA shifts **FLUX 3 Image Klein** into the native 4-channel RGBA latent space of the **AlphaVAE <code>AutoencoderKLFlux2</code> 4-channel RGBA VAE)**, allowing the diffusion model to natively synthesize clean alpha silhouettes, crisp cutouts, and soft edge transitions directly during sampling.---## 🔧 Required Components (Important!)Because standard VAEs only decode 3 channels (RGB), **you MUST use the 4-channel RGBA VAE and the RGBA Decode nodes** to get the alpha channel:1. **LoRA:** <code>flux3_klein_transparent_lora.safetensors</code> (place in <code>ComfyUI/models/loras/</code>)2. **4-Channel RGBA VAE:** <code>flux2_3_rgba_vae.safetensors</code>](<a target="_blank" rel="ugc" href="https://huggingface.co/trmz/flux2-klein-alpha/resolve/main/vae/diffusion_pytorch_model.safetensors">https://huggingface.co/trmz/flux2-klein-alpha/resolve/main/vae/diffusion_pytorch_model.safetensors</a>) (place in <code>ComfyUI/models/vae/</code> — auto-downloads automatically if enabled in the loader node)3. **ComfyUI Custom Nodes:** *<code>denrakeiw_nodes</code>** <code>🔧 Flux 2/3 RGBA VAE Loader DRE</code> <code>🔍 Flux 2/3 RGBA VAE Decode DRE</code>* <code>👁️ Preview Transparent Image</code> (Checkerboard / White / Black preview)* <code>💾 Save Transparent Image</code> (Saves true 32-bit RGBA PNGs)---## ⚡ Trigger Words &amp; Prompting TipsInclude one or more of the following phrases at the beginning or end of your prompt:* *<code>RGBA transparent background, isolated foreground object</code>** (recommended full trigger)* *<code>transparent background</code>****Example Prompt:**&gt; <code>RGBA transparent background, isolated foreground object, a cute 3D glossy red toy robot holding a golden star, studio lighting, transparent background.</code>---## ⚙️ Recommended Settings in ComfyUIThis LoRA works with **both** the fast distilled/quantized **FLUX 3 Image Klein** models and the undistilled **FLUX 3 Image Klein Base** model:### Option A: Fast Distilled / Quantized <code>flux3-image-klein-bf16</code> / <code>flux3-image-klein-int8-convrot</code>)* **LoRA Strength:** <code>0.95 – 1.00</code>* **Steps:** <code>8</code> (or <code>4 – 8</code>)* **CFG:** <code>1.0</code>* **Sampler / Scheduler:** <code>euler</code> / <code>simple</code>* **Resolution:** <code>1024x1024</code> (or <code>512x512</code> – <code>1024x1024</code> via <code>EmptyFlux2LatentImage</code>)### Option B: Base Model <code>flux3-image-klein-base-bf16</code>)* **LoRA Strength:** <code>1.00</code>* **Steps:** <code>20 – 25</code>* **CFG:** <code>3.5 – 4.0</code>* **Sampler / Scheduler:** <code>euler</code> / <code>simple</code>---## 🧠 Training &amp; Technical Details* **Base Architecture:** Trained on <code>flux3-image-klein-base-bf16.safetensors</code> (5.6B, 3072 hidden dim, 5 early mode blocks + 28 joint single-stream blocks) + <code>f3_qwen3vl_4b.safetensors</code> <code>CLIPType.FLUX3</code>).* **Network Rank / Alpha:** <code>Rank 32</code> / <code>Alpha 32</code> across **203 linear layers** (62.9M trainable parameters), including <code>emb_in.image</code> <code>128 -&gt; 3072</code>), <code>content_mode_blocks.image.0..4</code>, <code>single_blocks.0..27</code>, and <code>final_layer.image.linear</code> <code>3072 -&gt; 128</code>).* **Latent Space:** Premultiplied RGBA <code>[0, 0, 0, 0]</code> background) encoded into the 128-channel packed <code>AutoencoderKLFlux2</code> RGBA latent space.* **Loss Function:*** **128-Channel Alpha-Sensitivity Weighting:** Per-channel Jacobian weighting boosting latent channels that control the decoder's 4th (Alpha) channel <code>3.0x</code> boost).* **Edge-Weighted Flow-Matching Loss:** <code>3.0x</code> spatial weight on object boundaries and semi-transparent transition pixels.* **Spatial Gradient Sharpness Loss:** Preserves crisp cutout silhouettes without dark halos.* **Dataset:** 800 curated multi-style RGBA layers from PrismLayersPro (3D renders, designer toys, anime, cartoons, watercolor paintings, steampunk assets).---## 🙏 Credits &amp; Acknowledgements* **RGBA VAE <code>AutoencoderKLFlux2</code> 4-channel):** Adapted by **Xavier Jara** ([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>)), following the [AlphaVAE (Wang et al., 2025)](<a target="_blank" rel="ugc" href="https://arxiv.org/abs/2507.09308">https://arxiv.org/abs/2507.09308</a>) method (Apache-2.0 License).* **Dataset Source:** [PrismLayersPro](<a target="_blank" rel="ugc" href="https://huggingface.co/datasets/artplus/PrismLayersPro">https://huggingface.co/datasets/artplus/PrismLayersPro</a>).* **ComfyUI Nodes &amp; FLUX 3 Klein LoRA Training:** **DenRakEiw <code>denrakeiw_nodes</code>)**.

Files with SHA256: ab85012b8f4afd61949db33c5acce6539276c443dbf6280d640be594bd448aa9