DenRakEiw/FLUX3-Image-Klein-Native-RGBA-Transparent-LoRA

✨ FLUX 3 Image Klein – Native 4-Channel RGBA Transparency Suite (Text-to-RGBA & Foreground Extraction LoRAs)

Complete native 4-channel RGBA transparency suite for FLUX 3 Image Klein (5.6B) in ComfyUI — featuring two specialized LoRAs and a 4-channel RGBA AutoencoderKLFlux2 VAE:

1. 🎨 flux3_klein_transparent_lora.safetensors (Text-to-Transparent-Image LoRA): Generate brand-new 32-bit RGBA PNG assets directly from text prompts with native alpha transparency (0.000 background alpha).
2. ✂️ flux3_klein_extract_lora.safetensors (Foreground-Extraction & Matting LoRA): Extract subjects from existing RGB images directly into clean 4-channel RGBA cutouts using FLUX 3's native ReferenceLatent (image_cond) conditioning — either automatically or guided by your text prompt!

---

📦 Files Included in This Repository

File	Target Folder in ComfyUI	Description
flux3_klein_transparent_lora.safetensors	ComfyUI/models/loras/	Part 1: Rank-32 Text-to-Transparent-RGBA LoRA for FLUX 3 Image Klein (240 MB)
flux3_klein_extract_lora.safetensors	ComfyUI/models/loras/	Part 2: Foreground-Extraction & Matting RGBA LoRA using ReferenceLatent (87 MB)
vae/flux2_3_rgba_vae.safetensors	ComfyUI/models/vae/	Original 4-Channel RGBA AutoencoderKLFlux2 VAE by trmz/flux2-klein-alpha (320 MB)
vae/flux3_ultra_rgba_vae.safetensors	ComfyUI/models/vae/	Post-Trained 4-Channel RGBA VAE: Decoder fine-tuned for cleaner transparent edges, reduced hidden-background color bleed (-55%), and higher foreground PSNR (32.7 dB -> 39.3 dB)
vae/flux3_ultra_rgb_vae.safetensors	ComfyUI/models/vae/	Experimental 3-Channel RGB Photo VAE: Pure 3-channel ae_flux_3 decoder post-trained on 2K DSLR (DIV2K) & editorial (Vogue Runway) photography with phase-coherent YCbCr luminance-wavelet distillation (~+18% subtle micro-contrast on pores/hair, 0.00% color shift)
workflows/Flux3_Klein_Transparent_RGBA.json	Drag & drop into ComfyUI	Ready-to-run ComfyUI workflow for Text-to-RGBA Generation
workflows/Flux3_Klein_Extract_RGBA.json	Drag & drop into ComfyUI	Ready-to-run ComfyUI workflow for RGBA Foreground Extraction & Matting

---

🔧 Required ComfyUI Custom Nodes

Because standard ComfyUI VAE nodes only encode/decode 3 channels (RGB), install DenRakEiw_Nodes to load, encode, and decode all 4 channels (RGBA):

cd ComfyUI/custom_nodes
git clone https://github.com/DenRakEiw/DenRakEiw_Nodes.git

Included Nodes:

• 🔧 Flux 2/3 RGBA VAE Loader *DRE – Loads flux2_3_rgba_vae.safetensors with dropdown selection and optional auto-download.
• 🔍 Flux 2/3 RGBA VAE Decode *DRE – Decodes FLUX 2 / FLUX 3 128-channel latents into rgba_image ([B, H, W, 4]), rgb_image ([B, H, W, 3]), and alpha_mask (MASK).
• 📥 Flux 2/3 RGBA VAE Encode *DRE – Encodes RGB/RGBA images into 128-channel latents (used with ReferenceLatent for Foreground Extraction).
• 👁️ Preview Transparent Image – Live Checkerboard / White / Black background preview in ComfyUI.
• 💾 Save Transparent Image – Saves true 32-bit RGBA PNGs with intact alpha channel.
• 📊 Transparent Image Info – Alpha statistics & grayscale alpha_preview output.

---

🎨 LoRA 1: Text-to-Transparent-Image (flux3_klein_transparent_lora.safetensors)

Elven Sorceress (Translucent Silk)	Steampunk Mechanical Owl	Cute Astronaut Red Panda	3D Glossy Toy Robot

⚡ Trigger Words & Prompting (Text-to-RGBA)

Include the trigger phrase at the start or end of your prompt:

• RGBA transparent background, isolated foreground object (recommended)
• transparent background

RGBA transparent background, isolated foreground object, an enchanting cute elven sorceress with pointed ears and emerald eyes, wearing a fitted white and gold silk dress with sheer translucent sleeves, glowing magical butterflies around her hands, long platinum hair, soft rim light, masterpiece character art, transparent background.

---

✂️ LoRA 2: Foreground Extraction & Matting (flux3_klein_extract_lora.safetensors)

Leverages FLUX 3's native ReferenceLatent (image_cond) stream to extract subjects from any input RGB image into a clean 4-channel RGBA PNG (workflows/Flux3_Klein_Extract_RGBA.json).

💬 Prompting Guide (Foreground Extraction)

1. Automatic Main Subject Extraction (Generic):

   Extract the main foreground subject, RGBA transparent background, isolated foreground object
   

2. Targeted Subject Isolation from Multi-Object Scenes (Prompt-Guided):

   Extract foreground object, RGBA transparent background: a steampunk mechanical owl with brass clockwork wings
   

---

⚙️ Recommended ComfyUI Settings

Works with both the fast 8-step distilled/quantized FLUX 3 Image Klein models and the undistilled Base model:

1. Fast Distilled / Quantized (flux3-image-klein-bf16 / flux3-image-klein-int8-convrot)

• LoRA Strength: 0.95 – 1.00
• Steps: 8
• CFG: 1.0
• Sampler / Scheduler: euler / simple

2. Undistilled Base Model (flux3-image-klein-base-bf16)

• LoRA Strength: 1.00
• Steps: 20
• CFG: 3.5
• Sampler / Scheduler: euler / simple

---

🧪 Included VAEs & Experimental Post-Training Notes

1. vae/flux3_ultra_rgba_vae.safetensors (Recommended for 4-Channel RGBA Generation & Extraction):
   • Post-trained on top of trmz/flux2-klein-alpha using 1,500 multi-layer RGBA matting samples (MULAN + LayerDecomp) with visibility-weighted premultiplied reconstruction, hidden-background suppression ($\alpha < 0.01 \rightarrow \text{RGB}=0$), and 2D Haar wavelet alpha-edge refinement.
   • Reduces hidden-background color bleed by -55% and improves foreground RGB PSNR from 32.74 dB to 39.29 dB while keeping the encoder 100% frozen (drop-in compatible with both RGBA LoRAs).
2. vae/flux3_ultra_rgb_vae.safetensors (Experimental 3-Channel RGB Decoder Fine-Tune):
   • An experimental 3-channel decoder post-training study inspired by UltraFlux (arXiv:2511.18050), initialized 100% from native ae_flux_3.safetensors (flux2-vae) and trained on uncompressed 2K DSLR (DIV2K) and editorial (Vogue Runway) photography using Phase-Coherent YCbCr Luminance-Wavelet distillation.
   • What to expect realistically: Because FLUX 2/3's native 32-channel (128 packed) ae_flux_3 VAE is already very strong and the encoder/latent space remains 100% frozen, visual differences compared to ae_flux_3.safetensors are subtle / minimal at normal viewing distance (~+16–20% higher Laplacian micro-contrast on 1–2 px details like skin pores, fine hair, and fabric weave when zoomed in at 100–200% in 2K/4K renders, with 0.00% color shift). Provided as an open experiment for side-by-side testing in standard ComfyUI Load VAE workflows.

---

🙏 Credits, Attribution & License

• 4-Channel RGBA VAE (vae/flux2_3_rgba_vae.safetensors):
  • Original Author & Repository: Created and trained by Xavier Jara (trmz) — trmz/flux2-klein-alpha (Hugging Face)
  • Research Paper / Technical Report: Teaching a FLUX.2 VAE to See Transparency (building on AlphaVAE, Wang et al., 2025).
  • The weight file vae/flux2_3_rgba_vae.safetensors is an unmodified mirror of trmz/flux2-klein-alpha/vae/diffusion_pytorch_model.safetensors renamed for one-click ComfyUI models/vae/ compatibility. Full credit for training the 4-channel RGBA VAE goes to trmz.
• FLUX 3 Image Klein LoRAs & ComfyUI Custom Nodes:
  • Adapted and trained for FLUX 3 Image Klein (5.6B) (flux3_klein_transparent_lora.safetensors & flux3_klein_extract_lora.safetensors) + ComfyUI custom nodes by DenRakEiw (DenRakEiw_Nodes).
• License: Apache-2.0