<p><strong>Text-to-image with a real alpha channel for MOSS-Image.</strong> End your prompt with <code>isolated on a transparent background</code>, decode with the RGBA VAE from trmz/flux2-klein-alpha, and get a transparent PNG straight from the sampler: no chroma key, no background removal, and partial transparency where it belongs (glass, fairy wings, gems, the Eiffel Tower's lattice).</p><p>⚠️ <strong>Proof of concept</strong>: trained on only 39 images in ~40 minutes on one 16 GB GPU. Not perfect (expect occasional halos or semi-transparent patches), but it shows MOSS-Image can learn alpha with very little data.</p><h3 id="required-(comfyui)">Required (ComfyUI)</h3><ol><li><p>MOSS-Image SFT + Qwen3.5-9B text encoder (<a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MOSS-Image">Comfy-Org/MOSS-Image</a>)</p></li><li><p>RGBA VAE: <code>vae/diffusion_pytorch_model.safetensors</code> from <a target="_blank" rel="ugc" href="https://huggingface.co/trmz/flux2-klein-alpha">trmz/flux2-klein-alpha</a> (Apache-2.0)</p></li><li><p>The <strong>Load RGBA VAE (Flux2)</strong> custom node and the workflow, from <a target="_blank" rel="ugc" href="https://huggingface.co/paralaif/MOSS-Image-Alpha-LoRA">paralaif/MOSS-Image-Alpha-LoRA</a> (folder <code>comfyui/</code>). Stock ComfyUI can't load a 4-channel VAE.</p></li><li><p>Optional: an 8-step DMD LoRA for MOSS-Image SFT (e.g. <a target="_blank" rel="ugc" href="https://huggingface.co/lodestones/MOSS-Image-SFT-8-steps-dmd">lodestones/MOSS-Image-SFT-8-steps-dmd</a>)</p></li></ol><h3 id="tips">Tips</h3><ul><li><p>Always keep <code>isolated on a transparent background</code>. Without it, prompts that imply a place (e.g. "the Eiffel Tower") get a full scene and the alpha breaks.</p></li><li><p>LoRA strength 1.0 (0.8 is fine; at 0.6 subjects turn see-through). With DMD: 8 steps, CFG 1, er_sde/simple. Without DMD: 28 steps, CFG 4.</p></li><li><p>Preview images here are flattened onto backgrounds because Civitai doesn't show transparency. Raw RGBA PNGs are on Hugging Face.</p></li></ul><h3 id="how-it-was-made">How it was made</h3><p>The RGBA VAE shares FLUX.2's latent space, so it can decode MOSS latents, but plain MOSS latents give garbage alpha. We built a small dataset with MOSS as the image source and alpha "teachers" (Klein Base 4B + Extract-4B on composite/plate pairs, Qwen Image 2.1 for effects), filtered it automatically and trained a rank-32 LoRA at LR 3e-4 with ai-toolkit. Full research notes, training config and the ai-toolkit patch: <a target="_blank" rel="ugc" href="https://huggingface.co/paralaif/MOSS-Image-Alpha-LoRA">huggingface.co/paralaif/MOSS-Image-Alpha-LoRA</a></p><h3 id="licence">Licence</h3><p>Non-commercial research and evaluation use only. <strong>Built with Qwen</strong>: 8 of the 39 training images were generated with Qwen Image 2.1 (Qwen Research License). Credits: MOSS-Image (OpenMOSS, Apache-2.0), RGBA VAE and Extract-4B (Xavier Jara / trmz, Apache-2.0), FLUX.2 Klein Base 4B (Black Forest Labs, Apache-2.0), ai-toolkit (Ostris). All characters are original designs.</p>

MOSS-Image Alpha LoRA - transparent PNG (proof of concept)

<p>Proof of concept, checkpoint at step 1000 of 2000 (later steps overfit). 39 training images, rank 32, LR 3e-4, AdamW 8-bit.</p><p><strong>Use:</strong> end the prompt with <code>isolated on a transparent background</code> and decode with the RGBA VAE (trmz/flux2-klein-alpha) via the "Load RGBA VAE (Flux2)" node from <a target="_blank" rel="ugc" href="https://huggingface.co/paralaif/MOSS-Image-Alpha-LoRA">huggingface.co/paralaif/MOSS-Image-Alpha-LoRA</a>. With an 8-step DMD LoRA: 8 steps, CFG 1, er_sde/simple. Without DMD: 28 steps, CFG 4.</p>

Files with SHA256: 63e27e0cecf940babab28deebe74673e99bb3f18db281336f9e0e468c9684995