Redditor trains Krea2 to skip VAE Decode and output pixels directly using DWD paper
Remote_Net374 · reddit · 2026-11-22
A Reddit user applied the Decoding Without a Decoder (DWD) paper to Krea2 so inference skips VAE Decode entirely: the original denoising head still samples, while an added pixel head produces the image at the final step.
- Trained overnight on a single RTX 3090, only the pixel head and last transformer block (memory-limited setup)
- Early outputs had harsh grids and unrecognizable textures; images are now recognizable but grid artifacts and texture issues remain
- A ComfyUI workflow screenshot shows images generated without VAE decoding
- The author plans to release a ComfyUI node and patch after fixing the grid issue
More from Multimodal
- Dev uses Claude with his game engine plus fal, Tripo, Meshy and ElevenLabs to make a movie trailer — gorkem · 2026-11-22
- A 100% AI monster movie scene made with Seedance 2.5, plus Dreamina unlimited offer — JaynitMakwana · 2026-11-22
- Splatting-based real-time rendering in Rust/Bevy hints at the end of traditional 3D game pipelines — taiyasaki · 2026-11-22
- Running Minimax H3 locally on a 3090: how to seamlessly continue video segments — fire_abyss_guardian · 2026-11-22
- Steppe-Wan2.1: open-source latent upscaler for Wan/Qwen/Krea, arch found by 120 Claude-run experiments — TensorForger · 2026-11-22
- Walking a snow ridge in Genie 4, then rendering the same mountain four ways with MiniMax H3 in ComfyUI — PurzBeats · 2026-11-22