Krea 2 GGUF in ComfyUI: Getting fried psychedelic colors & green grid artifacts instead of clean images. Any working setup?
retained text
Hi everyone, I've been trying to run Krea 2 in ComfyUI using the GGUF quantizations, but I can't seem to get a clean generation. Depending on the settings, I get either a heavy green grid/checkerboard texture or deep-fried, iridescent psychedelic artifacts with blown-out colors (see attached images). Here is my current environment and workflow setup: \*\*System:\*\* \* GPU: NVIDIA RTX 4070 (12 GB VRAM) \* OS: Windows 11 \* ComfyUI: Latest portable build \*\*Nodes & Models:\*\* \* \*\*Loader:\*\* \`UnetLoaderGGUF\` (using molbal's maintained fork of \`ComfyUI-GGUF\`) \* \*\*Diffusion Model:\*\* \`krea2\_turbo-Q4\_K\_M.gguf\` (and tested with \`krea2\_turbo-Q4\_K.gguf\`) \* \*\*Text Encoder / CLIP:\*\* \`qwen3vl\_4b\_fp8\_scaled.safetensors\` with CLIPLoader set to type \`krea2\` \* \*\*VAE:\*\* Tested both \`wan\_2.1\_vae.safetensors\` and \`Wan2\_1\_VAE\_fp32.safetensors\` (as well as \`qwen\_image\_vae.safetensors\`) \* \*\*Latent:\*\* \`EmptySD3LatentImage\` (16-channel) set to \~1024x1024 or 1920x1280 (also tried Krea2 regional latent) \* \*\*Conditioning:\*\* Positive prompt encoded via CLIP, negative hooked into \`ConditioningZeroOut\` (CFG-free style) \*\*Sampler Settings:\*\* \* Steps: 8 to 9 \* CFG: 1.0 \* Samplers tested: \`er\_sde\`, \`euler\`, \`euler\_ancestral\_cfg\_pp\` \* Schedulers tested: \`simple\`, \`beta\` \*\*What happens:\*\* 1. In some configurations, the image produces a repeating knitted/green mesh grid pattern across the whole canvas. 2. When switching samplers/schedulers, shapes and text fragments start to appear, but the entire image is solarized with intense chromatic aberration, acid colors, and repeating tiling artifacts. Has anyone managed to get a completely clean output with Krea 2 in ComfyUI? Are there specific latent scaling factors, exact VAE models, or sampler/scheduler combinations required for Krea 2's DiT architecture? Any insights or working workflow JSONs would be greatly appreciated! Thanks in advance!
document fetched 2026-09-08 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1w9yvi8/krea_2_gguf_in_comfyui_getting_fried_psychedelic