⚡ Flux Klein High-Res Powerhouse (4K & 8K Workflows)
Unlock the full potential of Flux Klein. These workflows allow you to generate incredible high-resolution images, reaching 4K and even 45MP (8K) resolutions, utilizing the efficiency of Klein models.
I have included two versions of this workflow to suit your specific hardware and quality needs.
📂 The Two Versions
1. 🚀 Performance Hybrid (Speed)
Filename: Flux_Klein_9b_with4b_upscaler.json
Best for: Faster generation times while still maintaining excellent resolution.
How it works:
Stage 1: Generates the base composition using the Klein Base 9b model.
Stage 2: Upscales the image (2x or 4x) using the lightweight Klein 4b (4-step) model.
2. 💎 Ultimate Fidelity (Quality)
Filename: Flux_Klein_9b_with_9b_Upscaler.json
Best for: Maximum detail, texture consistency, and highest quality output.
How it works:
Stage 1: Generates the base composition using the Klein Base 9b model.
Stage 2: Upscales using the more robust Klein 9b (4-step) model for superior refinement during the upscale process.
✨ Key Features
Extreme Resolution: Capable of generating standard images and upscaling them to 4K or 8K (45MP).
Smart Tiled Upscaling: Uses a "Divide and Conquer" strategy combined with Florence2. The workflow automatically captions the image before upscaling to guide the diffusion model, ensuring the upscale respects the context of the image.
Efficient: By utilizing the Klein quantized models, you get Flux capabilities with a much friendlier VRAM footprint.
🛠️ Requirements & Setup
To use these workflows, you will need the following models in your ComfyUI models folders:
Checkpoints (UNET):
flux-2-klein-base-9b.safetensors
flux-2-klein-4b.safetensors (for the speed version)
Upscale Model:
4xNomosUniDAT_otf.pth (or your preferred 4x upscaler)
Nodes: Ensure you have the necessary custom nodes installed (specifically ComfyUI-Florence2, ComfyUI_Steudio (for Divide & Conquer), and ComfyUI-GGUF if using GGUF versions).
Description
FAQ
Comments (10)
why is the workflow using base? cuz they produce lower quality overall, possible better flexibility, but thats it.
现在做图片放大的话流行用哪个模型呢
Difficult question. In practice, it mostly needs more denoising, but to render with higher denoise you first need a tiled or upscaling ControlNet.
Both text encoders: qwen_3_8b_fp4mixed.safetensors and fp8 seems not be working, only the full qwen_3_8b model. Can u help me?
I'm trying to get the first image in 1920x1080 but I only can get x1072 or x1088. How do I fix it? ><
Also, I'm trying to make the final image(upscaled) to 4k(3840x2160) but I'm messing with something, my images are 2700x4835 e.e
@TheKnightsWhoSayNI x1072 and x1088 are both divisible by 16, 1080 is not. So it makes sense, and under the hood 1080 is altered to 1072 anyway.
@Spoonman2002 That was a clear and simple answer. Thank you mate
The landscape resolution is hardwired into this workflow and does not adjust automatically when the source resolution is anything but.