FLUX + Hyper Lora 8 Steps + LLM + Detailing + Inpainting + Upscaling + Low Vram

Made with Hyper Flux 8steps and Flux Dev Q4_0.GGUF. The main goal is to use FLUX with 8GB VRAM (my own configuration).

Multiple functions in the workflow are controlled by a switch in Group 0 (the black one).

1- Inpainting: it offers three options: (SAM Detector + Custom Mask or Only Custom Mask or Mask Uploaded)

2- Prompt Generation: A prompt can be generated either through text or an image via a LLM.

3- 1st Detailing: it is performed with FLUX.

4- Upscaling: it is done using the Tile Diffusion Node, SDXL Lightning, and CN SDXL Tile. It operates quickly and produces stunning results.

5- 2nd Detailing: a second pass is done with SDXL/FLUX for the background and the main subject.

6- Post-Processing: a final step of post-processing is done to adjust brightness, contrast, etc.

Make sure to download the exact models shown in the workflow. I selected them specifically to work with my 8GB VRAM card.

Models:

- flux1-dev-Q4_0.gguf

- Hyper-FLUX.1-dev-8steps-lora.safetensors

- t5xxl_fp8_e4m3fn.safetensors

- 4xLeexicaDat2_otf.pth

- 4x-UltraSharp.pth

- 4xFaceUpDAT.pth

- dreamshaperXL_lightningDPMSDE.safetensors

- ttplanetSDXLControlnet_v20Fp16.safetensors

Please, note the package required for Ollama is ComfyUI Ollama by stavsap.

ComfyUi-Ollama-YN is not required and create conflict with the Ollama nodes.