Qwen and Krea 2 + SAM3 Face Inpainting Workflow — Krea2_T2I_V1.1

This ComfyUI workflow is for generating portrait images from text prompts; it can take one prompt or scheduled prompts, automatically find and enhance faces, and output images with those edits blended back in.

It generates an image, uses SAM3 Smart Inpainter to find and enhance a woman’s and a man’s face, then blends both edits into the main image.

Batch mode runs through scheduled prompts, and AutoResetInt resets to 0 after each generation or cancellation.

Single Prompt mode takes one text prompt; Batch Prompts mode takes prompts keyed to frame positions such as “0” and “30”.

In Batch Prompts mode, max_frames controls how many items the workflow generates.

The SAM prompt can target a hat, logo, or hands instead of a face.

Install the node packs in ComfyUI/custom_nodes/, install their Python requirements, place the model files in the required folders, restart ComfyUI, and load the workflow JSON.

Required node packs include comfyui_fizznodes, comfyui_nvidia_rtx_nodes, comfyui-impact-pack, and sam3_smartinpainter.

Required node packs also include rgthree-comfy, comfyui-workflow-encrypt, comfyui-custom-scripts, and ComfyUI-Chibi-Nodes; model files include qwen3_vl_4b_fp8_abliterated.safetensors, sam3.pt, qwen_image_vae.safetensors, krea2_raw_fp8_scaled.safetensors, and Krea2_Turbo_FP8_T​​​​

Set the mode toggle to true for Single Prompt or false for Batch Prompts.

The suggested settings are min_mask_area 0.001, crop_padding 32–128, inpaint_resolution 1280, denoise 0.3–0.5, steps 6, and cfg 1.0.

The model-format toggle offers GGUF for lower-VRAM or quantized setups and FP8/BF16 for powerful GPUs.

CacheDiT is mentioned for acceleration, while CacheDiT Accelerator is listed as removed; verify its availability before relying on it.

Suggestion · not verified

Try one image before a large batch, and verify parameter settings if you switch models because they may need adjustment.

Suggestion · not verified