ComfyUI workflows · import into your own ComfyUI
This page doesn't run ComfyUI — it hands you ready-to-import workflow graphs. Download a .json and drag it onto your ComfyUI canvas (or paste it), set your checkpoint, and run. 9 starters across 6 kinds.
txt2img
Stable Diffusion 1.5 — Text to Image (basic) txt2img 7 nodes
The canonical starter graph: load a checkpoint, encode a positive + negative prompt, KSample, decode, save. Swap in any SD1.5 checkpoint.
Models: v1-5-pruned-emaonly.safetensors (or any SD1.5 checkpoint)
Models: v1-5-pruned-emaonly.safetensors (or any SD1.5 checkpoint)
SDXL — Text to Image txt2img 7 nodes
SDXL base checkpoint at 1024×1024. Higher fidelity than SD1.5; needs more VRAM. Same shape as the basic graph with an SDXL checkpoint and a 1024 latent.
Models: sd_xl_base_1.0.safetensors
Models: sd_xl_base_1.0.safetensors
FLUX.1 [dev] — Text to Image txt2img 8 nodes
FLUX uses a separate UNET + dual CLIP/T5 text encoders + a FLUX VAE. State-of-the-art prompt following and text-in-image. Heavy: best on a good GPU. Grab the model files from the linked example page.
Models: flux1-dev.safetensors, clip_l.safetensors, t5xxl_fp16.safetensors, ae.safetensors
Models: flux1-dev.safetensors, clip_l.safetensors, t5xxl_fp16.safetensors, ae.safetensors
upscale
Upscale — Latent 2× (hires fix) upscale 9 nodes
Generate at base resolution, upscale the latent 2×, then a second low-denoise KSampler pass for crisp detail. The classic "hires fix" pipeline.
Models: any SD1.5/SDXL checkpoint
Models: any SD1.5/SDXL checkpoint
Upscale — Model (ESRGAN / 4×) upscale 4 nodes
Take an existing image and upscale it with a pixel-space upscale model (e.g. 4x-UltraSharp / RealESRGAN). No diffusion — fast, deterministic, great for finishing.
Models: 4x-UltraSharp.pth (or any ESRGAN-family upscale model)
Models: 4x-UltraSharp.pth (or any ESRGAN-family upscale model)
inpaint
Inpaint — Mask & Regenerate inpaint 8 nodes
Load an image with a mask (paint the area to change in the Load Image node), encode it for inpainting, and KSample only the masked region. Good for fixing or replacing parts of an image.
Models: any SD1.5 inpainting checkpoint
Models: any SD1.5 inpainting checkpoint
img2img
Image to Image — Restyle img2img 8 nodes
Feed an existing image into the latent and run a partial-denoise KSampler to restyle it while keeping the composition. Lower denoise = closer to the original.
Models: any SD1.5/SDXL checkpoint
Models: any SD1.5/SDXL checkpoint
video
Stable Video Diffusion — Image to Video video 6 nodes
Animate a single still into a short clip with Stable Video Diffusion. Load the SVD checkpoint, feed your image, sample the video latents, decode to frames. Heavy — needs a capable GPU.
Models: svd_xt.safetensors
Models: svd_xt.safetensors
controlnet
ControlNet — Pose / Edge Guided controlnet 10 nodes
Steer generation with a control image (pose, depth, canny edges). Load a ControlNet model, apply it to the conditioning, and the output follows your control image structure.
Models: control_v11p_sd15_openpose.pth (or canny/depth), any SD1.5 checkpoint
Models: control_v11p_sd15_openpose.pth (or canny/depth), any SD1.5 checkpoint