🍩 Donut Workflow V5
Krea2 • DonutNodes 3.0.23+
Generate, edit, inpaint and outpaint in one workflow—with organized panels, flexible references and previews for every finishing stage.
💬 Discord — help, artwork & updates · ☕ Support on Ko-fi · 🛠️ DonutNodes on GitHub
✨ V5 highlights
Cleaner layout — separate panels for setup, generation, upscaling and face detail.
Inpainting & outpainting — paint or draw a selection, invert it, adjust the seam, or expand with directional canvas presets. Outpainting stays within your megapixel budget.
Smart masks — describe what to select with SAM3.1, or use BiRefNet for foreground/background separation. Refine masks manually.
Independent A/B crops — frame each reference separately, with clipboard paste and optional reference guidance.
Flexible finishing — first and second upscales, Face Detailer and SeedVR2, with individual result previews.
Prompt variants — add, duplicate and reorder prompts; keep one active or cycle automatically. Includes autosizing editors and shared wildcard controls.
Searchable LoRAs — find models by typing, then adjust strengths and block weights.
🚀 Quick start
Install/update DonutNodes 3.0.23+. Restart ComfyUI and refresh your browser.
Open DonutWF_v5.json.
Use Manager → Install Missing Nodes, then restart.
Install the models using one of the options below.
Review your models, LoRAs, prompt and resolution. Leave Editing off for your first text-to-image run, then hit Run.
📦 GitHub installation: use Download missing for the supported model catalog.
📦 Registry/Manager installation: download the optional DonutNodes-model-installer.zip from this version’s files. Extract it, follow the included README and run the Windows or Linux/macOS launcher. It downloads catalog models into the appropriate ComfyUI folders and reuses matching existing files. Some downloads may require an access token.
The Registry package provides model links and the separate installer link instead of the built-in downloader.
🎨 Editing, masks & outpainting
Enable Editing, add your base image to A, and optionally place a second reference in B. Describe the change in the prompt.
Open Paint / outpaint… to edit a selected area or extend the canvas. Use Overlap to let generation reach into the original image and Seam width to soften the transition.
For B, choose Prompt selection to mask a named object, or Auto subject for the foreground. Use full-resolution references for better results.
Reference guidance is a separate option for borrowing visual elements during generation; it pauses while Editing is enabled.
💾 Saving & moving workflows
The supplied preset saves to output/Final with seed-based filenames. Saving supports PNG/WebP workflow metadata.
Keep your old workflow when upgrading, then copy your personal settings into V5. Reference images and wildcard files are stored separately—copy those too when moving installations.
🐛 Feedback
Found a bug? Share your workflow with private content removed, ComfyUI/DonutNodes versions, GPU/VRAM and the full error on Discord.
High-resolution editing with multiple references and NAG can be VRAM-heavy. Seam quality depends on the image, prompt and overlap settings.
Description
krea version with krea nsfw loras and uncensor patch
FAQ
Comments (4)
this workflow impressed me so much, I had to comment on it
it's not complicated, and compared to the other ones I tested so far, this one is vastly superior, specially the donut helper node (Donut Prompt Injection), it's the first time seeing it and it made it so much easier to "engineer" the prompt, and the included upscalers and face detailer are doing a really good job
Thanks, i'm really happy i figured out the right solution for upscaling/img2img for this model cause it really beats the alternatives like ZiT and Ideogram.
Anyone else get this error when changing the "Everything Else" prompt?:
Face Detailer (Max Faces) ValueError: Expected NHWC tensor, but found 5 dimensions
Any solution? Thanks
I had the same, not sure if I found the real error but I narrowed it down to the vae... I used zimage with my zimage lora that I had previously and it worked as I didn't manage to find a different vae that would work with krea without some heavy glitches
or you can alternatively just use the normal face detailer and it will take in the krea model and vae as per normal



















