If you want help or share artwork please join the discord server where i also update everyone on new tools and versions: https://discord.gg/5djEeQ3wpU
My kofi (any amount helps with AI bills): https://ko-fi.com/donutsdelivery
# Donut Workflow V4 Beta
A Krea2 workflow built around DonutNodes, with a redesigned interface for model setup, LoRAs, editing, prompts, generation and saving. DonutNodes supplies the controls and processing nodes; this workflow brings them together.
Workflow file: DonutWF_v4_beta.json · DonutNodes on GitHub
## Quick start
1. Install or update DonutNodes to the code accompanying this beta. Restart ComfyUI and refresh the browser after updating. An older registry build that lacks DonutImageSave or DonutEditStudio cannot run this workflow.
2. Download DonutWF_v4_beta.json from this workflow’s download files. Drag it into ComfyUI or use Open. If GitHub shows a file preview, use Download raw file.
3. Choose Install Missing Nodes → Install All, accept the default pack versions, and Apply Changes/restart. No WAS package is required.
4. Select your models in 01 · Models, then use Download missing for files supported by Donut's model catalog. Install any reported uncatalogued files separately, or select models you already have.
5. Review the enabled LoRAs, prompts, seed, resolution and output settings. Leave Editing off for a first text-to-image run, then click Run.
The beta workflow must be distributed with the updated node code. The local installation test used the development checkout, not the older registry release. Workflow V4 Beta and the DonutNodes package version are separate version numbers.
### Companion packs
Manager detects these from the workflow, including the unconnected nodes inside Required node packs. Keep that subgraph: it makes internally used dependencies visible to the missing-node installer.
- ComfyUI-bleh: Sampler preset
- ComfyUI Impact Pack: Detection, SAM and face-detail processing
- ComfyUI Impact Subpack: Ultralytics detector provider
- Derfuu ComfyUI ModdedNodes: Text boxes
- ComfyUI Krea2 Edit: Identity-edit integration
- Krea2 NAG: Negative attention guidance
- Krea Seed Variance Enhancer: Seed-dependent conditioning variance
WAS and rgthree's label nodes are no longer needed by this workflow. Existing installations may keep those packs for other workflows; updating Donut does not uninstall them.
### Models and downloads
The supplied configuration selects the following files. These are the saved choices, not a promise that every file is needed in every mode or included in this repository.
- Primary diffusion model: finepornV4INT8NVFP4BF16_v4.safetensors
- Secondary diffusion model: krea2_turbo_bf16.safetensors
- Text encoder: qwen3vl_4b_fp8_scaled.safetensors
- VAE: qwen-image/qwen_image_vae.safetensors
- Upscaler: 4x_NickelbackFS_72000_G.pth
- Face detector: bbox/face_yolov8m.pt
- SAM: sam_vit_b_01ec64.pth
- Enabled generation LoRA: krea2/Krea2_NSFW_Aesthetics_V1.safetensors
- Edit LoRA: krea2/krea2_identity_edit_v1_2.safetensors
Use ComfyUI's corresponding model folders. Single-model mode does not load the secondary model. Editing uses the identity-edit LoRA; editing off requires no reference images. Review the saved model/LoRA choices before distributing your own preset.
Download missing only downloads after a click. It uses the repository's reviewed model catalog model_sources.json), checks file size and SHA-256, and can reuse matching renamed files. It reports missing catalog entries rather than guessing download links. Some hosts require your own access credentials: Civitai uses Donut's local API-key setting; Hugging Face can use HF_TOKEN.
## What’s in the workflow
Numbered cards expose everyday controls, with additional controls under Advanced. The source loaders and generation wiring live inside inspectable subgraphs. Graph and App Mode use the same underlying settings.
- Models: choose a single model or two-model merge, encoder, VAE and upscaler.
- LoRAs & block weights: add, remove, reorder and enable LoRAs; edit strengths and block weights through sliders or numeric fields.
- Image setup & editing: set the output size and optional editing references.
- Prompts: edit the general/face, scene and negative text, with wildcard tools.
- Seed & guidance: control the shared seed, NAG and seed variance.
- Generate & finish: adjust sampling, first/second upscale and face detail.
- Latest result: choose a stage preview or follow the latest output; inspect the final expanded prompt and stage progress.
- Save images: choose the destination, format, quality and filename behavior.
## Editing and reference guidance
For editing, enable Editing, upload/paste/drop image A as the base scene, and optionally use image B for the subject/identity. Enter the edit instruction and adjust the crop and output sizing. With B connected, face identity comes from B. Image A alone remains supported.
Reference guidance is a separate optional path for borrowing visual elements through Krea2's native image conditioning. Describe what to borrow in the prompt. It does not use the edit LoRA and pauses while Editing is enabled.
Save after changing references or crops. Reference images are stored separately under ComfyUI/user/donut/edit_references/; copy that folder too when moving a personal workflow to another installation. The distributed beta has empty reference slots.
## Wildcards
Use the Wildcard library card to create or edit one-choice-per-line text files. Insert haircolor* to use user/wildcards/haircolor.txt; nested names such as clothes/shirt* and __haircolor__ syntax also work. The shared seed drives expansion, so a fixed seed keeps choices repeatable. Copy your wildcard files when moving installations; they are not embedded in the workflow JSON.
## Saving images
Donut Image Save replaces WAS's save node. The supplied settings preserve output/Final, the seed-based filename, WebP quality 100 and numbered saves with overwrite disabled. The secondary resized output remains a core ComfyUI save.
The Donut saver supports 8-bit PNG, JPEG, WebP, TIFF, GIF and BMP; output/temp locations; delimiter and number placement; lossless WebP; and optional PNG/WebP workflow metadata. It does not include WAS's history browser, color-profile socket or high-bit-depth/EXR modes. Those features were unused in the migrated configuration.
## Moving from V3
Open V4 Beta as a separate workflow. Keep your V3 JSON and copy your own model choices, LoRAs, prompts and settings through the visible controls. Do not copy raw widget arrays between versions: the layout and inputs changed. Re-select references in Edit Studio and copy any personal wildcard/reference files.
V3 already supported editing, face detail, model merging, LoRA stacking and upscaling. V4 Beta reorganizes and extends that workflow; it does not introduce face detailing or promise better image quality simply from the version change. See the accompanying CHANGELOG.md for the actual differences.
## Beta validation and reporting
A fresh Linux/Python 3.12.7 venv with the current DonutNodes source passed default Install All → restart → full generation, without WAS or a manual package repair. Base generation, first upscale, two face refinements and WebP saving completed on an RTX 4070. The saved 1728 × 1344 image matched the previous WAS save byte-for-byte with the same configuration. Six saver tests and 34 registration/dependency-isolation tests passed.
That acceptance run used editing off and the second upscale disabled, and reused model files. It does not certify every configuration, model downloads, other GPUs, or Windows/macOS installation. Companion requirements still install two OpenCV variants in the tested environment; a warning remains, but imports and inference passed. No fixed minimum VRAM requirement has been established.
When reporting a problem, include your workflow file with private content removed, DonutNodes/ComfyUI versions, OS, GPU/VRAM, active mode and full error traceback. Donut Dependency Check can provide dependency diagnostics.
Detailed test notes accompany the node source in docs/validation/no-was-fresh-install-2026-09-09.md.
ComfyUI/
└── models/
├── diffusion_models/
│ ├── krea2_turbo_bf16.safetensors
│ └── finepornV4INT8NVFP4BF16_v4.safetensors
│
├── text_encoders/
│ └── qwen3vl_4b_fp8_scaled.safetensors
│
├── vae/
│ └── qwen-image/
│ └── qwen_image_vae.safetensors
│
├── loras/
│ └── krea2/
│ ├── krea2_identity_edit_v1_2.safetensors
│ └── Krea2_NSFW_Aesthetics_V1.safetensors
│
├── upscale_models/
│ └── 4x_NickelbackFS_72000_G.pth
│
├── sams/
│ └── sam_vit_b_01ec64.pth
│
└── ultralytics/
└── bbox/
└── face_yolov8m.pt
Description
compatibility (mostly)
FAQ
Comments (39)
Great WF! How do you change the resolution of what you want the final image to be? It's always stuck at 208x304.
in the SDXL latent empty image node you can change the aspect ratio and you can change upscaled resolution by changing rescale_factor in "tiled donut upscale".
@DonutsDelivery it still outputs at a low resolution. I want the end result to be at 1080x1920, and no matter what I change in the SDXL empty image node, it comes out at a much, much lower resolution. Changing rescale_factor doesn't change anything either.
@imo09 you do know that it saves the upscaled img to a subfolder right? what you are looking at is a small png with metadata it saves for you as a backup.
@DonutsDelivery Apologies, I forgot about the non SDXL workflow, and everything works great now, lol. Great workflow once again!
Thx for the great WF. I have some questions:
- What exactly are the 3 rows of DonatLoraStack Nodes? Are they in a particularly order? Or can i just load Loras in any of them? What is "Block Preset 1" Configuration?
- There is a Face Related only and Everything else prompt box. What if I only have a normal prompt? Can I leave face empty? What exactly do the single boxes they influence exactly?
- Why is there a deactivated CFG Tone Modifier Node? What happens when I activate it?
1. DonutLoraStack allows you to use 3 loras. ive stacked multiple of them so you can load up to 9 loras in the WF. Ive made the most right of the stacks only apply to the refiner (so it first kicks in after the first 5 steps (step-5-9). The Presets are for block weighted lora loading. It is a method to only apply the loras to some parts of a Model. The blocks that are first mess with composition/overall structure and the last blocks mess with finer detail, and the middle is everything inbetween. There is not a clearcut definition of what each block does but you can try to make the best of Loras by disabling less desirable features this way. For example if a lora is prone to making your image more or less saturated, you can disable the blocks in charge of that. use the preset where everyhing is set to 1 if you want it to work as a normal lora loader.
2. The everything else prompt is excluded from the Face Detailer prompt. The face detailer is working on just the face alone so there is no reason to included the full prompt for that. If you don't have a subject/face in your image, you can put the prompt in either node.
3. You can use the CFG tone modifier to set a higher CFG and then using this node to making the image less "burnt". The way I use it is to just enable it and put the CFG to 2. If you leave it at 100 its the same as it were off, then you lower it to tame your CFG so your output dosent look burned. It is non-conventional.
Gotta say this is prolly my favorite workflow right now, perfect for decently consistent faces!
Appreciate your work!
I love your workflow, but I’m having an issue. I have a LoRA trained on a specific face, and this workflow tends to change the face. What would you recommend I do in this workflow to preserve just the facial appearance?
all you gotta do is link the loraapplstack model link to the face detailer.
Do you have any plans to include controlnet with this type of workflow? Prolly the last thing it needs to be perfect in my opinion!
I make these for personal use and share, so if you can convince me to try controlnet then ill make it
Reading your other comments I cannot figure out where would I apply my lora for my character face. Do I add it to one of the doughnut stacks you have, or do I add a new lora loader for just my character? I see above the face detailer you have a SDXL lora which would not work with my Zimage lora. Awesome workflow btw, works great out of the box when you have the packages.
link donutapplylorastack to facedetailer (the first of the details which uses the ZIT model, not the last 2 which uses the SDXL model)
@DonutsDelivery so link the first donutapplylorastack in your workflow, or use one the the last two, or add a brand new one and link it? :)
@DirtyBeagles whichever one is linked to the loader that has your face lora loaded :)
Where can I find the two bbox .pt files?
i want know too, you find it?
ive uploaded it to resources in my discord server
ask chatgpt, it gives you the links.
@Yedust Yeah, in DonutsDelivery's discord
Nudifiers and noncon are not allowed here, and you made nude taylor swift with breasts and posted her on civitai! (image 113173710)
I think you're lost, that's an image of Loras that are on this site.
I get an error in the "BlehSetSamplerPreset" Node - he's missing the "any_input" - input. The workflow creates the first image, but doesn't continue over the right 3 detailler nodes. any advice?
also the faces turn out not too similar - also haircolors differ. what can i do?
also: could i modify the workflow to generate single images (not side by sides)?
This workflow has a node that will break your comfyui and you won't be able to run it.... careful
and that is?
Which node? Cuz i am using this workflow without a problem
I had an issue with donuts nodes before, but I think its all fixed with new version of confyui, make sure you update or your going to have a bad time.
Sometimes the manager can install conflicting dependencies depending on which custom nodes you got installed. Custom nodes don't break confyui on their own.
@DonutsDelivery Exactly what was going on, I can't rememeber the name of the node. but it was a conflict. never had the issue after last update. Or maybe I got rid of the conflicting node. In any case, the suggestion that the node itself breaks comfyui is incorrect at best.
@EldraVisionsExtreme idk ive had it happen many times. the easy solution is just to remove delete all custom node folders expect comfyui, delete your venv and make it again, the run "pip install requirements.txt" in your comfy directory then run comfy and install missing nodes with manager.
Possibly the conflict is because Manager is recommending installation of the retired WAS Node pack instead of the Revised version. This causes a break in the loading of nodes. Removing the conflicting WAS Node folder and installing the Revised pack fixes it.
@MrHara when conflicting dependencies has been installed in your venv, comfyui wont start, and its hard for most people to activate the venv and uninstall the conflicting packages without other things breaking. so the easiest solution is to recreate your venv and install from start. there are many nodes which has had issues in the past.
not true, just a skill issue on your part. educate yourself before commenting next time
How to change to i2i
load an image with an image load node, vae encode it, and the pass it into the first ksampler, and change the denoise from 1 to anything else.
can you make upscaled with controlnet tile union?
Thanks for sharing this, any chance you could add a note to the workflow with links to the LoRAs and other models?
ive attached most of them as suggested resoruces on my latest checkpoint.