Have a ComfyUI workflow? Get a readable Workflow Doc in your browser.Make a workflow readable.Try it now ↗
Load content -> Load style -> Load models -> Prep -> Diffuse -> Save -> Optional upscale
Runtime profile
Source description
Workflow overview:
Load content -> Load style -> Load models -> Prep -> Diffuse -> Save -> Optional upscale
Load the source image you want to restyle at the top, then load the style references in the lower two image nodes. Then you're good to go. Maybe you'll want to tweak the text prompt and batch size to suit your use case.
issue:
The USO process can do significant changes to change the style and it often also takes some liberty with the composition, changing the zoom, removing items, moving arms, etc. I found about one in nine of the generations is a good match to the source image while the rest will have various
You can add a prompt to improve the chances the pose is right. For example without prompt lots of distorted poses and limbs, only one or two are good:
AI-generated commentary
AI-generated explanation based on source and configuration details. Suggestions are clearly labeled.
Style-Shift_FLUX1-uso- I2I — v1.0 is for restyling a source image with one or two style-reference images; it can use an optional text prompt and batch-size setting, save generated images, and optionally upscale and refine them 2×.
It loads content, style references, and models, then prepares, diffuses, saves, and optionally upscales the result.
Inputs are a source image, one or two style-reference images, an optional text prompt, and a batch-size setting.
Outputs are saved generated images; the optional refinement stage can add noise, upscale an image 2×, and resample it.
The workflow is listed as v1.0 and uses Flux.1 D.
The listed node packs are comfyui_fill-nodes (2.2.3), comfyui-easy-use (1.3.5), comfyui-image-compare (2.0.4), and comfyui-kjnodes (1.2.4).
Listed model files include flux1-dev-fp8.safetensors, FLUX1/ae.safetensors, FLUX1/clip_l.safetensors, and FLUX1/t5xxl_fp16.safetensors.
Additional listed model files are RealESRGAN_x2plus.pth, sigclip_vision_patch14_384.safetensors, uso-flux1-dit-lora-v1.safetensors, and uso-flux1-projector-v1.safetensors.
If using one style image, bypass the second image load and USO node.
Suggestion · not verified
Add a text prompt to improve the chances of a correct pose, and adjust the batch size for your use case.
Suggestion · not verified
For refinement, enable the final group, use choose mode to filter dud generations, and lower denoise if resampling changes more than desired.
Suggestion · not verified
Restyling may change composition, zoom, or limb positions; about one in nine generations was reported as a good match to the source image.
Before use, verify that these five files are available in your setup: flux1-dev-fp8.safetensors, FLUX1/ae.safetensors, sigclip_vision_patch14_384.safetensors, uso-flux1-dit-lora-v1.safetensors, and uso-flux1-projector-v1.safetensors.
Suggestion · not verified
Does this need to be edited?
Sign in to send an edit request.
Sources
1 sourceSource excerpts
2 excerptsSource context: 180 downloads · Type Workflows · Base model Flux.1 D
Simple prompt, still some deformities, but much better odds, more like 4 or 5 good ones:
I recommend two style images but you could bypass the second image load and uso node if you just want to use one.
Bonus upscale and refine step
The resample tends to refine details and can fix small problems if your lucky (odd fingers, transitions, etc). Enable the last group to upscale the image x2, add in a bit of noise, and then resample. Optional choose mode allows you to filter out the dud generations. You might want to play with the denoise, and reduce it if the resample is changing more than you want. Bump it up if you like what it's doing but too high and you're just starting from scratch. If ram runs out try making the resample sampler slice the image 2/2, slower but usually works.
Runtime: approx 15min on a 4060 Ti (16g) to generate a batch of 9
Estimated VRAM requirement
Estimate unavailable
26.7 GB across 6 of 8 model files. Model file total + 25% loading overhead + 2 GB execution buffer, rounded up.
Requirements
12 requirementsCheckpoint · 16.1 GB · Hugging Face · Comfy-Org/flux1-dev
Text encoder · 235 MB · SAFETENSORS · Hugging Face · diamkan/comfyui_oldmodel
Text encoder · 9.12 GB · SAFETENSORS · Hugging Face · xiaofennuonuo/comfyui
Upscaler · Hugging Face · deAPI-ai/realesrgan-x2 · Realesrgan X2
Vision model · 817 MB · Hugging Face · Comfy-Org/sigclip_vision_384
LoRA · 456 MB · Hugging Face · Comfy-Org/USO_1.0_Repackaged
Checkpoint · 20.5 MB · Hugging Face · Comfy-Org/USO_1.0_Repackaged
Node pack · Registry
Node pack · Registry
Source context: 163 downloads · Type Workflows · Base model Flux.1 D
Node pack · Registry
Node pack · Registry
