pre_USO
Deprecated preset for ByteDance's USO style/subject model on Flux
- context
- image
- mask
- ref_image
- context
- model
- positive
- latent
Filed under the pack's Deprecated category, so this is a "here's what it wired together" reference more than a recommendation for new work. If you want USO on Flux today, you're probably better off going straight to ByteDance's own USO repo and its native ComfyUI support rather than this preset wrapper - native support landed directly in ComfyUI not long after USO's release, which is unusually fast and is presumably part of why a third-party pack's own preset around it stopped being maintained.
What USO is
USO is ByteDance's subject-identity-plus-style model for Flux - feed it a reference image and it can preserve who or what's in it while transferring a style onto the result, and it can also do pure style transfer or mix two style references together into a new blended look. Community reaction described it as combining what InstantID, IP-Adapter, DreamO and UNO each do separately into one model, and specifically called it stronger than InstantCharacter at holding character identity. This preset node bundles the pieces USO needs - a vision encoder, a projector "model patch," and the reference-handling logic - into one node instead of you assembling them from the raw USO nodes by hand.
The inputs and outputs that matter
context/image(required) - the run-context bundle and the base image you're conditioning on.reference_latents_method(required, one ofuxo/uno,offset,index) - which mechanism USO uses to fold the reference into the latent stream. If you don't know which to pick,uxo/unois the name closest to USO's own lineage (it explicitly builds on the UNO subject-consistency work) and is the reasonable default to try first.guidance(default 3.5, range 0ā100) - this is Flux's own guidance-distillation value, not classifier-free guidance in the classic sense; 3.5 is the standard Flux starting point and most people don't stray far from it.crop(centerornone) - how the reference image gets fit to the model's expected input size.clip_vision(defaultsigclip_vision_patch14_384.safetensors) andmodel_patch(defaultuso-flux1-projector-v1.safetensors) - the two model files USO actually needs. Neither ships with this node; you download them separately (see below).smoothness(default 0, range 0ā10) - note the schema lists this as an integer with a 0.1 step, which is an oddity in the node's own definition rather than something you're misreading; in practice expect it to behave as a small integer knob for how sharply the style/subject blend transitions.mask/ref_image(optional) - a mask to constrain where the effect applies, and a second reference image for style mixing between two references.- Outputs:
contextpassthrough,model(patched for USO),positive(CONDITIONING), andlatent.
How to install it (for archival purposes)
Search ComfyUI-Apt_Preset in ComfyUI Manager, or:
cd ComfyUI/custom_nodes
git clone https://github.com/cardenluo/ComfyUI-Apt_Preset.git
install.bat handles Windows dependencies; on Linux or Mac, read it and pip-install the listed packages yourself. The two model files this node needs aren't bundled - sigclip_vision_patch14_384.safetensors and uso-flux1-projector-v1.safetensors need to be downloaded from ByteDance's USO release and placed wherever your ComfyUI install expects CLIP-vision and style-patch models respectively; check the enum's empty choice list in your own install (it showed zero available options in ours) as a sign neither file is present yet.
Common issues & troubleshooting
clip_vision or model_patch dropdowns are empty. That means the files aren't in the expected folder. Download them from the official USO release and restart ComfyUI - these dropdowns populate from disk at startup, not on demand.
Style mixing between two references looks like a messy midpoint rather than either style. That's reportedly by design, not a bug: community testing of USO's mixing feature describes the result as landing at a "middle point" between two input styles rather than favoring either one - if you want closer fidelity to a single style, use ref_image for subject reference only and skip the mixing use case.
The node is gone after updating the pack. Expected, given its Deprecated status - pin an older commit if a saved workflow depends on it, or move to the native USO nodes instead.
Inputs (10)
| Name | Type | Default | Description |
|---|---|---|---|
| context | RUN_CONTEXT | ā | |
| image | IMAGE | ā | |
| reference_latents_method | COMBO | 3 options: uxo/uno, offset, index | |
| guidance | FLOAT | 3.50ā100 | ā |
| smoothness | INT | 00ā10 | ā |
| crop | COMBO | 2 options: center, none | |
| clip_vision | COMBO | sigclip_vision_patch14_384.safetensors | 0 options: |
| model_patch | COMBO | uso-flux1-projector-v1.safetensors | 0 options: |
| maskopt | MASK | ā | |
| ref_imageopt | IMAGE | ā |
Outputs (4)
| Name | Type | Description |
|---|---|---|
| context | RUN_CONTEXT | ā |
| model | MODEL | ā |
| positive | CONDITIONING | ā |
| latent | LATENT | ā |