ACE++ and UNO integration

[attashe]

Summary

Type of change: Feature

Redux is very powerful instrument for Flux generattion but have sometimes it's hard to find a balance with precision/flexibility/prompt cohere. Novel methods can outperform Redux on some tasks.

https://ali-vilab.github.io/ACE_plus_page/
https://bytedance.github.io/UNO/

What's Done:

1. ACE++ Integration:

   • Added image concatenation nodes (for left/right/top/bottom stitching; requires images of the same size).
   • Implemented generation using a Reference Image and Mask Editing.
   • Maintained backward compatibility with existing pipelines.

   ACE++ node:
   

Output width/height/pad settings are used to crop the result from the concatenated image.

Image Concatenation node:

Generation with reference example + Redux:

Mask editing with reference:

2. UNO Implementation:
   • Added support for up to 4 reference images via latent space concatenation.
   • Implemented Universal Positional Embeddings (UnoPE) using Rotary Position Embedding adaptation.
   • Added a new input port to the "FLUX denoise" node for reference sources.

UNO reference images loading node:

Results:

Only first image slot is require others are optional

Discussions

ACE++: Shows good compatibility. Additional inpainting nodes could be beneficial, for example, for crop-mask inpainting at higher resolutions. A good example is this ComfyUI node collection:
https://github.com/lquesada/ComfyUI-Inpaint-CropAndStitch

UNO: Compatibility is limited, and it can be slightly unstable at higher resolutions (> 512-768). Generation with UNO requires a special LoRA. In the original repository, it is saved in the Xlabs-lora format, which is not compatible with other frameworks. I converted it to the diffusers LoRA format.

• Original: https://huggingface.co/bytedance-research/UNO/tree/main - dit_lora.safetensors
• Converted: https://huggingface.co/attashe/uno_converted/tree/main - uno_lora_diffusers.safetensors

• ACE++: Processing takes 1.5-1.7 times longer because the input image tensor expands 1.5-2 times.
• UNO: Processing time increases by 50%-100%, depending on the input size and the number of reference images.

Merge Plan

Patch adds new nodes that doesn't interact with others. UNO encoding add some code into backend but they have no effects if no referece image was provided

Checklist

• ACE++ integration
• UNO integration
• New nodes

Workflows:

UNOExperimentsWorkflow.json
InpaintEditingACEpp.json
ReduxGenerationACEpp.json

[hipsterusername]

@attashe - I'm just now getting a chance to test both of these. ACE++ is working well, but I'm unable to get Uno to work. Can you share some examples of how the workflow should be configured to produce meaningful results?

[psychedelicious]

ACE++ is working well for me - pretty cool! Uno does something but I'm having a hard time getting it to do something useful... That said, I've only done some quick testing.

I did some minor cleanup:

• Format with ruff (make ruff from repo root)
• Update the Uno node and field to accept a list of images

[psychedelicious]

Note: It's not necessary to provide latents to the denoise node, unless we are doing partial denoising (i.e. denoise_start > 0).

[attashe]

@hipsterusername
This is workflow that works for me. Important to load uno_lora_diffusers.safetensors not converted
With multiple references it works better with smaller resolution - 768, 512.
test_uno.json