lllyasviel / lllyasviel/stable-diffusion-webui-forge

Feature Request: Add Normalized Attention Guidance (NAG) and Token Perturbation Guidance (TPG) as built-in extensions

Open
#3,029 0 comments 0 reactions 0 assignees View on GitHub
Dominant language
Python
Stars
13k
Forks
1.7k
PR merge metrics
No merged PRs in 30d

Description

### Description

It would be great to integrate two recent guidance methods directly into Forge:

* **TPG (Token Perturbation Guidance)** – [[Paper](https://arxiv.org/abs/2506.10036)](https://arxiv.org/abs/2506.10036) | [[Code](https://github.com/TaatiTeam/Token-Perturbation-Guidance)](https://github.com/TaatiTeam/Token-Perturbation-Guidance)
* **NAG (Normalized Attention Guidance)** – [[Paper](https://arxiv.org/abs/2505.21179)](https://arxiv.org/abs/2505.21179) | [[Code](https://github.com/ChenDarYen/Normalized-Attention-Guidance)](https://github.com/ChenDarYen/Normalized-Attention-Guidance)

**Why?**

* Complements existing built-in guidances (SAG, PAG).
* Improves prompt adherence (TPG) and stability/semantic faithfulness (NAG).
* Already tested in the community: see [[sd-perturbed-attention extension](https://github.com/pamparamm/sd-perturbed-attention)](https://github.com/pamparamm/sd-perturbed-attention) and [[this PR for NAG](https://github.com/pamparamm/sd-perturbed-attention/pull/44)](https://github.com/pamparamm/sd-perturbed-attention/pull/44).

Contributor guide

No contributing guide indexed for this repository

Research direction

No Forge files or tests are named. Start by reviewing the linked TPG and NAG implementations, then compare them with Forge's existing SAG and PAG guidances to identify integration points. Done means both methods are available as built-in Forge guidances.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.