ComfyUI-Spectrum-Qwen

A ComfyUI custom node that applies a Spectrum-style spectral hidden-state forecaster to native Qwen Image-family models, with the primary target being Qwen-Image-Edit-2511.

This repo patches the Qwen transformer core used by the loaded MODEL, captures the final pre-norm_out image-stream hidden state on real steps, fits a ridge-regularized Chebyshev model over recent real steps, and uses that forecast to skip selected expensive Qwen transformer passes.

What this supports

Targeted support:

• ComfyUI native Qwen-Image-Edit-2511 workflows
• Other native ComfyUI Qwen Image-family models that still expose the same transformer internals (img_in, txt_norm, txt_in, transformer_blocks, norm_out, proj_out, time_text_embed)

Intended scope:

• Qwen-Image
• Qwen-Image-Edit
• Qwen-Image-Edit-2511
• likely other near-identical native Qwen Image variants, as long as ComfyUI keeps the same core transformer layout

What this does not claim

This repo does not claim universal compatibility with:

• Nunchaku / quantized Qwen wrappers
• wrappers that hide or replace the native Qwen transformer core
• ControlNet-heavy runs where per-block residual injection must remain exact
• architectures outside the native Qwen Image-family transformer layout

When unsupported conditions are detected, the node falls back to real forwards instead of forcing a broken forecast path.

Node

Spectrum Qwen Model Patcher

Category: model/optimization

Inputs:

• model: the ComfyUI MODEL to patch
• warmup_steps: initial steps forced to run normally
• tail_actual_steps: final steps forced to run normally
• history_points: number of captured real hidden states used for fitting
• chebyshev_degree: polynomial degree for the fit
• max_consecutive_forecasts: max skipped Qwen passes in a row before forcing a refresh step
• ridge_lambda: ridge regularization strength
• cache_device: where hidden-state history is stored (main_device, offload_device, cpu)
• force_actual_on_control: force real forwards when control residuals are present
• debug: print per-step decisions and a summary

Output:

• patched MODEL

Recommended starting settings

Conservative:

• warmup_steps = 5
• tail_actual_steps = 2
• history_points = 5
• chebyshev_degree = 3
• max_consecutive_forecasts = 1
• ridge_lambda = 1e-4
• cache_device = main_device
• force_actual_on_control = true
• debug = false

More aggressive:

• warmup_steps = 5
• tail_actual_steps = 2
• history_points = 6
• chebyshev_degree = 3
• max_consecutive_forecasts = 2
• ridge_lambda = 1e-4

How it works

This implementation is deliberately narrow and reviewable.

1. It locates the native Qwen transformer core inside the loaded MODEL.
2. On real steps, it runs the original forward path unchanged and captures the hidden state immediately before norm_out.
3. It keeps the last history_points captured states.
4. On forecast steps, it predicts the next pre-norm_out hidden state using a ridge-regularized Chebyshev fit over recent real steps.
5. It then runs the lightweight tail only:
   • time_text_embed
   • norm_out
   • proj_out
6. Warmup, tail, refresh, and unsupported cases stay on the real path.

Why the forecast target is pre-norm_out

For Qwen Image-family models, the expensive part is the long transformer block stack. The norm_out + proj_out tail is cheap by comparison.

Forecasting the image stream before norm_out keeps the sampled step-specific modulation in the cheap tail while still bypassing the expensive transformer blocks.

Installation

Manual

Clone into your ComfyUI custom_nodes directory:

git clone https://github.com/xmarre/ComfyUI-Spectrum-Qwen-Proper.git

Restart ComfyUI.

No extra Python dependencies

This repo uses only Python stdlib + whatever ComfyUI already provides.

Usage

Typical placement:

Load Qwen model -> (optional LoRA/model modifications) -> Spectrum Qwen Model Patcher -> KSampler

Important:

• Put Spectrum after any node that mutates the model you want sampled.
• Start conservative first.
• Enable debug for the first test run so you can confirm the actual/forecast pattern.

Example workflow pattern

Qwen-Image-Edit-2511 basic edit

1. Load your Qwen Image Edit model normally
2. Build conditioning with the usual native Qwen Image Edit nodes
3. Insert Spectrum Qwen Model Patcher between the loaded model and the sampler
4. Sample with a normal KSampler setup
5. Inspect output quality before pushing forecast aggressiveness harder

Error handling / fallback behavior

The patcher intentionally prefers correctness over overclaiming support.

It falls back to real forwards when:

• the loaded model is not a supported native Qwen Image-family core
• not enough real history exists yet
• you are still in the warmup region
• you are in the protected tail
• zero_cond_t is active
• gradient checkpointing is active
• ControlNet/control residuals are present and force_actual_on_control = true
• forecast reconstruction fails for any reason

Assumptions

This repo assumes the inner Qwen transformer core still exposes fields compatible with the current native Qwen Image-family layout:

• img_in
• txt_norm
• txt_in
• transformer_blocks
• norm_out
• proj_out
• time_text_embed

If ComfyUI changes those internals materially, this repo will need updating.

Caveats and known limitations

1. This is not a paper-exact universal Qwen port. It is a practical native-ComfyUI Qwen Image-family patcher.

2. Forecasting target approximation. The implementation forecasts the final pre-norm_out image stream. That is the most useful cheap-tail split for this family, but it is still an engineering approximation of the general Spectrum method rather than a claim of paper-exact integration for every Qwen variant.

3. ControlNet path stays conservative. Per-block control residual injection is not reconstructed analytically here. By default those cases stay on the real path.

4. Native layout required. Quantized or heavily wrapped variants may expose a different forward path and are not promised here.

5. No automatic quality oracle. The node does not score image quality for you. You still need to compare results and tune aggression based on your workflow.

6. CPU cache mode trades VRAM for latency. cache_device = cpu can reduce GPU pressure, but it may erase some of the speed win depending on token count and resolution.

Repository layout

ComfyUI-Spectrum-Qwen-Proper/
├── __init__.py
├── nodes.py
├── README.md
├── requirements.txt
├── pyproject.toml
├── LICENSE
└── spectrum_qwen/
    ├── __init__.py
    ├── chebyshev.py
    ├── config.py
    ├── constants.py
    ├── controller.py
    ├── forward_qwen.py
    ├── model_introspection.py
    ├── patcher.py
    ├── state.py
    └── utils.py

Development notes

The implementation is intentionally small:

• one node
• one controller
• one forecasting core
• one inner-forward patch

No framework-like layer was added on top.