ComfyUI-Qwen-SEGA

Native ComfyUI integration of SEGA-style Qwen-Image sampling.

This package does not launch external scripts, does not load model weights inside the node, and returns a normal LATENT for downstream VAE Decode and Save Image nodes.

Scope

MVP supports:

• Qwen-Image text-to-image
• ComfyUI-loaded MODEL, CONDITIONING, and LATENT
• Per-step latent FFT analysis inside the denoising path
• Qwen attention RoPE patching through ComfyUI model patch APIs

Not in MVP:

• Qwen-Image-Edit
• image-to-image
• multi-reference editing
• masks
• Flux support

Installation

1. Copy this folder into ComfyUI/custom_nodes/ComfyUI-Qwen-SEGA.
2. Install any missing Python deps in the ComfyUI environment:

pip install -r custom_nodes/ComfyUI-Qwen-SEGA/requirements.txt

3. Restart ComfyUI.

Required ComfyUI Version

This node targets a recent ComfyUI build that includes native Qwen-Image support and the following internals:

• comfy.ldm.qwen_image.model.QwenImageTransformer2DModel
• ModelPatcher.set_model_sampler_calc_cond_batch_function
• ModelPatcher.set_model_attn1_patch

If your build predates native Qwen-Image support, this package will not work.

Required Models

• A native ComfyUI Qwen-Image diffusion model
• The matching Qwen text encoder loaded through normal ComfyUI nodes
• A compatible VAE loaded through normal ComfyUI nodes

This package does not bundle or download any model files.

Nodes

Qwen SEGA Settings

Builds a reusable SEGA_SETTINGS object.

Important controls:

• sega_strength: overall patch intensity
• sega_start_percent / sega_end_percent: active sampling window
• frequency_mode: adaptive or fixed weighting bias
• low_freq_weight / high_freq_weight: spectral weighting bias
• rope_scale_min / rope_scale_max: clamp range for per-frequency RoPE rescaling

Qwen SEGA Sampler

Inputs:

• model
• positive
• negative optional
• latent_image
• seed
• steps
• cfg
• sampler_name
• scheduler
• sega_settings

Output:

• LATENT

Graph

Recommended graph:

Load Diffusion Model -> Text Encode -> Empty Latent -> Qwen SEGA Sampler -> VAE Decode -> Save Image

Example Workflows

Example prompt graphs are in examples/qwen_sega_t2i_workflow.json and examples/qwen_baseline_comparison_workflow.json.

They are API-style workflow stubs intended as a starting point because node IDs and loader names vary across ComfyUI installs.

Performance Notes

• The patch computes FFT-derived statistics from the current latent on every denoising step.
• High resolutions amplify both FFT cost and Qwen attention cost.
• 4096x4096 is likely VRAM-limited on many systems.

VRAM Notes

• Expect higher peak memory than baseline Qwen sampling.
• Quantized weights may still work, but this path has not been validated across every fp8/q8/q4 variant.

Known Limitations

• The MVP uses a ComfyUI-native approximation of SEGA on top of Qwen’s existing RoPE path.
• Qwen-Image-Edit and reference-image flows are not implemented yet.
• Batch sizes greater than 1 are not the main target.

See docs/investigation.md, docs/design.md, and docs/limitations.md.