F2_k_Spectral_Graft Swaps Flux.2 Klein Clothing Sans Face Changes
F2_k_Spectral_Graft is a new ComfyUI custom node that performs regional image editing on the Flux.2 Klein model without altering the rest of the scene. Users paint a mask over an area they want to change, and the node swaps clothing, adds or removes objects, or restyles that region while keeping everything outside the mask faithful to the original. It works through frequency-domain feature grafting, nudging the diffusion model’s internal value features in the masked area toward a target rather than forcing pixels directly.
Magirad who also created Flux_ID_Adjuster_V2 built this node specifically to solve a common frustration with Flux.2 Klein, where asking the model to change something like clothing often shifts the subject’s pose, face, or lighting along with it. The tool takes a different approach from typical attention-swap or hard-paste methods by letting the model synthesize the edit naturally, which tends to produce more seamless results.
Editing modes and practical presets
Key capabilities
- Swap clothing using a reference garment image.
- Add or remove objects by painting masks.
- Restyle specific areas without changing the background.
- Face and identity edits with light touch settings.
- Background lock prevents surrounding regions from drifting.
- Built-in presets save and load common configurations.
- Expert controls reveal advanced frequency cutoffs.
- Works across single and dual-stream transformer blocks.
Privacy-conscious professionals and small agencies running local AI will find this useful for photo manipulation tasks where maintaining the original subject identity matters. Hobbyists with consumer GPUs can run it entirely offline since it targets the distilled Flux.2 Klein 9B model. The presets system means you can save, export, and reload settings rather than memorizing parameter combinations each time.
Developer notes and known limitations
Magirad describes this as a “vibe coded” project, meaning it was built through rapid iteration rather than formal engineering. Testing was done specifically on the FP8 distilled version of Flux.2 Klein 9B, so results on other quantizations aren’t guaranteed. The tool depends on the model supplying reference image tokens and will pass through unchanged with a warning if those are missing.
Spatial alignment works by position, not by meaning. If you drop a reference element into a mask that doesn’t roughly overlap where that element sits in the scene, the result weakens, and far-off-center additions are notably less reliable. Text-only addition also tends to be weaker than reference-image addition because it relies entirely on the prompt plus shape blocks to define what appears.
"The issue with flux.2 klein 9b model is if you ask it to change the clothes it will change the pose or face or the lighting. this node prevents that and also ensures the swap happens realistically." — Source: Reddit
Project Links
