FL Qwen Image Edit with Strength

FL_QwenImageEditStrength is a modified version of ComfyUI's stock TextEncodeQwenImageEditPlus node, and the modification is the whole point: Qwen-Image-Edit's native multi-image conditioning treats every reference image you feed it equally, and this node adds an individual strength dial per image so you can weight them differently in the same edit call.

That capability - multi-image editing, up to three inputs at once, combining person+person, person+product, or person+scene - arrived with Qwen-Image-Edit's 2509 revision and is a real, current thing people build workflows around. But the stock encoding node doesn't let you say "the face reference matters more than the background reference" - every image goes in at equal weight. FL_QwenImageEditStrength is Fill-Nodes' answer to that gap.

The inputs. clip and prompt are the standard text-conditioning path Qwen-Image-Edit needs; optional vae follows the same convention as the stock node. Then up to three reference images, each with its own strength: image1/image1_strength, image2/image2_strength, image3/image3_strength - all three images optional, and each strength ranges −10 to 10 (default 1.0). interpolation_method (weighted_sum/average/maximum/minimum/multiply, default weighted_sum) controls how the per-image conditionings get combined into one - weighted_sum is the mode that most directly leans on the strength values you set, while the other four are alternate combination math worth experimenting with if a straight weighted blend isn't giving you what you want.

The output is CONDITIONING only - this node builds the conditioning for Qwen-Image-Edit, it doesn't run the edit itself. You still need the model, sampler, and the rest of a normal Qwen-Image-Edit graph downstream; this just replaces the encoding step.

Two honest caveats before you lean on this. First, it's filed in the pack's own WIP category, so treat it as functional but not battle-tested - weighted_sum is the safer starting point if the other interpolation modes produce something unexpected. Second, and more important: Qwen-Image-Edit itself has a well-documented structural limitation that no amount of strength-tuning in this node will fix. The model re-emits the entire frame on every edit rather than patching only what changed, which means unedited regions can drift, and faces specifically are known to lose consistency across a chain of edits - badly enough that the community's standard workaround is inpainting the face back in afterward. If you're seeing drift or identity loss, that's the base model's known behavior, not something wrong with your strength values.

Installing the node itself is the standard Fill-Nodes routine - search "Fill-Nodes" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI_Fill-Nodes

• then restart ComfyUI. The node adds no model weights of its own; you need Qwen-Image-Edit's actual model, CLIP, and VAE files already set up in your workflow the normal way, since this only changes how the conditioning gets built.

Where people get burned. Start each _strength near 1.0 and adjust from there - the −10 to 10 range technically allows strongly negative weighting, but that's a genuinely unusual, more experimental setting rather than a sensible starting point, and it's easy to end up with a confusing result if you dive straight into negative territory without first understanding how the image behaves at normal weights. And if results look inconsistent switching between interpolation modes, remember only weighted_sum is the mode explicitly designed around the strength values - the others may respond to your per-image strengths differently than you'd expect from a simple weighted blend.