Eric Qwen Conditioning Blend

The smallest node in the conditioning system, and the one you'll reach for when you want two edit intents at once. Eric Qwen Conditioning Blend takes two QWEN_CONDITIONINGs - say "make it black and white" and "add strong contrast" - and produces a weighted mixture of their embeddings.

Mechanically it's the weighted sum (weight_A × cond_A + weight_B × cond_B) / (weight_A + weight_B). That's the same math as Conditioning Interpolate with alpha = weight_B / (weight_A + weight_B); Blend just expresses it more readably. Instead of remembering that alpha 0.7 means "70% toward B", you set weight_A and weight_B directly. For a compound expression like 60% neutral + 40% smile, that reads a lot better in a graph you'll come back to next week.

The inputs

• cond_A, cond_B - two conditionings from Encode nodes (or from any manipulation node, if you're blending blends).
• weight_A, weight_B - each 0.0–2.0, default 0.5/0.5.
• normalize - on by default, and the tooltip's word is "recommended". It rescales the weights to sum to 1.0, so 0.5/0.5 stays a true half-and-half and you don't accidentally push the combined embedding's magnitude up or down. Turn it off only if you're deliberately scaling overall strength.

Output is a single conditioning, which feeds Conditioned Edit directly, or another Blend if you want to stack three intents.

When to use it

Blend is the coarse-intent tool. In this system's mental model, the text prompt defines what kind of edit happens and the conditioning mix shapes it within that space. If you want a portrait that's simultaneously warm-toned and softly lit, blend two encodings rather than trying to cram both adjectives into one sentence - the model's Qwen2.5-VL encoder is an LLM, and tag-stacking is not how it responds. A weighted blend of two cleanly-encoded intents is often more controllable than one overloaded prompt.

It's also the friendlier sibling for A/B work: blend at weight 0/1 for pure A or B, then sweep between. For a wide interpolated sweep with smooth perceptual steps you'd want Interpolate with slerp instead - Blend's lerp-style mixing is for "a bit of both", not for animation.

Installing it

Same pack as the rest of the conditioning system. ComfyUI Manager → search "Eric Qwen-Edit", or:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Eric_Qwen_Edit_Experiments.git

Restart ComfyUI. It doesn't load any model itself - it only manipulates conditioning tensors - but you need the Qwen-Image-Edit model (Qwen/Qwen-Image-Edit-2511, ~54 GB) loaded by a pack loader before any Encode node upstream will work.

Gotchas

• Blend doesn't re-encode anything. It's pure tensor math on whatever the two inputs carried, so garbage in, garbage out - if cond_A came from a badly-worded prompt, blending won't fix it.
• Normalize off is rarely what you want. With normalization disabled, weights above 1 can inflate the embedding magnitude and change not just the mix but the overall guidance strength. Leave it on unless you have a reason.
• Scope doesn't exist here. Unlike Interpolate, Blend has no scope control - it mixes the whole token sequence, image and text tokens alike. If you need to interpolate only the text tokens (expression edits) or only the image tokens (style), that's Interpolate's job, not Blend's.