AI Image & Video

Dual-Reference Mosaic Corruption in AI Image Editing: A Root-Cause Diagnosis

The mosaic only appears when two reference images AND no acceleration model (cfg above 1) are both true. The 'just drop one reference image' advice I gave last month was a misdiagnosis; the real knob is the step count.

10/19/2026 ｜ 4 min read ｜ Model: Qwen-Image-Edit-2511. The table's logic transfers, its numbers may not. About 10 minutes to read

The one idea that matters

Changing the test material and the parameter amounts on the same symptom is how you separate “the method is broken” from “one parameter is off”. Test once and conclude, and you can easily misjudge a fixable parameter problem as a dead route, then take a long detour for nothing.

Steps

Step 1: Confirm your symptom

“Dual-reference mosaic corruption” means: you give the AI two reference images (say a scene photo and a character photo) and ask it to place the character in the scene, and the faces, clothes and hair come out as shattered, pixelated noise that no prompt rewording can rescue.

If what you see is “two people sharing one face” or “an extra person appearing from nowhere”, that is a different multi-reference bug (attribute leakage, identity duplication) and not what this note fixes. Mosaic is a texture-level disaster; attribute leakage is an identity-level mix-up.

Step 2: List three variables

1. Whether an acceleration or distilled model/LoRA is on (for example Lightning-style add-ons that squeeze denoising into 4 steps)
2. The cfg value: usually locked to 1 with an acceleration model, commonly 2 or more without
3. Number of reference images: 1, or 2 and more

Step 3: Place your combination in this table

The mosaic appears only when “two images AND cfg above 1 (no acceleration model)” are both true. With acceleration on, one or two images never produce mosaic; too few steps only makes the character look less like the person.

The same applies to any other model: list every switch you can turn, change one at a time, and compare the broken combinations against the working ones. The culprit is usually the intersection of two variables, not a single variable.

Step 4: Pick the fix for your case

• Acceleration model + two images, but the character does not look enough like the person: do not drop a reference image, raise the steps. In our test, going from the default 4 steps to 10 fixed it: right likeness, clean quality, and both the scene detail from two images and the speed of the acceleration model stay.
• No acceleration model + two images, and you get mosaic: switch to the acceleration model (cfg=1) with enough steps first. Only if your framework cannot switch, fall back to dropping an image and compensating with a text description.
• Not sure which case you are in: run all 4 combinations (acceleration on with low steps, on with high steps, off with two images, off with one image) and compare by eye. Different base models and different acceleration LoRAs may have different thresholds.

Step 5: Verify the fix is stable

• Test once with the real material you will actually use, not a random web image.
• Change the seed and test again to rule out a single lucky sample.

Traps you will hit

1. A flawed reference image corrupts your judgment. The character reference in our first round was AI-generated and already had odd texture on the face. After the fix the output looked “more like the reference”, faithfully copying the flaws too, so it felt no better. A clean real photo made it clear the fix worked. Before judging quality, check that the thing you use as the answer key is clean.
2. “Looks like the person” and “has mosaic” are two separate problems. Too few steps on the acceleration model only makes the character less like the person; mosaic comes from another parameter combination. Mix them in one diagnosis and nothing seems to work, because you are chasing two rabbits at once.

Universal rescue: how to ask an AI when you are stuck

I am using [your model / tool] for multi-reference image composition and I see [symptom: mosaic corruption / identity duplication / other]. I suspect these parameters but cannot tell which intersection causes it: whether an acceleration or distilled model is on, the cfg value, the number of reference images, the step count. Please design a minimal cross-check experiment (change one variable at a time), list which configurations I should test and what to record for each, so I can isolate the true root cause. Do not just tell me to reword the prompt.

Run the experiment your AI proposes and compare results by eye. It is much faster than continuing to guess at prompts.