Connect the model whose identity and behavior you want as the starting point to model1; connect the model whose key patches you want to add to model2. ComfyUI’s documentation says the node clones model1 and applies key patches from model2. It does not describe this as a blanket rule that one entire model overwrites another. When the intended direction is uncertain, compare A→B and B→A under the same generation conditions before adopting either result.
Decide what should remain and what should be added
Write down two observable goals before wiring the node: which characteristics should remain from the base, and which target characteristics should be contributed by the patch source. For example, a creator might want to retain A’s established subject rendering while exploring a visual behavior associated with B. That is a hypothesis to inspect, not a guarantee that B will contribute only that behavior.
The ModelMergeAdd documentation assigns the cloned base role to model1 and the patch-source role to model2. Use the node’s actual input type labels when connecting models. Color can help locate ports, but it is only a visual cue; do not infer compatibility or direction from color alone.
- Choose one prompt that makes both the retained characteristic and the target characteristic visible. Avoid changing the prompt between direction tests.
- Connect A to
model1and B tomodel2. Record this explicitly as A→B, rather than relying on node position or memory. - Generate one output using fixed conditions: keep the prompt, seed, sampler, steps, dimensions, and other relevant workflow settings unchanged. Record those settings so the comparison can be interpreted.
- Reverse only the model inputs: connect B to
model1and A tomodel2. Generate the comparison output under the same recorded conditions. - Judge each output against the written goals. Mark retained features, target features, and unexpected changes separately. Continue with a direction only if its observed trade-off fits the project; do not treat the arrow as proof of a specific visual result.
This two-direction comparison is an editorial verification method. The documentation defines the node’s patch direction, but does not promise what a given pair of models will visibly produce.
Record an acceptance worksheet
The following is a hypothetical example, not a ComfyUI test or a prediction of guaranteed output. It shows how to record observations without treating model order as sufficient evidence.
| Direction | Fixed test prompt and conditions | Retained subject feature | Target feature | Unexpected observation | Decision |
|---|---|---|---|---|---|
| A→B | “Portrait of a seated ceramicist”; same seed and workflow settings | Hypothetically, the seated pose remains recognizable | Hypothetically, a surface quality associated with B is more visible | Hypothetically, background lighting also shifts | Keep only if the lighting change is acceptable |
| B→A | Same prompt, seed, and workflow settings | Hypothetically, B’s subject rendering becomes more prominent | Hypothetically, the target surface quality is less apparent | Hypothetically, facial proportions differ from the brief | Reject if the base identity is the priority |
Replace these example observations with what you actually see. If the prompt cannot reveal a required trait, revise the prompt once and rerun both directions under matching conditions; otherwise, differences in the test itself can be mistaken for differences in merge direction.
Limits and what to do when a result fails
The documentation explains the operation at the key-patch level: it clones the first model and adds patches from the second. It does not specify a guarantee that all features from model1 remain unchanged, that every patch from model2 is visually apparent, or that the output will be compatible with every downstream workflow. It also does not define a patch-count or visual-strength calculation. Avoid describing the result as whole-model replacement or assuming that the node provides a predictable blend percentage.
If neither direction meets the brief, first check that the intended models are connected to the labeled inputs and that both outputs used the same generation conditions. Then choose a prompt that exposes the failed requirement and compare again. If a direction preserves the subject but introduces an unacceptable change, record that trade-off rather than claiming the patch source was cleanly isolated. If the behavior remains unclear, treat the pair as unsuitable for this decision until a more discriminating acceptance prompt or another documented workflow is available. The official node reference supports the input roles; it does not establish a universal visual acceptance rule.
Common questions
Does `model2` overwrite `model1`? The documentation describes adding key patches from model2 to a clone of model1; it does not characterize the operation as a blanket replacement of the first model.
Which model is the base? The one connected to model1, according to the documented input description.
Can I know the winner from node colors? No. Read the model1 and model2 labels and verify the connections; color is only a cue.
Sources and next step
This guide interprets public documentation and uses explicitly hypothetical acceptance examples, not application tests. Controls and availability can change; verify the cited manual against your installed version.
For work created in Imagild, carry the input requirements and acceptance checklist into the photo editor. The external software controls discussed above belong to their respective products.