Treat each merge node’s MODEL output as a working result you can feed into another merge, and connect Save Checkpoint only when a stage has passed your own acceptance check. This keeps exploration in the node graph while making the checkpoint file a deliberate deliverable rather than a save after every operation. The saved checkpoint also needs the required CLIP and VAE inputs; a MODEL output alone is not the complete set of inputs documented for that node.
Decide what earns a save
Start with a goal you can judge, such as retaining a chosen visual trait while keeping a baseline’s other tested behavior acceptable. The goal is your evaluation criterion, not a behavior guaranteed by a merge node. Before merging, write down the exact source models, their input order, the node type, and each ratio or multiplier. Input order matters: the supplied documentation describes the first input as the base in ModelMergeAdd and ModelMergeSubtract, and defines ModelMergeSimple’s ratio relative to model1 and model2.
- Load the source models and keep their identities fixed during a comparison. Connect them to
ModelMergeSimple,ModelMergeBlocks,ModelMergeAdd, orModelMergeSubtractaccording to the operation you intend to investigate. - Connect the first merge node’s
MODELoutput to a later merge node’sMODELinput when you want to continue the chain. Record the other input and settings at each step. A node output is an intermediate graph result; it is not, by itself, the checkpoint file described bySave Checkpoint. - Compare each stage with the same prompt and declared generation settings, changing only the model stage under review. This is an author-recommended comparison method, not an official guarantee; record other conditions you can control, such as CLIP, VAE, and sampler settings, and keep them fixed. Judge against the written goal and note desired and unacceptable observations.
- If a stage passes, connect that stage’s
MODELoutput toSave Checkpoint, provide its associatedCLIPandVAE, and set a distinct filename prefix. If it fails, continue from the last passing graph output or revise the next merge; do not label an unaccepted stage as the final asset.
The official ModelMergeSimple documentation specifies the ratio endpoints: 1 yields all of model1, while 0 yields all of model2. ModelMergeBlocks exposes separate input, middle, and out ratios. These are distinct controls, so record them separately rather than summarizing them as one blend value. For add and subtract, record the multiplier where applicable; do not assume that a value has the same meaning across node types.
Stage acceptance worksheet
The following is a hypothetical example, not an observed ComfyUI run or a prediction of what these settings will produce. Suppose the goal is to retain a baseline’s tested composition while checking whether a second model contributes a desired style. The creator uses one unchanged prompt and records observations before deciding where to save.
| Stage | Node, ordered inputs, settings | Hypothetical observation against goal | Decision and recovery point |
|---|---|---|---|
| Inputs | Baseline A, candidate B | Baseline prompt result recorded as comparison reference | Keep both source identities fixed |
| First merge | ModelMergeSimple; A then B; ratio 0.7 | Composition appears acceptable; requested style change is not clear in this comparison | Do not save; retain A and B as the restart point |
| Second merge | ModelMergeBlocks; prior output then B; input 0.7, middle 0.5, out 0.7 | Hypothetically, the desired change is more evident, but one required composition check fails | Reject this stage; return to the first-merge output |
| Accepted stage | Revised block settings; same ordered inputs and comparison prompt | Hypothetically, both written checks pass in the reviewed samples | Connect this MODEL output to Save Checkpoint and record its prefix |
Use the worksheet to name the last passing output explicitly. The example’s values and observations are invented to show how the decision is recorded; they do not establish a preferred ratio or expected visual effect.
Limits and what to do when a stage fails
Save Checkpoint documents .safetensors output under output/checkpoints/ and notes that architecture support and standardized metadata can vary. See Save Checkpoint documentation before assuming metadata is standardized for your architecture.
If a comparison fails, first identify which written check failed. Return to the last passing graph output and change one recorded operation or setting for the next comparison. For ModelMergeSubtract, document which model is the base and which is subtracted, plus the multiplier; for ModelMergeAdd, document the base and patch-contributing model. The official descriptions explain these input roles, but they do not promise a particular qualitative result. If a merge cannot accept the intended inputs or the resulting file cannot be loaded, check the model architecture and node input types instead of inferring a repair from a node’s color. Colors can be a visual clue; use the actual type labels to confirm compatibility.
Acceptance is not complete when the save node is connected. After saving, locate the resulting checkpoint, load it again in a separate validation workflow, and check it against the same written goal and comparison setup. This is a recommended validation step, not a claim that the checkpoint was tested here. If it does not reload or fails the goal, keep the last accepted graph stage as the recovery point and investigate the missing input, architecture compatibility, filename, or failed acceptance check before saving another candidate.
Three quick questions
Can I save every intermediate result? You can choose to attach a save node to a stage, but this workflow reserves checkpoint files for stages that passed your stated checks. Keep earlier stages in the graph when they remain useful for continuing the experiment.
Does `ModelMergeSimple` ratio `0.7` mean 70% of the second model? No. The documented endpoints establish that `1` is all `model1` and `0` is all `model2`; read the value with that ordering in mind, and record the inputs as connected.
What does the saved checkpoint contain? The node documents a complete checkpoint with UNet, CLIP, and VAE. Supply the `CLIP` and `VAE` associated with the intended model.
Sources and next step
- ModelMergeSimple - ComfyUI Built-in Node Documentation
- ModelMergeBlocks - ComfyUI Built-in Node Documentation
- ModelMergeSubtract - ComfyUI Built-in Node Documentation
- ModelMergeAdd - ComfyUI Built-in Node Documentation
- CheckpointSave - ComfyUI Built-in Node Documentation
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.