The reliable way to verify a ComfyUI checkpoint is to audit the three inputs separately: MODEL, CLIP, and VAE. Do not infer completeness from the merged MODEL connection alone. The official Save Checkpoint documentation defines these as separate inputs and describes the node as saving a complete Stable Diffusion checkpoint in .safetensors format. Your acceptance decision should therefore be: each input has a recorded source, each source has an intentional role in the merge, and the saved file name and location match that record.
Specific judgment and steps
Use the node’s actual type labels as the authority. Node color, layout, and visual proximity are only clues. A practical trigger for review is any workflow where ModelMergeSimple, ModelMergeBlocks, or another merge node produces MODEL while CLIP or VAE comes from a different branch, loader, or earlier experiment.
- Identify the
MODELpath. Record the merge node, its input models, and the output that reachesSave Checkpoint.model. Write down whether this is the intended merged result or an accidental bypass. - Trace
CLIPindependently. Record the loader or node that reachesSave Checkpoint.clip, including whether it was intentionally selected for this merge and base-model family. Do not label it “the model’s CLIP” unless the graph actually shows that source. - Trace
VAEindependently. Record the VAE loader or upstream source connected toSave Checkpoint.vae. State why this VAE belongs with the intended checkpoint, or mark the relationship as deliberate external pairing. - Enter a distinctive
filename_prefix, such asmerge_042_sdxl_baseA_clipB_vaeC. The field controls the filename prefix; it does not prove that the three inputs are compatible. - Run the save node and compare the resulting file under
output/checkpoints/with the source table. Confirm that the file path and prefix identify the intended run. This is an identity check, not a compatibility proof.
The node also has hidden metadata inputs for workflow prompt information and additional PNG information. Those fields do not replace the three component-source records. The official documentation describes the output location and the three typed inputs; it does not define a built-in semantic test that certifies that their pairing is intentional. See the related model-merging source reference for the implementation context, while keeping your acceptance rule narrower than an unsupported claim about automatic compatibility validation.
A new-information acceptance worksheet
The table below is a hypothetical worksheet, not a software guarantee. The observations show how a creator could make a decision when the graph contains mixed branches.
| Field | Hypothetical observation | Decision |
|---|---|---|
| Experiment ID | merge_042 appears in the graph note and filename prefix | Identity is traceable |
| MODEL source | ModelMergeBlocks receives baseA and detailB; its output reaches Save Checkpoint.model | Accept as the intended merge path |
| CLIP source | Load Checkpoint for baseA reaches Save Checkpoint.clip; no connection comes from detailB | Accept only if retaining baseA CLIP was an explicit choice |
| VAE source | Load VAE points to vaeC, documented in the worksheet as the chosen external VAE | Accept as deliberate pairing, not as an automatic match |
| Filename and location | output/checkpoints/merge_042_sdxl_baseA_clipA_vaeC...safetensors is present | File identity matches; component correctness remains a separate decision |
| Missing or unclear source | Any row says “probably from the same model” without a visible connection or record | Stop and trace the graph before reuse |
This worksheet adds a useful distinction: “the file exists with the expected name” and “the three inputs were intentionally selected” are different observations. A filename can document an intention, but it cannot reconstruct an unrecorded connection.
Limits and the next step after failure
If the MODEL source is clear but CLIP or VAE is missing, ambiguous, or connected from an unreviewed branch, do not treat the checkpoint as complete for reuse. Return to the graph, identify the actual upstream node, and update the worksheet before saving again. If the source is intentional but comes from a different family, record that as a deliberate pairing and subject the resulting checkpoint to the creator’s own downstream compatibility review; the supplied documentation does not promise that Save Checkpoint will reject or repair such a choice.
The documentation says SDXL, SD3, SVD, and other mainstream architectures are fully supported, while other architectures may be saved without standardized metadata information. That is a limitation on what the saved artifact can communicate, not evidence that a visually plausible output proves component coherence. Do not claim that metadata, file presence, node color, or the MODEL output alone validates the CLIP-VAE relationship.
Three short questions
Does a merged MODEL automatically include the right CLIP and VAE? No. Save Checkpoint has separate `MODEL`, `CLIP`, and `VAE` inputs. Trace and record all three.
Does a correct `.safetensors` filename prove the checkpoint is coherent? No. `filename_prefix` identifies the saved run; it does not certify semantic compatibility.
Should an external VAE always be rejected? No. It may be an intentional choice. Record its source and reason, then distinguish that decision from anything guaranteed by the node documentation.
Sources and next step
- CheckpointSave - ComfyUI Built-in Node Documentation
- ComfyUI/comfy_extras/nodes_model_merging.py at master · Comfy-Org/ComfyUI · GitHub
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.