When a FLUX multi-reference workflow already has the intended conditioning and latent inputs, compare reference_latents_method by making otherwise identical workflow branches and changing only that setting. Record the option actually saved, the order of the reference inputs, and observations tied to your edit brief. The ComfyUI node is marked experimental, and its documentation does not define what each method name guarantees; treat the result as evidence for your workflow, not a general ranking.
Set up a comparison with one controlled change
Use this procedure when you can run the same prompt and reference inputs through multiple branches. First, save a working copy of the workflow so the original setup remains available. Check that the input types and connections are the intended ones; use the displayed type labels, not node color, as the check. This comparison assumes the conditioning and reference latents are already correct. It is not a test of whether references were appended or connected properly.
- Write down the edit brief and its pass conditions. For example, specify which object should change and which recognizable details should remain. Do not score general “quality” without naming the evidence you will inspect.
- Duplicate the conditioning branch for each option you intend to compare. Keep the prompt, reference latent inputs, their order, other conditioning, downstream generation path, and other settings unchanged. Change only
reference_latents_methodbetween branches. - Use the options shown in the node documentation:
offset,index,uxo/uno, andindex_timestep_zero. The node stores the selected method in conditioning for later reference-latent handling; the documentation does not explain the individual algorithms or promise a particular visual effect. See the ComfyUI node documentation. - Run the planned comparison and label each output with the branch setting and the reference order. If the interface changes a value containing
uxoorusotouxo, record the value actually stored or displayed after saving, rather than only the original menu text. - Compare outputs against the pass conditions you wrote first. Note visible evidence and failures, and retain the prompt and settings with the observation. Avoid changing a failed branch and then comparing it with an unchanged branch as though only one variable differed.
Use this hypothetical acceptance worksheet
The entries below are invented examples of how to record a comparison, not test results or predictions about ComfyUI. Assume a brief asks for a jacket-color edit while keeping the same person recognizable. Suppose all four branches use the same prompt, two reference latents in the order front portrait → jacket detail, identical conditioning, and the same generation settings. The observations are deliberately task-specific; your own results may differ.
| Saved method | Reference order recorded | Hypothetical observation | Decision for this brief |
|---|---|---|---|
offset | Portrait, jacket detail | Jacket color appears changed; face looks less like the approved portrait to the reviewer | Fails the identity condition; keep as a comparison record, not a preferred method |
index | Portrait, jacket detail | Face seems closer; jacket change is less clear in the reviewed output | Repeat only if the brief’s change condition is ambiguous |
uxo | Portrait, jacket detail | Reviewer notes both conditions appear met in this output | Candidate for this brief; not evidence of general superiority |
index_timestep_zero | Portrait, jacket detail | Output has a noticeable change outside the requested area | Fails this brief’s preservation condition |
These hypothetical rows do not describe expected behavior for the named methods. In particular, they do not establish how reference count, position, or latent dimensions affect processing. Base a decision on the saved setting, the actual output, and the criteria for the task at hand. If two options both pass, consider the number of acceptable outputs from a fixed, recorded set of runs before choosing; do not silently give one option more attempts.
Limits and what to do after a failure
The official description says the node modifies conditioning by storing the selected reference-latent method, which is used later when reference latents are handled. It also lists the four options and marks the node experimental. It does not provide a behavioral definition for each option, establish which is newest or strongest, or say which works best for a particular reference type. The node’s input and output description supports the controlled comparison, but not a promised outcome.
If outputs differ in several ways, check that your branches share the same prompt, reference order, conditioning, and generation settings before drawing a conclusion. If the saved value differs from the menu label, report the saved value and note the conversion. If no option meets the brief, do not declare a winner: review whether the prompt, references, or acceptance conditions need revision, then run a new comparison with the revised setup held constant across branches. A comparison of these settings cannot by itself establish that the underlying reference connections are correct.
Three quick questions
Should I try all four options? Not necessarily. Compare the documented choices that are available in your node and relevant to your decision. Record which ones you omitted and why; do not imply that unrun options were evaluated.
Does `uxo/uno` mean that two methods are available? The documentation lists that as an option and says values containing uxo or uso are converted to uxo before storage. Record the value after saving; the documentation does not explain additional meaning for those names.
Can I reuse the winner for another workflow? Treat it as a candidate, not a transferable guarantee. Recompare when the prompt, reference inputs, their order, conditioning, or task changes enough to alter what counts as success.
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.