When a FLUX edit uses several reference images, first confirm that the edit model supports multiple reference latents; the presence of ReferenceLatent nodes alone does not establish compatibility. For a supported model, connect each node’s CONDITIONING output to the next node’s conditioning input, and give each node its own reference LATENT. ComfyUI documents this sequence as accumulating reference samples, not replacing the earlier ones. Verify the wiring by comparing an A-only branch with an A-then-B branch.
Decide and wire the chain
- Check the model or workflow documentation for explicit support for multiple reference latents. If you cannot confirm support, treat a multi-reference result as unverified rather than inferring support from the node being available.
- Identify the initial CONDITIONING source and the LATENT source for each reference image. Follow the connection’s type label, not its color alone; colors are only a visual clue.
- Add a ReferenceLatent node for each reference. Connect the initial conditioning to A, A’s output to B’s conditioning input, and B’s output to C if needed. Connect each reference image’s latent to its corresponding node’s latent input.
- Trace the final CONDITIONING output to the downstream conditioning input used by the edit workflow. Record that destination so you know which branch is actually being used.
- Build a comparison branch with only A. Keep the starting conditioning and downstream route comparable to the A-to-B branch, so the changed structure is the addition of B rather than an unrelated rewiring.
The node documentation says a connected latent modifies the conditioning and stores reference latents as a list of sample tensors. It also says that with no latent connected, the incoming conditioning passes through unchanged. That pass-through behavior is not evidence that an absent reference was added.
Acceptance worksheet
The entries below are an explicitly hypothetical wiring review, not a software test or a guarantee about generated images. “Observed” means the connection visible in the workflow graph; use the actual node names and destinations in your own graph.
| Branch or node | Reference LATENT source | CONDITIONING input | CONDITIONING output goes to | Wiring decision |
|---|---|---|---|---|
| A-only: node A | Image A latent | Initial conditioning | Edit conditioning input | Baseline contains A node only |
| A-to-B: node A | Image A latent | Same initial conditioning | Node B conditioning input | A is not bypassed |
| A-to-B: node B | Image B latent | Node A output | Edit conditioning input | B receives A-processed conditioning |
| Hypothetical faulty B | Image B latent | Initial conditioning | Edit conditioning input | A is bypassed; repair the input wire |
| Hypothetical unconnected node | No latent connection | Incoming conditioning | Next conditioning input | Pass-through only; do not count it as a reference |
Before running either branch, follow the CONDITIONING path from the source to the edit input and check each handoff against the table. In the A-to-B branch, the decisive observation is that B’s conditioning input traces back to A’s output, not directly to the initial source. This checks the graph’s transmission structure; it does not by itself prove model compatibility or a particular visual result.
Limits and next steps
If the model does not support multiple reference latents, correct wiring alone cannot establish that it will use them as intended. Check the model’s own documentation or a model-specific workflow before interpreting output. The supplied ReferenceLatent documentation explains the node’s connection behavior, but does not identify which FLUX variants support multiple references or promise how much influence each reference has.
If the A-only branch is wired as intended but the chained branch still looks like it uses only one reference, first check the model-support question, then inspect whether the final edit input receives B’s output and whether B receives A’s output. If those links are correct, the graph alone cannot settle what the model did with the references; report the result as unresolved rather than claiming that a reference was definitely retained or ignored. Do not infer success from node count.
The official ReferenceLatent node documentation describes the optional latent input, pass-through case, and output type. Its notes on chained nodes describe reference accumulation and make model support a condition, not a universal FLUX guarantee.
Common questions
Does adding a second node replace the first reference? The documented sequence appends reference samples. That conclusion applies when the model supports multiple reference latents and the second node receives the first node’s conditioning output.
What if B is connected to the original conditioning? Then the graph does not show B receiving A’s processed conditioning. Reconnect B’s conditioning input to A’s output, then trace the final output to the edit input.
Does an unconnected latent input count as a reference? No. The node documentation says conditioning passes through unchanged when no latent is connected. Check the LATENT input itself, not just the node’s presence.
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.