A still connected only to a MiniMax H3 Add Guide node anchors a composition at a timeline frame; it does not automatically become an image that the text encoder can refer to in the prompt. If the prompt should also identify or describe that still, connect the same image to an additional ref_images slot and refer to the corresponding tag. Use both connections only when the image has both jobs.
Decide what each image must do
Treat timeline anchoring and prompt reference as separate inputs, not two names for the same function. In ComfyUI’s documented multiframe workflow, Add Guide supplies a still, a frame index, and optional audio as a guide at a point in the generated video. The ref_images inputs supply images visible to the text encoder and addressable by prompt tags. The official H3 multiframe workflow documentation explicitly distinguishes these roles.
For each still, answer two questions: must the video be anchored to its composition at a particular frame, and must the prompt be able to name or describe that still? A “yes” to both calls for two connections to the same image. A “yes” only to the first calls for Add Guide alone. A “yes” only to the second calls for ref_images alone. Neither input should be added just because another image uses it.
| Image role | Add Guide | ref_images | Prompt treatment |
|---|---|---|---|
| Shot 2 composition at 1.5 seconds | Yes | Only if the prompt must refer to it | Use the assigned reference tag if connected |
| Identity image used as a prompt reference, no separate timed anchor | No | Yes | Refer to its assigned tag |
| Image that must both anchor Shot 3 and be named in the prompt | Yes | Yes | Describe its shot or content using its tag |
The table is a hypothetical planning example, not a generation result. The template’s identity image is also its first frame, but that example does not make every Add Guide image a prompt reference.
Connect in a deliberate order
- Choose the still and write down its intended job: timeline anchor, prompt reference, or both.
- For each timeline anchor, connect the image through its guide block to an Add Guide node. Set
frame_idxfor the intended point. The documented template converts seconds at 24 frames per second withround(seconds * 24); for a hypothetical 1.5-second anchor, that gives frame 36. - If the text encoder must see that same still, connect it separately to an available
ref_imagesslot. In the image audit table, record the slot, its corresponding reference tag, and the shot or visual role named by that tag. Only use a reference tag for an image connected toref_images; an Add Guide connection alone does not make the image visible to the text encoder. - Check the guide chain and its timing: each guide’s
positiveoutput is chained in series into the Basic Guider in the documented workflow. Keep the frame index plus guide length within the generated duration.
The official H3 prompt guide is the linked guidance for writing full-reference prompts. The precise tags and supported inputs should be taken from the active H3 workflow, not inferred from another interface.
Audit image, frame, and tag together
Use one row per image, not one row per shot: a still can be an anchor and a prompt reference at once. In this hypothetical plan, the clip is 6 seconds at 24 fps, and the creator intends anchors at 1.5 and 3.0 seconds. The corresponding example frame calculations are 36 and 72; they illustrate the documented conversion, not a generated result.
| Image | Add Guide connection and anchor | ref_images connection | Prompt tag and intended meaning | Acceptance check |
|---|---|---|---|---|
| Shot 2 still | Yes; 1.5 s, frame 36 | Yes; first slot | First reference tag identifies Shot 2 still | Graph, frame value, and prompt agree |
| Shot 3 still | Yes; 3.0 s, frame 72 | No | None; prompt does not claim to reference it | Anchor exists; no reference claim |
| Identity still | No separate anchor | Yes; second slot | Second reference tag identifies identity image | Slot order and prompt meaning agree |
Before generation, check that each loaded file is the intended image, its Add Guide connection and frame value match the planned anchor, and any ref_images slot and prompt tag agree with the image’s intended shot or visual role. This is a wiring and mapping check, not a guarantee of visual fidelity.
Diagnose a missing or misplaced reference
If a prompt reference seems missing or points to the wrong image, use the image audit table to verify that the image is connected to ref_images and that its recorded slot matches the tag in the prompt. An Add Guide connection alone does not provide prompt visibility. Correct the connection or mapping before changing shot descriptions.
If the anchor is missing or appears at an unexpected point, inspect the Add Guide image, frame_idx, guide length, and generated duration. The documented template computes frame positions from seconds at 24 fps; negative frame indices count from the end, and the index plus guide length must remain within the output duration. Also verify that guide positive outputs are chained in series into the Basic Guider.
A prompt reference is not a substitute for a timed anchor, and an anchor is not a substitute for a prompt reference. Slot availability and settings can differ between workflows, so verify the active H3 workflow before relying on a particular connection.
Common questions
Does every Add Guide image need a `ref_images` connection?
No. Add it only when the prompt should also be able to refer to that image. A still used solely as a timeline anchor can remain connected only to Add Guide.
If the same still is connected twice, is that a duplicate mistake?
Not when it has two intended roles. The Add Guide branch supplies a timed composition anchor; the ref_images branch makes the image available to text-encoder prompt references. Keep one row in your image audit and record both connections.
What should I check if the tag is correct but the video still differs?
Use the image audit table to confirm the slot-to-tag mapping and intended role; check Add Guide timing separately if the image is also an anchor.
Sources and next step
- ComfyUI MiniMax H3 Multiframe Reference Workflow
- docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md · MiniMaxAI/MiniMax-H3 at main
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 film workspace. The external software controls discussed above belong to their respective products.