MiniMaxH3AddGuide is not a whole-clip video-to-video restyler. It adds an image or short clip as conditioning at a chosen frame_idx; sampling still begins from the workflow’s empty latent. Use it to anchor a generated shot at selected moments. For a whole-video reference-driven restyle, route the job to MiniMax H3 R2V. For changes limited to a region, use the documented mask workflow instead.
Decide what the reference must control
Before wiring nodes, state the desired contract in one sentence: “Generate a new sequence that passes through these poses,” “use this clip to guide the whole result,” or “change only this masked area.” These requests need different routes. A guide anchor constrains a point or supplied segment on the output timeline; it does not mean every frame of a source video is being transformed while the rest is preserved.
| Goal | Route | What the documented workflow establishes |
|---|---|---|
| New video with chosen moments pinned | H3 workflow plus MiniMaxH3AddGuide | Image or clip conditioning at a timeline index; generation starts from an empty latent |
| Whole result guided by reference video | H3 R2V | Reference videos can guide video generation; the guide recommends assigning reference roles |
| Regenerate a selected area | H3 Fun ControlNet Union masks, or latent noise masks where appropriate | Region-oriented control or masked regeneration; not a guarantee of unchanged pixels outside the mask |
Build a timeline anchor deliberately
- Choose the multiframe reference template if the task is a new sequence with pinned compositions. It combines an H3 R2V node with chained Add Guide nodes; consult the multiframe workflow.
- Connect the H3 node’s
positiveand latent outputs toMiniMaxH3AddGuide. For image guides, connect the VAE. For audio at the same anchor, connect the audio VAE and set the audio guide accordingly. - Set
frame_idxin frames at 24 fps. For a time in seconds, the template usesround(seconds * 24): a hypothetical 1.5-second anchor is frame 36. Ensure the anchor and guide length fit within the generated duration. Negative indices count from the end. - For a multi-frame clip, check the supported guide length: the node crops batches to valid
17k+5lengths, and a clip shorter than five frames uses only its first image. Chain additional guides through theirpositiveoutputs when you need multiple anchors. - If the text encoder should refer to a guide image by a reference tag, also connect that image to an extra
ref_imagesslot. An image attached only to Add Guide pins a composition but is not thereby exposed to the text encoder.
The H3 native workflow guide explicitly distinguishes this conditioning path from video-to-video: the sampler still starts with the workflow’s empty latent. Do not treat denoise as a V2V control for scaling the guide’s strength.
Route whole-clip and regional edits separately
For a whole-clip reference-driven result, use the H3 R2V template and assign the source video a clear role in the prompt, such as motion or camera reference. R2V accepts reference videos, but the documentation describes reference-guided generation, not a guarantee that every source frame, object, or camera move will be preserved exactly. If timing matters, inspect the source and output at corresponding events.
For a localized change, select a mask-oriented workflow rather than adding a still guide and expecting it to isolate a region. The native guide points to Fun ControlNet Union masks for regional editing. It separately documents latent noise masks for inpainting, removal, or extension: on the sampler’s denoise_mask, 0 means keep and 1 means regenerate; video masks snap to the 2×2 latent patch grid. These controls define a regeneration area, not a pixel-level preservation promise.
Diagnose with a hypothetical acceptance check
Suppose a creator has a 6-second, 24-fps street clip and wants to change its overall visual style while retaining its sequence of movement. That is a whole-clip reference request, not a request to pin one pose. If instead the brief is to generate a new 6-second scene with a supplied still at 1.5 seconds and another at 4 seconds, Add Guide is the relevant route: the hypothetical anchors are frames 36 and 96.
| Hypothetical observation | Diagnosis | Next action |
|---|---|---|
| Output starts from a new scene and only approaches the supplied image near frame 36 | Expected anchor-conditioning behavior, not evidence of V2V failure | Keep Add Guide for a keyframe task; use R2V if the source clip should guide the whole sequence |
| A short clip guide contributes only its opening image | The guide batch may be below the documented five-frame threshold | Use a still intentionally or provide a guide length compatible with the documented 17k+5 rule |
| Prompt mentions a guide image but the text encoder does not seem to refer to it | The image may be connected only to Add Guide | Also wire it to a ref_images input if the prompt needs a reference tag |
| Only a defined region should be regenerated | A timeline anchor is the wrong control for the request | Move to the documented mask route and review the full output for unwanted changes |
These are routing checks based on documented behavior, not results from a software test. If an R2V result does not preserve a required event or camera move, treat that as an acceptance failure, revise the reference role or brief, and review another generation; do not claim that changing Add Guide’s denoise restores V2V behavior.
Common questions
Can I put the whole source video into Add Guide? The node accepts a clip as a multi-frame guide, but that makes it conditioning at an anchor, with documented cropping rules. It does not change the empty-latent sampling contract into whole-video transformation.
Will an Add Guide image be visible to the prompt encoder? Not if it is connected only to Add Guide. The multiframe workflow says guide-only images are not visible to the text encoder; connect the image to an additional `ref_images` slot when prompt-side reference tags are needed.
Does a mask guarantee everything outside it stays identical? The documentation describes masked regeneration and defines mask values, but does not guarantee exact unchanged pixels outside the mask. Inspect the resulting clip against the requirement rather than assuming the mask is a preservation warranty.
Sources and next step
- ComfyUI MiniMax H3 Text-to-Video, Image-to-Video, and Reference-to-Video Workflows
- ComfyUI MiniMax H3 Multiframe Reference Workflow
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.