For most H3 R2V runs, start with match: it fits the reference to the canvas and is the documented default. Choose max only when you have a reason to retain a larger reference size and can accept a much slower run. Crucially, max does not enlarge a small file: if its short edge is below 2048 pixels, pre-scale it yourself when that size is needed, then check that both sides are multiples of 32.
Decide what the reference needs to contribute
Set the target video canvas first. Then identify the job of each reference image: for example, is it mainly establishing a character’s identity, a visual style, or a subject detail that needs close inspection? The setting changes how the image is fitted for reference use; it does not by itself guarantee better adherence or restore detail that is absent from the source.
Use match when fitting the reference to the canvas is sufficient and a faster workflow matters. Consider max when the reference contains detail you need to retain at a larger size and the added processing time is acceptable. The ComfyUI guide describes match as the default fit-to-canvas behavior and max as keeping a 2048-pixel short edge while running much more slowly. It does not promise a quality improvement for every image.
Check the input before choosing max
- Open the R2V workflow and locate the
ref_image_sizecontrol on the H3 reference-to-video node. Confirm that the reference image is connected as intended. - Record the source width and height in pixels. The short edge is the smaller dimension, regardless of whether the image is portrait or landscape.
- If choosing
max, compare the short edge with 2048 pixels. If it is below 2048 and you intend to use a 2048-pixel short edge, pre-scale the image in an image editor or another suitable preprocessing step; do not expect the node to upscale it. - After any scaling, check both dimensions. The guide says to round the sides to multiples of 32. Record the dimensions you will actually use rather than assuming the control will make this adjustment for you.
- Set
matchormax, then run the same target canvas and reference setup for your comparison. Note the choice and reason in a preprocessing record.
For a hypothetical 1200 × 1800 portrait image, the short edge is 1200 pixels. If the decision is to try max at a 2048-pixel short edge, pre-scaling is needed first. One possible proportional resize is approximately 2048 × 3072; both sides are divisible by 32. This is a calculation example, not a claim that those dimensions guarantee a better result.
Compare with a useful acceptance record
Keep the target canvas and prompt fixed while comparing the options. Judge the result against the reference’s intended role, and record runtime as an observation from your own workflow, not a promised benchmark.
| Hypothetical input and need | Setting and preparation | What to record at review | Decision signal |
|---|---|---|---|
| 1200 × 1800 portrait; reference mainly guides overall appearance | match; no pre-scaling | Whether the intended identity or style is recognizable; elapsed run time | Keep match if it meets the brief and speed matters |
| Same image; fine source details matter; short edge target is 2048 | Pre-scale to about 2048 × 3072; verify multiples of 32; use max | Whether those details are useful in the output; elapsed run time | Keep max only if the observed benefit justifies its slower processing |
| 1600 × 900 landscape; do not want to enlarge the source | match, or reconsider whether max is appropriate | Canvas fit and whether the reference still serves its role | Do not choose max expecting it to upscale the 900-pixel short edge |
These are hypothetical decision cases, not reported tests. Use the same output canvas before comparing: otherwise, a change in canvas can confound the comparison with the ref_image_size setting.
If max is slow or the reference still falls short
If max takes too long for your workflow, return to match and decide whether the reference remains adequate for its assigned role. If it does not, first inspect whether the source image itself contains the detail you need. Pre-scaling changes pixel dimensions; the documentation does not say that it creates missing source detail or guarantees a stronger reference result.
If you pre-scale, verify the resulting width and height and record how the source was transformed. The documented guidance specifies 2048 pixels on the short edge for max and sides rounded to multiples of 32; it does not specify a particular resizing algorithm or a universal rule for rounding dimensions. Do not infer a crop, aspect-ratio correction, or automatic upscale from the control name.
Common questions
Does max always produce a more faithful reference?
The documentation establishes its size behavior and slower processing, not a guaranteed improvement in fidelity. Compare outputs against the reference’s intended role and keep the setting only if the observed result justifies the cost.
What if my short edge is already greater than 2048 pixels?
The provided guide does not explain how max handles every above-target input. Do not assume it preserves every original pixel or applies a particular downscale. Record the source dimensions and inspect the result in your workflow.
Do I need to make both sides multiples of 32?
The guide says to round sides to multiples of 32. Apply that check to both width and height after pre-scaling, and record the actual dimensions. It does not document an automatic correction performed by the node.
Sources and scope
The behavior described here is from ComfyUI’s official MiniMax H3 workflow guide. It concerns the H3 R2V ref_image_size choice, not general resolution capabilities across image or video tasks.
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 film workspace. The external software controls discussed above belong to their respective products.