MiniMaxAI

/

MiniMax-H3

Like 5.43k

Follow

MiniMax 10.7k

I suspect that its due to higher resolutions making it so there are more tokens for the transformer backbone to process on a single pass,
More steps at a fixed lower res is a linear increase of compute, more res is quadratic increase of tokens (doesnt mean its quadratic in compute, but it still means that the context is bigger and bigger)

Hey guys,

I'm running the official MiniMax H3 model from Hugging Face on 4xB300, and I'm seeing a pretty significant difference in prompt adherence depending on the generation resolution.

With the same REF2VA prompt:

At 352p / 416p, the model generally follows the requested shot structure, camera angles, character placement, and opening-frame conditioning correctly.

At 768p, prompt adherence becomes dramatically worse. It may ignore requested camera angles, change character placement, ignoring duplicate interdiction...

Increasing the number of inference steps at 768p seems to improve/refine visual coherence somewhat, but doesn't really fix the structural guidance or composition.

I understand that changing resolution changes the latent dimensions, so I don't expect the same seed to produce an identical video across resolutions. But what surprises me is the systematic difference in instruction adherence, not just the resulting pixels/composition.

Has anyone else observed H3 following REF2VA / keyframe / shot guidance significantly better around 352p–416p than at 768p?

Is 768p native inference expected to behave differently, or are there resolution-specific settings / scheduler / RoPE / positional encoding / preprocessing parameters that need to be adjusted when scaling up the resolution?

I'm running that with a low step count, but increasing the steps with same seed and prompt will just make it a bit better looking, it will not remove undesired characters or problems...

Yes, I have definitely observed too that prompt adherence is almost spot on at low resolutions than high resolution. Almost like it skips some of the request being made in the prompt. I also found that strange. But even so this is still overall a big leap for open source video models. They did a great job with prompt understanding of this model.

Does higher resolution require higher steps perhaps? or steps really doesnt impact prompt adherence?

I have also noticed this issue as well

Can confirm. Although it's not dramatic and can be caused by model now factoring in things that were too small and ignored, so it needs even more strict prompting with higher res? We need an upsampler with reference support.

I noticed this as well, but for my case I find that 576p is best for prompt adherence

I too am experiencing significant differences between my lo-res trial generations and the final hi-res generation. It is frustrating as my lo-res version is exactly what I want! To run a lo-res, my gen time is about 4 minutes. To run a hi-res takes about half an hour! To then see something I don't want, frankly, pisses me off.
I have only recently got into AI video so it could be I am doing something wrong. Perhaps as suggested earlier in the thread H3 is stepping to a more elaborate language interpreter (eg: 3B to 7B)?
@Decadence-1 : Your suggestion (I think) of using the lo-res generation to act as a guide (ref2VA workflow) seems to be a really good idea... but as yet I've got no idea how to implement it as I am starting with a reference (first) frame and generating from there. So sort of a mix between F2VA and Ref2VA?

I would separate two questions here: image refinement and structural adherence. More steps can improve local coherence, but it is useful to test whether the requested shot structure survives the resolution change independently.

For a reproducible comparison, keep the same reference inputs, duration, scheduler, seed where applicable, and one compact prompt. Change only the resolution. Then score a small checklist for each run:

• subject count and placement
• camera direction
• first-frame / reference preservation
• one required action or temporal change
• unwanted additions or duplicated subjects

At the higher-resolution pass, I would also reduce prompt competition: one dominant camera movement, one primary action, and a short explicit preservation clause for only the elements that truly must stay fixed. For example:

Preserve the subject identity, composition, and camera framing.
The camera performs one slow forward push-in.
The subject takes one step forward; no additional people enter the scene.
Keep the opening-frame placement stable while background motion stays subtle.

For a production workflow, a lower-resolution run that has the correct shot structure can be useful as a reference for the next motion or refinement experiment, rather than treating it only as a disposable preview. The important part is documenting the exact configuration beside each result; otherwise it is hard to tell whether the loss of adherence comes from resolution, a changed scheduler, reference preprocessing, or a longer competing prompt.

I would be interested to see a minimal reproduction with the same input and only the resolution changed. That would make the comparison much easier for others to validate.